Deep learning for secure imaging and video surveillance in smart cities: from bibliometric mapping to real-world implementations

1 Introduction

The integration of deep learning with video analytics enables automated perception in large-scale visual sensing systems. Shifting inference from centralized cloud infrastructure to edge-based processing reduces latency and bandwidth requirements while supporting deployment in resource-constrained environments (Badidi et al., 2023). Contemporary systems operate as distributed architectures combining edge and cloud resources, with some implementations incorporating blockchain-based mechanisms for data integrity, access control, and auditability (Myagmar-Ochir and Kim, 2023).

Deep neural inference introduces security risks within imaging systems. Models are vulnerable to adversarial perturbations, i.e., inputs constructed to induce incorrect outputs. These vulnerabilities are relevant in UAV and satellite-based image interpretation frameworks and geospatial inference workflows, where prediction errors may propagate to downstream geospatial decision processes (Kazmi et al., 2023). In parallel, identity-centric tasks such as person search rely on multi-modal feature representations (visual, textual, and attribute-based) across distributed sensing nodes, introducing privacy risks associated with cross modal alignment, re-identification, and cross-domain inference (Irene et al., 2024).

Secure imaging systems therefore require guarantees on data confidentiality, output integrity, and interpretability of model decisions. In remote sensing (RS), Trustworthy Remote Sensing Interpretation (TRSI) integrates data-driven learning with knowledge-based representations to improve interpretability and robustness in geospatial analysis (Wang S. et al., 2024). These properties are critical in mission critical applications such as urban planning, disaster response, and infrastructure monitoring. Deep learning-based change detection (CD) further requires controlled data provenance, semantic label consistency, and validation protocols to ensure reliability in security-sensitive monitoring tasks (Wang L. et al., 2024).

Recent imaging studies further demonstrate that reliability challenges in visual intelligence extend across heterogeneous sensing modalities and task formulations. Graph-structured representations have been applied to compact high-frequency surface wave radar to maintain plot classification performance under low-SINR conditions (Li et al., 2025b). Liu et al. (2026) developed a visual–textual mutual guidance fusion network for remote sensing visual question answering, emphasizing explicit cross-modal semantic alignment between small image objects and textual queries. Wang et al. (2023) further showed that degraded acquisition conditions remain a key limitation in underwater image enhancement before high-level recognition or surveillance inference is performed. These studies support the broader premise that secure visual intelligence must account for sensing heterogeneity, non-ideal acquisition conditions, and cross-modal reliability as models are integrated into multi-sensor UAV and satellite platforms.

The increasing deployment of UAV platforms equipped with heterogeneous sensing modalities, including optical, infrared, and LiDAR, expands spatial and temporal coverage. However, this expansion also increases the scale of sensitive data acquisition and the associated privacy exposure (Zhang and Zhu, 2023). As sensing resolution and coverage increase, constraints on data governance and access control become more stringent.

Security and privacy requirements are closely coupled with system deployment models. Edge-based video analytics improves responsiveness by performing inference near the sensor, but introduces additional system-level considerations, including distributed attack surfaces, model update consistency, and data synchronization constraints (Badidi et al., 2023). Generative models for anomaly detection improve event recognition in streaming data but require evaluation under operational conditions, including robustness to noise, distribution shift, and adversarial inputs (Duong et al., 2023).

Despite rapid development, the research landscape remains fragmented. Existing surveys address individual components, including edge analytics (Badidi et al., 2023), person search (Irene et al., 2024), and adversarial robustness in aerial imagery (Kazmi et al., 2023). However, limited work provides a quantitative mapping of interactions between these domains or examines the interplay of privacy, trust, and learning-based methods within integrated imaging systems.

This study applies bibliometric science-mapping to characterize the development of secure imaging and video surveillance. Citation and keyword networks across smart cities, remote sensing, and computer vision are analyzed to identify thematic clusters and their evolution. The objective is to provide an evidence-based characterization of research directions relevant to secure imaging systems.

The study follows a three-phase framework (Figure 1). Phase I applies bibliometric analysis to identify thematic clusters. Phase II evaluates models using two criteria: hardware readiness (Tier 1) and operational validation under real-world conditions (Tier 2). Phase III synthesizes results across surveillance, remote sensing, and smart-city domains to derive research directions. The framework distinguishes between functional application of AI (“security by AI”) and assurance of model and system integrity (“security of AI”).

The remainder of this paper is organized as shown in Figure 2. Section 2 reviews the relevant literature, covering bibliometric foundations, prior reviews on secure visual intelligence, and the main research gaps related to robustness, privacy, integrity, and trust. Section 3 describes the bibliometric framework, including the search strategy, screening process, subgroup dataset construction, keyword normalization, and cluster generation. The subsequent section presents the procedure used to identify real-world implementation studies, distinguishing between physical or edge deployments and studies validated on real-world datasets. The results and discussion section, Section 4 then reports the bibliometric findings, including publication trends, geographical and institutional distributions, publication venues and overview of keyword co-occurrence structure. Subsequently, the analysis of prominent keywords in Section 5, and the review of deployment-oriented studies across the thematic subgroups in Section 6. This is followed by a discussion of design principles and implementation-level requirements for secure visual intelligence under deployment constraints in Section 7. The paper then outlines future challenges and research directions in Section 8, with emphasis on adversarial robustness, privacy-preserving surveillance, validation under real-world conditions, and trusted model architectures. Finally, Section 9 concludes by summarizing the main findings and their implications for secure imaging and video surveillance in smart city environments.

2 Related work

This section reviews prior work on AI-enabled imaging systems with respect to two technical dimensions: (i) functional security enabled by AI, (ii) model-level vulnerabilities. A comparison of the proposed bibliometric study with existing reviews is provided in Table 1.

Table 1. Representative related reviews and their primary focus areas compared to this study.

2.1 Security by AI: deep learning for urban safety

Deep learning (DL) underpins smart-city surveillance and monitoring systems. Edge-AI video analytics reduces end-to-end latency and network load by distributing compute-intensive tasks to edge servers, enabling real-time operation in transportation, healthcare, and public safety domains (Badidi et al., 2023). Integration with cloud–edge architectures and distributed ledgers has been reported to improve system coordination and data management in CCTV deployments (Myagmar-Ochir and Kim, 2023).

Representative application-level development include person search, which combines visual, textual, and attribute-based descriptors for localized identification (Irene et al., 2024), and generative models for anomaly detection in streaming video (Duong et al., 2023). In these systems, “security” is primarily defined at the application level (e.g., detection, monitoring, and response). In other words, it is not merely treated as a property of the underlying models.

2.2 Security of AI: adversarial robustness in imaging

DL-based perception systems are susceptible to adversarial perturbations. In UAV and satellite imaging, perturbations derived from constrained optimization can induce misclassification or misinterpretation in deep neural networks (DNNs), degrading geospatial inference (Kazmi et al., 2023). Object detection and tracking pipelines, including those based on generative adversarial architectures, are similarly vulnerable to input manipulation that affects real-time reliability (Al Jaberi et al., 2023).

Existing reviews provide detailed attack and defense taxonomies but are typically restricted to specific sensing modalities (e.g., aerial or satellite imagery). However, a multi-domain analysis and review of real-world use cases of security and adversarial spanning urban surveillance and remote sensing remains underexplored, which is the focus of this work.

2.3 Research gaps and taxonomic framework

Existing reviews are largely organized by application domain or threat category. Smart-city surveys emphasize system infrastructure without addressing the co-evolution of data, models, and knowledge representations (Myagmar-Ochir and Kim, 2023). Some of the studies treat security as a functional objective. In other words, it is not treated as an intrinsic property of model design (Irene et al., 2024; Duong et al., 2023). In contrast, adversarial robustness studies focus on specific sensing modalities and do not generalize across heterogeneous imaging ecosystems (Kazmi et al., 2023; Al Jaberi et al., 2023).

As summarized in Table 1, quantitative cross-domain studies remain limited, with existing reviews providing little coverage of integrated security, privacy, and adversarial robustness across surveillance and remote-sensing domains. To address this, we classify existing related literature into two main categories: (i) Security by AI (application-level functionality) and (ii) Security of AI (model robustness). This taxonomy provides a basis for analyzing thematic clusters using keyword co-occurrence structures using the proposed bibliometric study.

3 Methodology and bibliometric framework

The overall methodological workflow is shown in Figure 3. The framework consists of dual-database retrieval, thematic subgroup construction, keyword extraction, thesaurus-based normalization, co-occurrence mapping, cross-dataset analysis, and integrated interpretation of the bibliometric findings.

3.1 Data source, scope, and access date

Bibliographic records analyzed in this study were retrieved from the Scopus database (𝒟_*S*) and the Web of Science Core Collection (𝒟_*W*). Scopus is used as the primary dataset due to its broader indexing coverage, while Web of Science is used for cross-database validation.

The Scopus dataset 𝒟_*S* was accessed on 18 January 2026 at 16:13 (Kuala Lumpur time). The Web of Science dataset 𝒟_*W* was independently retrieved on 1 April 2026 at approximately 08:20 (Kuala Lumpur time) using an equivalent query formulation. All bibliometric analyses were conducted independently on each dataset using identical parameter settings to ensure comparability of results.

The analysis includes records indexed from 2010 to early 2026, covering a period of rapid growth in deep learning for computer vision, surveillance, and remote sensing. However, because data collection took place before the end of 2026, the 2026 records represent only part of the year. Therefore, annual publication trends were interpreted only for the complete years 2010–2025. The 2026 records were kept in the corpus to reflect the most recent research topics, but they were not used to determine whether publication output increased or decreased in 2026. This avoids a premature interpretation of the partial 2026 publication count while keeping the keyword co-occurrence analysis up to date.

To preserve database-specific indexing characteristics and avoid ambiguity in record matching, the datasets were not merged:

DS≠DW(1)

No deduplication was performed. Each dataset was processed separately under the same methodological framework, enabling structural comparison of bibliometric patterns across databases.

3.2 Search strategy and query formulation

A unified Boolean query was constructed and applied to both databases using their respective metadata fields (Scopus: TITLE–ABS–KEY; WoS: Topic Search) (Tables 2, 3). The query integrates four semantic components: (i) Application contexts (smart cities, surveillance, remote sensing); (ii) Imaging and video modalities (image/video and computer-vision analysis); (iii) security-, privacy-, and trust-oriented mechanisms (e.g., encryption, authentication, watermarking, privacy preservation); and (iv) Advanced learning paradigms (deep learning, neural networks, CNNs, transformers, and adversarial learning).

Table 2. Complete Scopus search string used for first-round literature retrieval on secure imaging for surveillance, remote sensing, and smart-city applications (prior to Scopus EXACTKEYWORD subgroup filtering).

The query allowed records indexed up to early 2026 because retrieval was performed on 18 January 2026; however, 2026 was treated as a partial retrieval year and excluded from annual trend interpretation.

Table 3. Complete Web of Science (WoS) search string used for literature retrieval on secure imaging for surveillance, remote sensing, and smart-city applications.

The query allowed records indexed up to early 2026 because retrieval was performed on 1 April 2026; however, 2026 was treated as a partial retrieval year and excluded from annual trend interpretation.

The query was designed to capture the fragmented terminology used across surveillance engineering, remote sensing, computer vision, and cybersecurity communities. Only English-language publications indexed by Scopus and Web of Science as journal articles or conference papers (WoS: articles and proceedings papers) were retained to emphasize original technical contributions.

The Scopus query yielded 9,206 records, reduced to 9,076 after restricting the corpus to records indexed from 2010 to early 2026. The WoS query returned 2,571 records, with 2,565 retained after removing retracted publications.

3.3 Document-type and language screening

Language filtering was applied during the query stage to exclude non-English publications. Document-type filtering retained only journal articles and conference papers, supporting a focus on methodological- and system-level research in computer vision and surveillance. Other document types (e.g. editorials, notes, book chapters, and conference reviews) were excluded to maintain emphasis on original technical contributions.

3.4 Thematic subgroup dataset construction

Beyond the full corpus used to construct the global keyword co-occurrence network, additional subgroup datasets were extracted using Scopus keyword-based filtering aligned with Scopus-indexed subject descriptors (Table 4). These subgroup datasets were used to support targeted review of practical implementations.

Table 4. Scopus retrieval search logic for the main corpus and subgroup datasets.

The core query block is shared across all datasets; subgroup datasets apply additional EXACTKEYWORD filters. Full verbatim strings for the core query block are provided in Table 2.

The extracted subgroup datasets included Learning and Algorithms (7,785 records), Surveillance Systems (6,881 records), Detection/Recognition/Tracking (4,019 records), and Computer Vision and Image Processing (3,828 records). Smaller subsets comprise Data and Algorithms (2,588 records), Smart Cities (1,614 records), Remote Sensing (1,387 records), and Video Analysis (1,127 records).

These subgroups are not mutually exclusive; individual papers may appear in multiple subsets. In this study, we focus on video analysis, smart cities, detection/recognition/tracking, computer vision and image processing, and surveillance systems, with the implicit assumption that relevant works employ learning-based methods, as reflected in query constraints.

3.5 Bibliometric processing and network construction

Bibliometric mapping was performed using VOSviewer. Author-supplied keywords exported from Scopus were used to construct keyword co-occurrence networks. Total Link Strength (TLS) was used to quantify connectivity between keywords and to assess their integration within the broader semantic network.

3.6 Overall keyword set and threshold selection (𝒟_*S*)

The initial Scopus keyword set was processed using a thesaurus-based normalization procedure before co-occurrence analysis. After normalization, the Scopus keyword universe comprised 15,765 keywords. An exploratory VOSviewer analysis with a minimum occurrence threshold of five retained 759 keywords. Although this configuration provided broad coverage, the resulting network remained dense and less suitable for interpretable cluster-level discussion.

To improve thematic separation and readability, a stricter minimum occurrence threshold of ten was applied. Under this configuration, 322 keywords were retained and organized into 12 clusters. Both thresholds are reported for transparency, but the analytical discussion focuses on the ten-occurrence configuration because it provides a more compact and interpretable representation of the main thematic structure.

3.7 Keyword normalization and cleaning (𝒟_*S*)

A thesaurus-based preprocessing procedure was applied before network construction to harmonize spelling variants, plural and singular forms, hyphenation variants, and unambiguous acronym/full-form equivalents. The thesaurus reduced lexical redundancy while preserving terms with distinct methodological, task-specific, platform-specific, modality-specific, or application-specific meaning.

For example, acronymic and full-form variants such as cnn, convolutional neural networks, and deep convolutional neural network were consolidated under convolutional neural network; gan and related plural/full-form variants were consolidated under generative adversarial network; uav, uavs, and expanded unmanned-aerial-vehicle variants were consolidated under unmanned aerial vehicle; and YOLO variants were consolidated under yolo. Similarly, face-recognition variants, VGG variants, abnormal-behavior/anomaly-detection variants, and privacy-preservation variants were normalized according to the final thesaurus.

Table 5 provides examples of raw labels consolidated into selected family-level labels. These examples show that normalized nodes such as yolo, mobilenet, vgg, and resnet represent families of related terms rather than single algorithm versions. In the final post-thesaurus Scopus network generated with a minimum occurrence threshold of ten, the normalized yolo node has 589 occurrences and a total link strength of 1,657. This value is the standardized yolo node used in the VOSviewer map after thesaurus consolidation.

Table 5. Examples of raw labels mapped to selected normalized labels during thesaurus-based preprocessing.

The normalized labels are therefore interpreted as keyword families rather than individual model versions. This reduces unnecessary fragmentation in the co-occurrence network and clarifies cluster-level relationships. Each family-level node may still include multiple lexical variants or version-specific labels.

3.8 Cluster generation and thematic structure

3.8.1 Scopus (𝒟_*S*)

Using the normalized Scopus keyword set and a minimum occurrence threshold of ten, VOSviewer generated a co-occurrence network comprising 322 keywords organized into 12 clusters. The cluster structure separates the literature into the following thematic groups: Cluster 1—remote sensing, robust learning, and geospatial image enhancement; Cluster 2—video surveillance, action recognition, and video anomaly analysis; Cluster 3—CNN-based biometric recognition and identity-centric surveillance analytics; Cluster 4—real-time object detection, traffic surveillance, and intelligent transportation systems; Cluster 5— security surveillance, UAV sensing, detection, and sensor fusion; Cluster 6—edge/cloud smart-city security and privacy-preserving analytics; Cluster 7—crowd analytics, public-health monitoring, and edge implementation; Cluster 8—re-identification, segmentation, and video surveillance scene analysis; Cluster 9—computer vision, gait recognition, and event/attention modeling; Cluster 10— tracking, target detection, and real-time surveillance calibration; Cluster 11—multi-object/pedestrian tracking and spatial-attention systems; and Cluster 12—foreground/motion detection, human tracking, and thermal imaging. The complete keyword list for each cluster is provided in the Supplementary file.

Security and privacy were analyzed as cross-cutting concerns. In the Scopus structure, security-oriented terms appear across several clusters. Cluster 5 captures surveillance, detection, UAV sensing, radar, sensor fusion, cybersecurity, intrusion detection, encryption, privacy, safety, and real-time monitoring. Cluster 6 captures privacy-preserving and infrastructure-oriented smart-city analytics through edge computing, cloud computing, fog computing, IoT, federated learning, blockchain, privacy protection, explainable AI, and lightweight models. Cluster 2 contributes video-anomaly and surveillance-oriented security terms, while Cluster 7 contributes public-safety, public-health, and edge-implementation terms.

Remote sensing remains a major imaging domain in the network. Cluster 1 groups remote sensing, remote sensing imagery, satellite imagery, SAR, LiDAR, hyperspectral data, Sentinel-2, semantic segmentation, image enhancement, image fusion, image restoration, super-resolution, precision agriculture, crop monitoring, ship detection, and adversarially robust learning. This cluster indicates that remote-sensing image analysis forms a distinct but connected branch of secure visual intelligence, extending the review beyond fixed surveillance cameras toward aerial, satellite, agricultural, and maritime sensing contexts.

3.8.2 WoS (𝒟_*W*)

The thesaurus-normalized WoS keyword set was analyzed in VOSviewer using two minimum occurrence thresholds. The lower threshold of 5 retained 204 keywords and provided a fine-grained view of the WoS thematic structure, while the higher threshold of 15 retained 60 keywords and was used as a compact visualization of the dominant themes. At the threshold of 5, the WoS co-occurrence network yielded 11 clusters. These clusters represent the main thematic structure of the WoS dataset: Cluster 1 covers remote sensing, agriculture, and geospatial modeling; Cluster 2 captures deep learning, biometrics, and surveillance perception; Cluster 3 represents video surveillance, anomaly analysis, and real-time detection; Cluster 4 captures adversarial robustness, reconstruction, and security modeling; Cluster 5 covers multimodal sensing, transformers, and anomaly/event analysis; Cluster 6 represents edge/cloud security, privacy, and smart-city analytics; Cluster 7 captures domain adaptation, feature extraction, and semantic image analysis; Cluster 8 reflects accuracy, computational efficiency, and monitoring; Cluster 9 covers SAR, maritime surveillance, and optical/radar sensing; Cluster 10 captures attention, crowd analytics, and object detection; and Cluster 11 represents authentication, subject-sensitive hashing, and U-Net segmentation. The complete keyword list for each cluster is provided in the Supplementary file.

Security and privacy are distributed across multiple clusters rather than being confined to a single thematic group. Cluster 4 captures adversarial robustness and attack-related modeling, while Cluster 6 emphasizes privacy, encryption, blockchain, federated learning, edge/cloud computing, and IoT-enabled analytics. Cluster 11 further captures authentication and subject-sensitive hashing. These security-related aspects also intersect with surveillance perception (Clusters 2 and 3), tracking and object-detection themes (Clusters 3, 6, and 10), and multimodal sensing (Cluster 5), indicating that security concerns are embedded across both model-level and system-level research themes.

Remote sensing constitutes a major thematic component of the WoS network. Cluster 1 covers satellite imagery, hyperspectral data, LiDAR, semantic segmentation, and agricultural/geospatial modeling, while Cluster 9 captures SAR, maritime surveillance, ship detection, optical sensing, and radar polarimetry. UAV-related terms appear in Cluster 1 and Cluster 2, indicating that UAV platforms connect geospatial remote sensing with surveillance perception. This structure extends the scope of secure imaging beyond fixed ground-based camera networks toward large-scale aerial, satellite, radar, and multimodal sensing systems.

3.9 Integration of subgroup analyses (scopus)

The main Scopus dataset was further organized into thematic subgroup datasets using Scopus EXACTKEYWORD-based filtering, as shown in Table 4. These sub-datasets were used to identify relevant papers for review in Section 6. The subgroup analyses supported the selection of practical implementation studies.

3.10 Systematic identification of real-world implementation studies

As a core essence of this paper, based on the data (abstracts and full text), we aim to review works that have considered real-world implementations or are close to real implementations as opposed to those focusing mainly on simulation or pure algorithmic proposals without real-world datasets. We thus applied a validation framework comprising two phases. The first phase evaluates titles and abstracts within the large Scopus datasets for each subgroup, while the second phase assesses full-text content, with only studies that pass this stage retained. Then the selected papers are categorized into tiers as explained in the next subsection. Based on the tiered selection, studies were excluded if they relied solely on (i) simulations without empirical visual data (e.g., purely network simulators or robotics simulators without real imagery/video), (ii) purely conceptual frameworks with no experimental validation, or (iii) synthetic datasets with no real-data evaluation relevant to surveillance or remote sensing.

3.10.1 Tier 1 (physical/edge implementation)

A study is categorized as Tier 1 if it reports hardware-level or edge-level validation. Eligible evidence includes: (i) deployment on embedded or edge platforms (e.g., FPGA, NVIDIA Jetson-class devices, Raspberry Pi, UAV/robot controllers, smart cameras); (ii) edge–cloud prototypes with measured runtime characteristics (e.g., latency, throughput/FPS, energy, memory footprint); or (iii) on-board sensing demonstrations (e.g., UAV-captured video processed on-device). Purely conceptual “edge” architectures without measured implementation evidence are not considered Tier 1.

3.10.2 Tier 2 (real-world dataset validation)

A study is categorized as Tier 2 if it is validated on real, non-synthetic imaging or video data relevant to surveillance, smart cities, or remote sensing (e.g., CCTV footage, body-worn camera datasets, satellite/UAV imagery, and hyperspectral datasets), but does not report hardware/edge deployment. Tier 2 studies typically demonstrate algorithmic effectiveness through standard quantitative metrics (e.g., mAP, accuracy, PSNR/SSIM, and detection rates) using real datasets.

3.11 Data synthesis and analysis

The final verified corpus was used to quantify the works that fall between real-dataset-based algorithmic solutions (Tier 2) and hardware/edge-based solutions (Tier 1). This enables analysis of security objectives (privacy, confidentiality, integrity, robustness) across application domains within this area.

3.12 AI assistance and transparency

Generative AI tools (ChatGPT, OpenAI; Gemini, Google) were used as auxiliary aids for initial screening support, interpretation of bibliometric keyword co-occurrence patterns, and manuscript structuring. Retrieval, filtering, mapping, clustering, final study selection, and all interpretations were manually verified by the authors.

4 Results and discussion

This section reports (i) performance indicators (related to publications, leading countries, institutions, and sources) and (ii) science mapping via keyword co-occurrence and thematic clustering.

4.1 Publication growth and temporal dynamics

4.1.1 Scopus-based analysis (𝒟_*S*)

Figure 4 illustrates the annual publication output of the Scopus-derived corpus for complete calendar years from 2010 to 2025. Although early-2026 records were retained in the bibliometric corpus for thematic coverage, they were excluded from annual trend interpretation because the Scopus retrieval was conducted on 18 January 2026. The data reveal a low-output phase from 2010 to 2016, followed by a pronounced acceleration beginning in 2017, mainly due to the proliferation of the applications of deep learning. Particularly, this could be a result of the widespread adoption of deep learning architectures for visual perception tasks, particularly within security-sensitive surveillance and remote-sensing applications.

From 2021 onward, the number of publications increased in a pronounced fashion, with annual publication counts consistently exceeding 900 and peaking in 2025. This period perhaps indicates the maturity of computer-vision research and applications, the integration of learning-based security and privacy mechanisms, and the expansion of smart-city and aerial-imaging use cases, requiring demand for scalable, real-time, and trustworthy imaging solutions.

The publication count for 2026 was excluded from the annual trend plots because the retrieval dates provide only partial-year coverage. Therefore, the lower number of records in 2026 should not be interpreted as a decline in research activity. Overall, the observed growth pattern indicates a transition from exploratory research toward sustained, application-driven development in secure imaging for surveillance, remote sensing, and smart urban environments.

4.1.2 WoS-based analysis (𝒟_*W*)

The analysis based on the WoS dataset (N_*W* = 2, 565) exhibits comparable structural patterns. The corresponding results are presented to enable cross-database validation. Figure 5 presents the annual publication output of the Web of Science (WoS)-derived corpus over the period 2010–2025. A relatively low and stable publication rate is observed from 2010 to 2016, followed by a marked increase from 2017 onwards. This transition coincides with the widespread adoption of deep learning techniques in visual perception tasks, particularly in surveillance and remote sensing domains. The category labeled “Others” includes document types that do not correspond to primary research outputs, such as editorial materials, letters, reviews, and other indexed publication formats within WoS.

4.1.3 Cross-database comparison

A comparison between 𝒟_*S* (Scopus) and 𝒟_*W* (Web of Science) indicates a generally consistent temporal trend, characterized by a low publication output prior to 2017 followed by a marked increase in subsequent years. Minor discrepancies in annual publication counts are observed, which can be attributed to differences in database coverage, indexing policies, and document type classification schemes. In particular, WoS employs a more granular document type taxonomy, where certain records may be categorized under non-primary types (e.g., editorial material, letters, or other indexed formats), thereby contributing to variations in aggregated counts across databases.

4.2 Geographical distribution of research output

4.2.1 Scopus-based analysis (𝒟_*S*)

Figure 6 summarizes the top 10 contributing countries in the corpus obtained from Scopus. Here, the top 10 ranking confirms that although research output is globally distributed, it is also strongly concentrated within the top two countries. China (n = 2, 850) and India (n = 2, 465) dominate publication volume by a wide margin, indicating that a substantial proportion of research activity and published works in secure imaging for surveillance, smart cities, and remote sensing. This is possibly due to lots of engagement in research in these areas, with sustained interest and investment in computer vision, AI, and cyber-physical security.

The United States follows these two countries by a large margin (n = 732), followed by South Korea (n = 490) and the United Kingdom (n = 291). A notable feature of the top-10 group is the presence of multiple countries with comparatively equal output in research on security-aware sensing and public-safety technologies, including Saudi Arabia (n = 275) and Pakistan (n = 274), alongside countries such as Italy (n = 213), Canada (n = 195), and Australia (n = 179).

Beyond the top 10, the next tier of contributors in the top-20 list (not shown in the figure to save space) includes: Japan (n = 178), Spain (n = 163), Taiwan (n = 152), Turkey (n = 142), Malaysia (n = 133), France (n = 132), and Germany (n = 127). The countries, among others, show that there is a diverse international engagement on this subject. Collectively, these patterns may be a result of national interest in (i) broad AI and smart-infrastructure initiatives and (ii) researchers' interests in intelligent surveillance, remote sensing, and public safety in these countries.

4.2.2 WoS-based analysis (𝒟_*W*)

As shown in Figure 7, the WoS country collaboration network is centered on the People's Republic of China, which forms the largest and most connected node. India also appears as a prominent contributor within the central cluster, with strong links to China and other major countries. The United States acts as another major collaboration hub, particularly through links with European contributors such as England, Italy, Spain, Germany, and France, while Saudi Arabia, Pakistan, South Korea, and Australia connect additional regional clusters. These patterns indicate that WoS-indexed research activity is concentrated around a small set of high-output countries with strong international collaboration ties. The country-level analysis based on the Web of Science dataset 𝒟_*W* indicates a strong dominance of the People's Republic of China, contributing 571 documents and 7,976 citations, followed by the United States with 174 documents and 5,825 citations, demonstrating high citation impact relative to publication volume. India ranks third with 263 documents and 1,696 citations, reflecting substantial research activity.

Other notable contributors include Saudi Arabia (86 documents, 1,121 citations), Pakistan (81 documents, 871 citations), and South Korea (97 documents, 955 citations), alongside European countries such as Italy (49 documents, 542 citations), England (46 documents, 561 citations), and Spain (43 documents, 508 citations) also maintain consistent research output.

4.2.3 Cross-database country comparison

The comparison between 𝒟_*S* and 𝒟_*W* demonstrates strong agreement in the identification of leading countries, particularly the dominance of China, the United States, and India. While the overall ranking structure remains consistent, minor variations in absolute publication counts and ranking positions are attributed to differences in database indexing coverage, regional journal inclusion policies, and citation aggregation mechanisms.

Importantly, the alignment between both datasets supports the robustness of the dual-database framework and reinforces the validity of subsequent bibliometric and thematic analyses.

4.3 Institutional productivity

4.3.1 Scopus-based analysis (𝒟_*S*)

Table 6 reports leading affiliations. These include the presence of national research organizations and large research-intensive universities, indicating that research in surveillance systems, security mechanisms, and AI-based imaging analytics are fundamental research area of interest across major institutions in the globe.

Table 6. Top affiliations by number of documents (Scopus).

The Chinese Academy of Sciences leads with 228 documents, followed by the Ministry of Education of the People's Republic of China (163 documents) and the SRM Institute of Science and Technology (119 documents). Other prominent institutions include the University of Chinese Academy of Sciences (108 documents), Wuhan University (82 documents), and Zhejiang University (73 documents).

4.3.2 WoS-based analysis (𝒟_*W*)

The institutional contribution analysis based on 𝒟_*W* highlights a strong dominance of Chinese research organizations. The Chinese Academy of Sciences leads with 69 documents and 996 citations, followed by the University of Chinese Academy of Sciences with 33 documents and 559 citations. Other notable contributors include Wuhan University (27 documents, 467 citations), Zhejiang University (25 documents, 854 citations), and Xidian University (22 documents, 343 citations).

In addition to Chinese institutions, international contributors such as King Saud University (19 documents, 425 citations) and Nanyang Technological University (14 documents, 797 citations) demonstrate strong research impact, particularly in terms of citation performance.

4.3.3 Cross-database comparison

A comparison between 𝒟_*S* and 𝒟_*W* reveals strong agreement in institutional leadership, particularly the dominance of Chinese research organizations. The recurrence of institutions such as the Chinese Academy of Sciences, University of Chinese Academy of Sciences, and Zhejiang University across both datasets reinforces the robustness of the observed patterns.

4.4 Leading publication venues

4.4.1 Scopus-based analysis (𝒟_*S*)

Journal or Conference (Venue) productivity is summarized in Table 7. Popular multidisciplinary and engineering journals (e.g., IEEE Access) and conference proceedings series (e.g., SPIE; LNCS) dominate. This is relatable, as deep learning–enabled surveillance, security mechanisms, and remote-sensing analytics cut across science and engineering disciplines and applications. The presence of Remote Sensing and major IEEE geoscience outlets indicates a substantial stream of research on security and surveillance also falls within the scope of mainstream remote-sensing outlets.

Table 7. Top sources by number of documents.

4.4.2 WoS-based analysis (𝒟_*W*)

The WoS dataset (N_*W* = 2565) exhibits a consistent distribution of publications across high-impact journals. The most productive source is IEEE Access with 108 documents and 1267 citations, followed closely by Remote Sensing with 100 documents and 1231 citations. Other prominent sources include Sensors (53 documents, 636 citations), IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (54 documents, 859 citations), and IEEE Transactions on Geoscience and Remote Sensing (51 documents, 947 citations).

Additional sources such as Multimedia Tools and Applications (36 documents, 369 citations), Scientific Reports (32 documents, 174 citations), and Applied Sciences-Basel (29 documents, 154 citations) further contribute to the publication landscape. These results indicate a strong concentration of research output within IEEE and remote sensing-oriented venues.

4.4.3 Cross-database comparison

A comparison between 𝒟_*S* and 𝒟_*W* indicates strong agreement in the ranking of dominant publication venues, particularly the prominence of IEEE journals and remote sensing-focused outlets. Minor discrepancies in source rankings and document counts can be attributed to differences in database coverage policies, especially for multidisciplinary journals and open-access platforms.

4.5 Top authors

4.5.1 Scopus-based analysis (𝒟_*S*)

The Scopus-based author productivity analysis, illustrated in Figure 8, highlights Baik, S.W. as the leading contributor with 37 publications, followed by Muhammad, K. (33) and Granger, E. (20). Other leading contributors include Khan, M.A. (19), Rho, S. (18), Hussain, T. (16), López-Rubio, E. (16), De Albuquerque, V.H.C. (15), and Ullah, A. (15).

Figure 8 focuses on the top 13 authors (minimum of 13 publications). This threshold includes Jalal, A. (14), Park, K.R. (14), Ullah, W. (14), and Ahmed, I. (13). Authors situated immediately below this threshold, including Chen, Y., Jo, K.H., and Khan, N. (each with 12 publications) —are recognized for their contributions but are excluded from the figure to maintain the Top-13 threshold and preserve visual readability.

4.5.2 WoS-based analysis

The WoS dataset (𝒟_*W*) is utilized to provide an explicit evaluation of top contributing authors, capturing nuances not reflected in the Scopus-based analysis. As illustrated in Figure 9, the authorship structure is characterized by a high degree of distribution. Mahmood, Arif leads as the most prolific author with 8 publications and 92 citations. He is followed by several researchers including Chen, Yu, Ding, Kaimeng, Khan, Muhammad Attique, Liu, Yang, and Zhang, Peng, each contributing 7 publications. The Top-10 representation was chosen to ensure visual readability while preserving the principal authorship patterns of the dataset. In terms of citation impact, Huang, J. demonstrates notable influence, accumulating 153 citations across 5 publications.

4.5.3 Cross-database comparison

A comparison between Scopus and WoS author productivity reveals partial agreement in leading contributors, particularly for authors such as Mahmood, A., Chen, Y., and Khan, M.A. Minor discrepancies in ranking positions can be attributed to differences in database coverage across 𝒟_*S* and 𝒟_*W*. (Figure 10).

4.6 Keyword co-occurrence analysis

4.6.1 Scopus-based analysis (𝒟_*S*)

After thesaurus-based normalization and application of the minimum occurrence threshold of ten, VOSviewer produced a Scopus keyword co-occurrence network comprising 322 keywords organized into 12 clusters. The network was obtained from 15,765 post-thesaurus keywords. At the lower occurrence threshold of five, 759 keywords were retained. The 12-cluster structure obtained using a minimum occurrence threshold of ten is summarized in Table 8.

Table 8. Integrated summary of Scopus keyword co-occurrence clusters derived from VOSviewer after thesaurus-based normalization (minimum occurrence threshold ≥10).

The network retains 322 keywords from 15,765 post-thesaurus keywords and organizes them into 12 thematic clusters.

Table 9 summarizes the most prominent keywords in the cleaned co-occurrence network using both occurrence counts and Total Link Strength (TLS). Occurrences indicate topical prominence, whereas TLS indicates the extent to which a keyword is connected to other keywords in the co-occurrence network. The leading terms after thesaurus normalization are deep learning, convolutional neural network, video surveillance, object detection, computer vision, and yolo. Their high occurrence and TLS values indicate that the field is centered on learning-based visual perception, surveillance video analysis, and object-level detection. The high ranks of anomaly detection, artificial neural network, machine learning, face recognition, remote sensing, unmanned aerial vehicle, and generative adversarial network further show that the corpus spans temporal anomaly analysis, biometric recognition, geospatial sensing, aerial platforms, and generative modeling.

Table 9. Top keywords (Scopus) in the cleaned VOSviewer network after thesaurus-based normalization (minimum occurrence threshold ≥10), ranked by occurrences.

TLS denotes Total Link Strength, i.e., keyword connectivity within the co-occurrence network.

The 12-cluster structure shows a research landscape organized around both methodological and deployment-specific themes. Cluster 1 groups remote sensing, remote sensing imagery, satellite imagery, SAR, LiDAR, hyperspectral data, Sentinel-2, semantic segmentation, image enhancement, image fusion, image restoration, super-resolution, precision agriculture, ship detection, and adversarial robustness. This cluster establishes remote sensing and geospatial image enhancement as a major component of the corpus, with robust learning and restoration-related methods embedded in the same thematic region.

Cluster 2 captures video surveillance, action recognition, video anomaly detection, violence detection, activity recognition, optical flow, recurrent neural networks, transformers, autoencoders, generative adversarial networks, UCF-Crime, and weakly supervised learning. This cluster represents the temporal video-understanding component of the field, where learning models are used to infer activities, anomalies, and security-relevant events from surveillance video.

Cluster 3 is centered on convolutional neural networks and biometric recognition. It includes face detection, face recognition, face verification, FaceNet, MTCNN, gait, gender classification, gender recognition, age estimation, person identification, transfer learning, ResNet, VGG, MobileNet, EfficientNet, support vector machine, local binary pattern, and histogram of oriented gradients. This cluster indicates that CNN-based recognition remains closely associated with biometric and identity-centric surveillance analytics.

Cluster 4 covers real-time object detection, traffic surveillance, and intelligent transportation systems. It includes object detection, object localization, object recognition, YOLO, Faster R-CNN, SSD, RetinaNet, Mask R-CNN, instance segmentation, CCTV, surveillance systems, traffic monitoring, traffic surveillance, vehicle detection, vehicle tracking, vehicle counting, license-plate recognition, helmet detection, and road safety. This cluster provides the object-level perception layer of the Scopus network.

Clusters 5 and 6 show that security, privacy, and infrastructure are distributed across the network. Thus, they are not confined to a single cluster. Cluster 5 links surveillance, unmanned aerial vehicle, detection, behavior analysis, crime detection, drone detection, radar, sensor fusion, cybersecurity, intrusion detection, encryption, privacy, safety, tracking, and video analysis. Cluster 6 links smart cities, artificial intelligence, edge computing, edge devices, cloud computing, fog computing, IoT, federated learning, blockchain, privacy protection, explainable AI, image encryption, lightweight models, video analytics, and video compression. Together, these clusters represent security-oriented monitoring and privacy-aware smart-city deployment.

The smaller clusters identify more specialized but coherent subthemes. Cluster 7 groups crowd analytics, density estimation, public safety, public-health monitoring, person detection, pose estimation, OpenCV, Python, TensorFlow, COVID-19, face-mask detection, social distancing, and weapon detection. Cluster 8 links deep learning with re-identification, metric learning, Siamese networks, triplet loss, foreground segmentation, background subtraction, image segmentation, video segmentation, moving-object detection, and scene understanding. Cluster 9 combines computer vision, gait recognition, graph neural networks, multi-task learning, reinforcement learning, pedestrian attribute recognition, soft biometrics, trajectory, visual attention, and real-time systems. Clusters 10–12 further separate tracking and calibration, multi-object/pedestrian tracking with spatial attention, and foreground/motion detection with human tracking and thermal imaging.

Overall, the Scopus co-occurrence structure supports an interpretation of secure imaging as a multi-layered visual-intelligence field. The network separates remote sensing and image enhancement, temporal surveillance-video understanding, biometric recognition, object detection, UAV/security monitoring, smart-city edge/cloud infrastructure, crowd and public-health analytics, re-identification, event/attention modeling, tracking, and thermal motion analysis. This structure provides the basis for the selected-keyword analysis in the next section, where individual high-frequency terms are examined through their VOSviewer maps and cluster-neighborhood relationships.

4.6.2 WoS-based analysis and keyword augmentation (𝒟_*W*)

The clustering structure derived from 𝒟_*W* provides an independent view of the thematic organization observed in the Scopus dataset. Following thesaurus-based normalization, the WoS keyword co-occurrence analysis was examined using two complementary thresholds. The lower-threshold configuration retained 204 keywords and produced 11 thematic clusters, as summarized in the Supplementary file. This configuration provides a broader thematic inventory of the WoS corpus. A stricter minimum occurrence threshold of 15 retained 60 keywords and was used for the compact WoS map in Figure 11.

The 204-keyword WoS structure separates the literature into 11 thematic groups: remote sensing, agriculture, and geospatial modeling; deep learning, biometrics, and surveillance perception; video surveillance, anomaly analysis, and real-time detection; adversarial robustness, reconstruction, and security modeling; multimodal sensing, transformers, and anomaly/event analysis; edge/cloud security, privacy, and smart-city analytics; domain adaptation, feature extraction, and semantic image analysis; accuracy, computational efficiency, and monitoring; SAR, maritime surveillance, and optical/radar sensing; attention, crowd analytics, and object detection; and authentication, hashing, and U-Net segmentation.

At the stricter 15-occurrence threshold, the most prominent WoS terms include deep learning (543 occurrences), convolutional neural network (186), remote sensing (174), object detection (114), feature extraction (113), computer vision (103), video surveillance (91), machine learning (78), transformer (71), unmanned aerial vehicle (71), yolo (70), surveillance (69), and face recognition (66). These terms indicate that the WoS corpus preserves the main methodological and application-oriented structure observed in Scopus, although with fewer retained terms and a more compact network.

As shown in Figure 11, the high-threshold WoS map emphasizes deep learning, convolutional neural networks, object detection, remote sensing, computer vision, transformer-based methods, surveillance, privacy, and security. The lower-threshold cluster inventory in the Supplementary file provides a higher-resolution view of related topics, including federated learning, edge computing, SAR, maritime surveillance, domain adaptation, authentication, and hashing. Together, both threshold settings show that the dominant WoS themes remain consistent while the lower threshold reveals finer-grained thematic subdivisions.

4.6.3 Cross-database comparison

The clustering structures of 𝒟_*S* and 𝒟_*W* demonstrate strong thematic consistency. Both datasets identify deep learning, convolutional neural networks, object detection, computer vision, surveillance, video surveillance, remote sensing, UAV-based sensing, privacy, and security as recurring components of the research landscape. The Scopus network provides a finer-grained structure, retaining 322 keywords across 12 clusters, whereas the WoS analysis provides a more compact but thematically consistent structure, retaining 204 keywords across 11 clusters at the 5-occurrence threshold and 60 keywords at the stricter 15-occurrence threshold.

The observed differences between both databases are primarily structural. Scopus provides higher keyword granularity and separates several subthemes more explicitly, including re-identification, crowd analytics, foreground/motion detection, tracking, edge implementation, and thermal imaging. WoS preserves the same broad thematic organization but groups several related topics into more compact clusters. These differences are consistent with database-specific indexing coverage and Keyword Plus augmentation in WoS.

4.6.4 Thematic and cluster consistency

The overlap between the dominant Scopus and WoS keyword sets, shows substantial alignment in the main thematic anchors. Shared high-frequency terms include deep learning, convolutional neural network, object detection, computer vision, remote sensing, video surveillance, machine learning, surveillance, transformer, unmanned aerial vehicle, yolo, feature extraction, privacy, and security. This agreement indicates that both databases identify similar methodological and application-oriented cores despite differences in coverage and indexing.

Quantitatively, the high-threshold WoS network is led by deep learning (543 occurrences), convolutional neural network (186), remote sensing (174), object detection (114), feature extraction (113), and computer vision (103). These terms correspond directly to highly ranked Scopus terms, including deep learning, convolutional neural network, object detection, computer vision, video surveillance, and remote sensing. The agreement supports the interpretation that the observed research structure is not merely an artifact of a single bibliographic source.

4.6.5 Summary

Ranking comparisons across publication years, countries, institutions, sources, and keyword structures indicate broad correspondence between 𝒟_*S* and 𝒟_*W*. The dominance of China, the United States, and India, together with the prominence of IEEE and remote-sensing venues, is observed across both datasets. The keyword structures also converge around learning-based visual perception, surveillance analytics, remote sensing, object detection, privacy, and security. These results indicate that the main bibliometric and thematic patterns are consistent across the two independently constructed datasets.

Observed discrepancies arise from database-specific differences in source coverage, keyword indexing, and Keyword Plus enrichment in WoS. These differences affect the number of retained terms and the granularity of clusters, but they do not alter the dominant interpretation of the research domain. The cross-database agreement therefore, supports the robustness of the main bibliometric findings while preserving the methodological distinction between Scopus as the higher-granularity source and WoS as an independent validation source.

5 Analysis of prominent keywords and inter-cluster connectivity based on scopus

The keyword maps in this section are constructed from the Scopus co-occurrence network, considering 𝒟_*S* provides the most detailed keyword structure. The selected terms—deep learning, convolutional neural network, machine learning, computer vision, object detection, surveillance, and video surveillance—are emphasized because of their prominence within the corpus and their relevance to sensing, detection, behavioral analysis, biometrics, deployment, tracking, and security themes in the final 12-cluster representation. Transfer learning is discussed within the machine-learning subsection as an emerging learning paradigm within this field.

Unless otherwise stated, larger nodes in the VOSviewer maps shown in this Section indicate higher keyword occurrence, while thicker edges indicate stronger co-occurrence. The accompanying tables group the neighborhood (connected keywords) to each selected keyword by their clusters based on the Scopus clustering earlier presented in the previous section, after thesaurus-based normalization. These tables support thematic interpretation.

Before interpreting the keyword relationships, we would like to reiterate that, as explained in Section 3.7, selected terms were merged during thesaurus development. Several keywords correspond to variants, acronyms, plural forms, or version-specific labels of the same methodological family. Reporting these cases is aimed at explaining the prominence of consolidated nodes, such as YOLO, and clarifying how individual algorithm versions or naming variants contribute to the normalized co-occurrence network.

5.1 Deep learning

Figure 12 and Table 10 summarize the keyword neighborhood centered on deep learning. Deep learning is located in Cluster 8, while its associated or linked keywords span several other clusters. The largest associated theme relates to sensing and is represented mainly by Cluster 1. This group includes remote sensing, geospatial image analysis, image enhancement, fusion, restoration, super-resolution, semantic segmentation, agricultural monitoring, maritime surveillance, and robustness-oriented learning. This distribution indicates that deep learning plays a fundamental role in image analysis across diverse sensing sources.

Table 10. Cluster-neighborhood terms associated with Deep Learning.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

Behavioral Analysis is represented by Cluster 2, where deep learning is positioned near action recognition, activity recognition, anomaly detection, video anomaly detection, video surveillance, violence detection, recurrent models, autoencoders, attention mechanisms, and transformer-based video models. Detection and Biometrics appear through Clusters 3 and 4. These clusters include convolutional neural network, face detection, face recognition, transfer learning, object detection, YOLO, Faster R-CNN, SSD, instance segmentation, vehicle detection, license-plate detection, and traffic surveillance.

Deployment-, tracking-, and security-related terms appear across multiple clusters, particularly Clusters 5–11. Terms such as unmanned aerial vehicle, sensor fusion, cybersecurity, and edge computing, together with implementation frameworks such as OpenCV and TensorFlow, position deep learning within operational and system-oriented research themes. Thus, deep learning functions as a cross-cutting methodological theme linking sensing, temporal analysis, deployment, and security monitoring. Within the mapped literature, its cross-cluster position indicates that deep learning bridges model design and deployable surveillance functions by linking image-level representation learning with multi-sensor fusion, identity analysis, temporal modeling, and edge-oriented implementation.

5.2 Convolutional neural networks

Figure 13 and Table 11 summarize the keyword neighborhood centered on convolutional neural network. The map represents the CNN acronym family after thesaurus normalization. The primary placement of convolutional neural network is Cluster 3, which corresponds to Biometrics and identity-centric visual analytics. Associated terms include biometrics, face detection, face recognition, face mask detection, DenseNet, AlexNet, MobileNet, ResNet, VGG, support vector machine, histogram of oriented gradients, local binary pattern, principal component analysis, age estimation, gender recognition, and recognition. This placement indicates that convolutional models remain closely associated with recognition-oriented visual tasks.

Table 11. Cluster-neighborhood terms associated with Convolutional Neural Network.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

Behavioral Analysis appears through Cluster 2, where convolutional neural network is positioned near 3D CNN, LSTM, ConvLSTM, autoencoder, attention mechanism, transformer, vision transformer, activity recognition, human action recognition, video classification, video processing, video summarization, video surveillance, anomaly detection, and violence detection. Sensing appears through Cluster 1, which includes remote sensing, satellite imagery, Sentinel-2, semantic segmentation, U-Net, image enhancement, image recognition, feature fusion, random forest, crop classification, maritime surveillance, and ship detection.

The CNN-centered neighborhood is more recognition-centric than the broader deep learning node, with stronger proximity to biometric features, face analysis, and backbone architectures. Its links to unmanned aerial vehicle, edge computing, IoT, privacy protection, OpenCV, TensorFlow, and Siamese networks extend this recognition role toward deployment-, tracking-, and security-oriented settings across Clusters 5–12. Thus, CNNs remain a persistent architectural substrate for surveillance-oriented visual analytics, connecting identity recognition with temporal video analysis, geospatial interpretation, object-level perception, and edge deployment.

5.3 Machine learning and transfer learning

Figure 14 and Table 12 summarize the keyword neighborhood centered on machine learning. The primary placement of machine learning is Cluster 1, which corresponds to Sensing and geospatial image analysis. The associated terms include agriculture, crop yield prediction, random forest, and remote sensing. This placement indicates that machine learning remains a broad methodological term for remote-sensing and agricultural image-analysis tasks.

Table 12. Cluster-neighborhood terms associated with Machine Learning.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

Behavioral Analysis appears through Cluster 2, with activity recognition, anomaly detection, human activity recognition, video surveillance, violence detection, LSTM, and generative adversarial network. Biometrics and Recognition appear through Cluster 3, which includes biometrics, face detection, face recognition, feature extraction, support vector machine, and transfer learning. Detection and Deployment are represented by Clusters 4–6, where machine learning is positioned near CCTV, image processing, surveillance system, YOLO, artificial intelligence, edge computing, IoT, image classification, smart cities, and video analysis.

The machine-learning neighborhood is broader and less architecture-specific than the CNN-centered neighborhood. Its proximity to remote sensing and random forests in Cluster 1, biometric recognition in Cluster 3, and YOLO in Cluster 4, together with links to edge computing and smart cities in Cluster 6, positions machine learning as an broad methodological foundation. It underpins both classical classifiers and modern neural models across sensing, recognition, and deployment-oriented visual analytics.

Figure 15 and Table 13 summarize the keyword neighborhood centered on transfer learning.

Table 13. Cluster-neighborhood terms associated with Transfer Learning.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

In the Scopus network, transfer learning is located in Cluster 3, which corresponds to Biometrics and CNN-based visual recognition. Its associated terms include convolutional neural network, face recognition, VGG, and transfer learning. This placement indicates proximity to pre-trained visual backbones and recognition-oriented applications.

Detection and Sensing appear through Clusters 4 and 1, where transfer learning is positioned near object detection, YOLO, Faster R-CNN, deep neural network, machine learning, and remote sensing. Behavioral Analysis and Deployment appear through Clusters 2, 5, 7, 8, and 9, including anomaly detection, video surveillance, violence detection, artificial neural network, COVID-19, background subtraction, deep learning, and computer vision.

The transfer-learning neighborhood is narrower than the machine-learning neighborhood but more closely associated with domain adaptation. Its proximity to backbone architectures such as VGG and CNNs in Cluster 3 and specialized detectors such as YOLO and Faster R-CNN in Cluster 4 indicates that transfer learning is primarily used to reuse pretrained representations across surveillance tasks with limited labeled data. Thus, it functions as a technical bridge between general-purpose visual recognition and specialized application domains such as violence detection, remote sensing, and public-health monitoring.

5.4 Computer vision

Figure 16 and Table 14 summarize the keyword neighborhood centered on computer vision. In the Scopus network, computer vision is located in Cluster 9, which encompasses gait recognition, intelligent video surveillance, pattern recognition, and embedded or low-cost implementation terms such as Raspberry Pi. Its links to object detection, YOLO, face recognition, anomaly detection, edge computing, unmanned aerial vehicle, and person re-identification indicate that computer vision interconnects perception methods, surveillance applications, sensing platforms, and deployment contexts.

Table 14. Cluster-neighborhood terms associated with Computer Vision.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

Biometrics and Recognition appear through Cluster 3, where computer vision is positioned near convolutional neural network, biometrics, face detection, face recognition, feature extraction, MobileNet, and transfer learning. Behavioral Analysis appears through Cluster 2, with action recognition, activity recognition, anomaly detection, human action recognition, LSTM, optical flow, attention mechanism, transformer, video anomaly detection, and violence detection.

Detection and Security are represented by Clusters 4 and 5. These clusters include object detection, YOLO, Faster R-CNN, RCNN, real-time detection, vehicle detection, traffic surveillance, CCTV, surveillance system, detection, tracking, surveillance, security, behavior analysis, video analysis, and unmanned aerial vehicle. Deployment appears through Cluster 6 and Cluster 7, including smart cities, IoT, edge computing, federated learning, privacy protection, video analytics, OpenCV, TensorFlow, crowd counting, pose estimation, and weapon detection.

Tracking and re-identification are represented by Clusters 8, 10, and 11, where computer vision is positioned near deep learning, person re-identification, and multi-object tracking. In contrast to architecture-specific keywords, computer vision appears across biometric recognition, temporal video understanding, object detection, and smart-city deployment. Its neighborhood therefore captures the breadth of visual-intelligence applications within the Scopus co-occurrence structure.

5.5 Object detection

Figure 17 and Table 15 summarize the keyword neighborhood centered on object detection. The primary placement of object detection is Cluster 4, which corresponds to Detection, traffic surveillance, and intelligent transportation systems. Associated terms include YOLO, Faster R-CNN, SSD, RCNN, instance segmentation, real-time detection, CCTV, surveillance system, traffic monitoring, traffic surveillance, and license-plate recognition. This placement indicates that object detection links detector architectures with surveillance, traffic-monitoring, and real-time visual analytics themes within the Scopus co-occurrence network.

Table 15. Cluster-neighborhood terms associated with Object Detection.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

Sensing appears through Cluster 1, where object detection is positioned near remote sensing, remote sensing imagery, semantic segmentation, super-resolution, ship detection, deep neural network, and adversarial attack. Behavioral Analysis and Biometrics appear through Clusters 2 and 3, which include anomaly detection, attention mechanism, transformer, video surveillance, convolutional neural network, MobileNet, face recognition, and transfer learning.

Deployment, tracking, and security terms appear across Clusters 5–12, including surveillance, UAV sensing, smart-city deployment, and human detection. The object detection neighborhood links detector families and architectures, such as YOLO and Faster R-CNN, with traffic analytics, remote sensing, and edge deployment. This structure indicates that object detection connects visual perception algorithms with operational monitoring applications in the Scopus co-occurrence network. Its neighborhood links detector families and architectures, such as YOLO and Faster R-CNN, with traffic analytics, remote sensing, and edge deployment, positioning object detection as a practical interface between visual perception algorithms and operational monitoring tasks.

5.6 Surveillance

Figure 18 and Table 16 summarize the keyword neighborhood centered on surveillance. In the Scopus network, surveillance is located in Cluster 5, which corresponds to Security, UAV sensing, detection, and sensor fusion. The main associated terms include surveillance, detection, security, tracking, unmanned aerial vehicle, video, video analysis, classification, artificial neural network, and anomaly. This places surveillance at the heart of security-monitoring.

Table 16. Cluster-neighborhood terms associated with Surveillance.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

Behavioral Analysis appears through Cluster 2, with action recognition, anomaly detection, human activity recognition, LSTM, and violence detection. Detection and Biometrics appear through Clusters 3 and 4, which include a convolutional neural network, feature extraction, face recognition, CCTV, image processing, and object detection. Deployment and tracking appear through Clusters 6–9, including smart cities, real-time, deep learning, person re-identification, and computer vision.

The surveillance neighborhood is more system-oriented than the object-detection neighborhood. It brings together security monitoring, UAV sensing, temporal analysis, identity analytics, and smart-city deployment across multiple clusters, indicating that surveillance functions as an operational context rather than a single perception task. It represents the system-level environment where perception, tracking, sensor fusion, and security-oriented analytics are integrated for monitoring and situational awareness.

5.7 Video surveillance

Figure 19 and Table 17 summarize the keyword neighborhood centered on video surveillance. In the Scopus network, video surveillance is located in Cluster 2, which corresponds to Behavioral Analysis and video anomaly analysis. The primary associated terms include 3D CNN, LSTM, autoencoder, attention mechanism, generative adversarial network, recurrent neural network, vision transformer, action recognition, anomaly detection, video anomaly detection, violence detection, suspicious activity, video summarization, and optical flow. This placement indicates that video surveillance is closely associated with temporal video interpretation and event-level security analytics.

Table 17. Cluster-neighborhood terms associated with Video Surveillance.

Terms are mapped to the Scopus cluster representation after thesaurus normalization and indicate thematic proximity.

Biometrics and Detection appear through Clusters 3 and 4, with convolutional neural network, face detection, face recognition, VGG, CCTV, Faster R-CNN, image processing, and vehicle detection. Deployment, Tracking, and Security appear across Clusters 5–12, including edge computing, privacy protection, smart cities, person re-identification, vehicle re-identification, object tracking, event detection, pedestrian detection, and motion detection.

The video-surveillance neighborhood is closely associated with event- and sequence-level video analysis. Its links to action recognition, optical flow, recurrent models, anomaly detection, video summarization, and motion detection indicate an emphasis on dynamic event interpretation. Additional links to object detection, YOLO, re-identification, edge devices, edge computing, privacy protection, and smart cities connect this video-centered keyword neighborhood to deployment and security-oriented monitoring. This structure distinguishes video surveillance from broader surveillance terminology by emphasizing temporal analysis, identity-aware monitoring, and edge-enabled implementation.

5.8 Summary of keyword co-occurrence patterns

The selected-keyword analysis supports five observations. First, Detection is concentrated around object detection, YOLO, Faster R-CNN, SSD, RCNN, instance segmentation, traffic surveillance, vehicle detection, and license plate recognition. These terms are mainly located in Cluster 4, with tracking-related terms in Clusters 10 and 11.

Second, Behavioral Analysis is concentrated around video surveillance, anomaly detection, action recognition, activity recognition, video anomaly detection, violence detection, LSTM, autoencoder, transformer, vision transformer, video summarization, and optical flow. These terms are mainly located in Cluster 2.

Third, Biometrics and identity-centric visual analytics are represented by convolutional neural network, face detection, face recognition, face verification, FaceNet, MTCNN, biometrics, transfer learning, ResNet, VGG, MobileNet, support vector machine, histogram of oriented gradients, and local binary pattern. These terms are mainly located in Cluster 3.

Fourth, the sensing-related group is mainly concentrated in Cluster 1 and includes machine learning, remote sensing, remote sensing imagery, satellite imagery, SAR, Sentinel-2, semantic segmentation, U-Net, image enhancement, image fusion, super-resolution, precision agriculture, crop classification, crop mapping, winter wheat, and ship detection. Although deep learning is located in Cluster 8, its links to these terms indicate strong association with sensing-oriented image analysis.

Fifth, Deployment and Security terms span several clusters. Surveillance, unmanned aerial vehicle, sensor fusion, cybersecurity, intrusion detection, privacy, security, and real-time monitoring are concentrated in Cluster 5. Edge computing, IoT, federated learning, privacy protection, blockchain, smart cities, and video analytics are concentrated in Cluster 6. Crowd analytics, public-health monitoring, re-identification, foreground segmentation, tracking, and thermal motion analysis are represented by Clusters 7–12.

Overall, the selected-keyword analysis indicates that the Scopus network is organized around sensing, detection, behavioral analysis, biometrics, deployment, tracking, and security. These relationships support the interpretation of secure imaging as a multi-layered visual-intelligence domain in which learning models, sensing platforms, surveillance functions, and deployment constraints are jointly represented.

6 Review of real-world deployment systems and field-dataset studies

This section examines the papers selected from the datasets in Table 4 that have been categorized into surveillance, smart cities, remote sensing, video analytics, computer vision and image processing, and detection. Papers were selected using predefined criteria requiring the application of deep learning to secure imaging or video surveillance in smart-city, UAV, or remote sensing contexts. Only studies with real-world validation were included, specifically hardware implementations or evaluations on non-synthetic datasets. Simulation-only studies, purely theoretical works, and papers without an explicit security component were excluded.

6.1 Surveillance

Table 18 summarizes the screened surveillance studies across tiers, relating deep learning architectures to security objectives, validation settings, datasets, and reported performance within their application domains. The comparison covers UAV, body-worn, CCTV, maritime, and remote-sensing surveillance settings, with security objectives spanning privacy, confidentiality, integrity, and robustness. This subsection summarizes the selected studies and identifies the main lessons derived from their validation settings, security objectives, and reported outcomes.

Table 18. Comparison of surveillance-related studies, combining tier classification, deep learning architectures, security objectives, validation characteristics, datasets, and performance outcomes.

Recent surveillance systems are implemented using distributed or resource-constrained sensing platforms, where energy efficiency and secure data handling are primary design considerations in the reviewed systems. Unmanned aerial vehicles are increasingly used for monitoring and surveillance applications. Sarkar et al. (2021) introduces an Energy-aware Secure Internet of Drones (ESIoD) architecture that encrypts UAV-captured images using AES or RSA before securely offloading them to the cloud server for cognitive processing using Haar cascade and Faster R-CNN classifiers to reduce the computational burden on the drone. Two experiments performed using drone-captured sample images and videos show that the proposed architecture conserves up to 80% of onboard processing time and extends battery usage by three times compared with conventional onboard processing.

As surveillance technologies expand into wearable and mobile platforms, face recognition systems must balance accuracy with strict resource and privacy constraints. Face recognition is increasingly deployed in surveillance and law enforcement scenarios, including body-worn cameras. Almadan and Rattani (2021) investigate on-device face recognition for body-worn cameras by evaluating lightweight MobileNet-V2, EfficientNet-B0, LightCNN-9, and LightCNN-29 models on the BWCFace dataset. Their analysis compares these models with heavier architectures such as VGG-16 and ResNet-50, as well as six hand-crafted features, to assess the trade-off between performance and model size. Experimental results show that LightCNN-29 achieves performance close to ResNet-50 with significantly fewer parameters and faster inference time on mobile devices.

Beyond single-device deployment, collaborative learning has emerged as a key paradigm for privacy-sensitive surveillance scenarios involving distributed data ownership. Violence detection in surveillance footage is challenging due to its sudden occurrence and the need to protect sensitive visual information. Gautam et al. (2024) propose a federated learning–based framework for recognizing violent behavior in CCTV videos while preserving user privacy. The approach combines YOLO for extracting relevant frames from client-side footage with ResNet-50, allowing trained models from clients to be transferred to the server and aggregated into a global model without sharing raw video data. Experimental results show that the proposed method achieves an accuracy of 98.73% under uniformly distributed data and outperforms existing deep learning approaches.

Privacy challenges are further amplified in anomaly detection tasks, where large-scale video aggregation across organizations is often required. Semi-supervised video anomaly detection (SS-VAD) is a core task in intelligent surveillance, yet aggregating large-scale video data across organizations poses serious privacy concerns. Xie et al. (2025) propose DLPP, a distributed learning framework that applies federated learning to privacy-preserving SS-VAD. The framework explicitly tackles non-IID data and imbalanced data volumes by dynamically updating client models based on gradient divergence and by adaptively weighting client contributions during aggregation. Extensive experiments on benchmark datasets show that DLPP consistently outperforms baseline and state-of-the-art methods, achieving up to a 3.89% performance gain while improving communication efficiency by three times compared with FedAvg.

In parallel with privacy concerns, ensuring the integrity of surveillance video streams is addressed in edge–cloud deployments. Video tampering poses a serious threat to the integrity and security within Internet of Multimedia Things environments, especially under resource-constrained edge–cloud settings. Shao et al. (2025) propose MGMA-DSCNN, a deep learning framework that combines lightweight convolutional neural networks with multi-head attention mechanisms to enable real-time tampered video detection. The model enhances feature representation while maintaining computational efficiency, allowing effective identification of deepfake manipulations, frame insertion, splicing, and adversarial forgeries. Experimental evaluations on multiple forensic video datasets, including HTVD, show that the proposed approach achieves a detection accuracy of 98.1%. By leveraging edge–cloud collaboration, the framework reduces latency and energy consumption, making it suitable for real-time surveillance and forensic applications.

Surveillance systems also increasingly rely on remote sensing imagery, where secure storage and controlled access are essential in sensitive operational contexts. Cloud-based storage of satellite and aerial images introduces significant security concerns in domains such as border surveillance, monitoring, and reconnaissance. Praveen Kumar et al. (2025) present a comprehensive end-to-end security framework for the secure storage and retrieval of remote sensing images. The architecture follows a three-phase design comprising secure authentication, storage, and retrieval, combining Zero-Knowledge Proof and Quantum Key Distribution for user verification and quantum-based cryptography for image protection. An adversarially robust deep hashing model supports efficient indexing and secure retrieval, while watermark embedding enables insider threat detection and forensic tracing. Experimental evaluation shows a retrieval accuracy of 94.77%.

Secure distribution of sensitive remote sensing imagery is critical in surveillance contexts, where encrypted transmission may still expose communication patterns. Luo et al. (2024) propose DIH4RSID, a deep information hiding scheme that enables secure RSI distribution without revealing encrypted data flows. The method employs an encoder–decoder architecture with a parallel attention mechanism and adversarial training. Experimental results show improved visual quality and security compared with related approaches, and the extraction of hidden content does not degrade average classification accuracy.

Collaborative surveillance analytics have also been extended to maritime domains, where data sharing constraints and heterogeneous sensing modalities pose additional challenges. Maritime surveillance requires effective situational awareness while addressing challenges related to data sharing, privacy, and communication constraints. Tienin et al. (2024) introduce FedRS-Net, a federated deep learning framework designed for collaborative maritime monitoring using multi-modal remote sensing data. The framework enables decentralized training on SAR and optical satellite imagery without exposing client data, employing a communication-efficient federated averaging strategy and a redesigned skip-connection CNN architecture. Experimental results show a testing accuracy of 99.8%, while secure aggregation and momentum-based gradient compression reduce communication costs by 7%.

At finer spectral resolutions, secure access to hyperspectral imagery is crucial due to the sensitivity of spectral signatures in surveillance applications. The secure retrieval of hyperspectral images is critical in surveillance, where sensitive spectral information may be exposed during open-network sharing. Zhang et al. (2019) propose a hyperspectral image secure retrieval method that combines deep spectral–spatial feature extraction with encryption to protect image content during retrieval. Experimental results on hyperspectral datasets demonstrate that the method effectively preserves data confidentiality while maintaining reliable retrieval accuracy.

Finally, robustness against adversarial manipulation has become a central concern for detection models deployed in security-critical surveillance settings. Adversarial attacks pose a serious threat to remote sensing object detection systems used in surveillance and military reconnaissance, as imperceptible perturbations can significantly degrade detection performance. Ci et al. (2025) propose a multi-task, class-aware adversarial training framework that jointly optimizes classification, bounding box regression, and confidence prediction to improve robustness. Experimental results on the PASCAL VOC and DIOR datasets demonstrate improved resistance to both white-box and black-box attacks, with notable gains in mean Average Precision under PGD attacks while maintaining strong accuracy on clean data.

Overall, the surveyed surveillance studies demonstrate the use of decentralized and deployment-aware architectures that combine edge processing, federated learning, and lightweight deep models to achieve real-time performance under energy, bandwidth, and privacy constraints. At the same time, the integration of encryption, secure retrieval, information hiding, and adversarially robust training integrates mechanisms addressing integrity, confidentiality, and robustness as core requirements for modern surveillance systems operating across aerial, ground-based, maritime, and remote sensing platforms.

6.2 Smart cities

Table 19 summarizes the smart city–related studies across Tier 2, relating deep learning architectures to security objectives, validation settings, data, and reported performance within their application contexts. The comparison covers secure image retrieval, privacy-aware re-identification, cross-modal forgery detection, federated crowd analytics, and edge-based traffic monitoring in urban sensing environments. This subsection summarizes the selected studies and identifies the main lessons derived from their validation settings, security objectives, and reported outcomes.

Table 19. Comparison of smart city-related studies, combining tier classification, deep learning architectures, security objectives, validation settings, data, and performance outcomes.

The integration of remote sensing and edge computing within smart-city infrastructures introduces concerns related to secure access, privacy preservation, and trustworthy data retrieval. The growing use of remote sensing imagery in sensitive surveillance applications has raised concerns over image security, privacy, and reliable retrieval. Amaithi Rajan et al. (2024) proposes a privacy-preserving content-based image retrieval (PPCBIR) system that introduces a trusted edge computing layer to handle image encryption and feature extraction, thereby reducing user-side processing overload. The system employs MobileNetV2 for feature extraction to improve retrieval accuracy under high inter-class similarity. Trusted third-party auditing is incorporated to verify image integrity throughout storage and retrieval, and experimental results demonstrate a high mean average precision of 0.889, highlighting the system's effectiveness for secure and efficient remote sensing image retrieval.

Beyond sensing infrastructures, smart-city environments increasingly intersect with social platforms, where large-scale image sharing introduces additional privacy risks. The widespread dissemination of images through social media platforms has raised privacy concerns related to user surveillance and cloud security breaches. Ahmed et al. (2018) propose a privacy-preserving image-centric friend recommendation framework that protects user images by applying obfuscation before sharing. The framework classifies obfuscated images directly using deep neural networks, enabling profile construction for friend recommendations while preventing service providers from accessing the original image content.

Identity management in dense urban environments further amplifies privacy challenges, particularly when real-time reidentification is required across multiple cameras. Real-time identity reidentification in crowded surveillance environments, such as smart cities and social media platforms, remains a challenging task. Miraftabzadeh et al. (2017) presents a privacy-aware parameterized function that enables identity reidentification without explicitly determining true identities. The approach employs concurrent ResNet-based embedding pipelines at the camera level to generate facial feature vectors in real time, which are matched using multivariate kernel density estimation at the fog cloud for identity reidentification and security verification. The system maintains individual privacy while supporting effective reidentification, achieving improved accuracy compared with existing methods in real-time evaluations.

In addition to privacy risks, smart-city surveillance systems are vulnerable to attacks that compromise the integrity of live video streams. Fang et al. (2023) introduce WiSil, a vision–WiFi cross-modal system designed to detect video forgery attacks in surveillance environments where video streams may be tampered with or replaced. The approach reconstructs wavefront information of monitored objects from WiFi signals and recovers corresponding silhouettes using a deep learning network. A Siamese network then compares semantic features extracted from video silhouettes with WiFi-derived silhouettes to identify manipulated frames. Extensive experiments demonstrate that WiSil achieves up to 95% detection accuracy and remains robust under varying environmental conditions.

Digital image manipulation within urban environments also poses challenges for public safety, particularly when forged content is used to mislead surveillance or monitoring systems. Sedik et al. (2022) focus on copy–move forgery detection (CMFD) in digital image transmission, a common manipulation encountered in surveillance scenarios such as streets, railway stations, and festivals within smart cities. The study proposes a deep-learning-based CMFD approach employing CNN and ConvLSTM models to identify forged regions. The models are evaluated on the MICC-220, MICC-600, and MICC-2000 datasets, achieving detection accuracies of 95%, 73%, and 94%, respectively, with an overall accuracy of 85% on the combined dataset.

Smart-city surveillance increasingly relies on aggregate analytics, where privacy-preserving learning is required to limit exposure of sensitive visual data. Pang et al. (2023) address privacy concerns in smart surveillance by proposing a horizontal federated learning framework for crowd counting that avoids sharing raw visual data. The approach enables collaborative model training through aggregation while preserving private data on local devices, reducing communication costs and preventing data leakage. The authors further design four non-independent and identically distributed partitioning strategies and introduce an encoder–decoder e-FCN model tailored to the federated learning setting. Extensive evaluations establish a benchmark for future federated learning research in smart surveillance systems.

Finally, traffic monitoring represents a core smart-city application where efficiency and privacy must be balanced at scale. Leroux et al. (2022) investigate privacy-aware traffic counting under surveillance cameras by deploying deep neural networks directly on edge devices, where only aggregated counts are transmitted instead of raw video frames. The study introduces location-specific models trained for individual cameras, eliminating the need for cross-location generalization. An automated training procedure enables small models to be trained without human intervention. Experimental results show that these models achieve counting accuracy comparable to state-of-the-art approaches while using up to five times fewer parameters.

Taken together, the reviewed smart-city studies highlight and demonstrate the use of edge-assisted and federated analytics that enable secure image retrieval, identity management, crowd analysis, and traffic monitoring without exposing raw visual data. At the same time, the incorporation of obfuscation, cross-modal verification, forgery detection, and injection-attack defenses underscores the growing emphasis on integrated privacy-preserving mechanisms and integrity assurance as foundational requirements for trustworthy smart-city surveillance infrastructures.

6.3 Remote sensing

Table 20 summarizes the remote sensing–related studies across Tier 1 and Tier 2, relating deep learning architectures to security objectives, validation settings, data, and reported performance within their application contexts. The comparison covers on-board detection, cloud-based retrieval, federated recognition, watermarking, selective encryption, hyperspectral protection, and secure image hiding.

Table 20. Comparison of remote sensing-related studies, combining tier classification, deep learning architectures, security objectives, validation settings, data, and performance outcomes.

Remote sensing surveillance systems increasingly operate under strict power, latency, and reliability constraints, particularly in satellite and aerial platforms where onboard computation is limited. Small-object detection in satellite remote sensing images is challenging due to the small number of target pixels, complex backgrounds, and noise interference, while real-time performance is constrained by the high power consumption of GPU-based methods on small satellites. To address these issues, Fang et al. (2025) propose an FPGA-based hybrid overlapping acceleration architecture combined with a lightweight YOLOv5s-derived model for small-object detection in remote sensing images. The model integrates a GhostBottleneckV2 module to reduce parameters and computational complexity. Experiments on the Thermal Infrared Small-Object Dataset (TIFAD) dataset achieve a mean average precision of 67.8% with an average power consumption of 2.8 W, and robustness is further validated on the High-Resolution SAR Images Dataset (HRSID) dataset, demonstrating suitability for on-board remote sensing image processing systems.

Ensuring the trustworthiness of retrieval results is also important in cloud-based remote sensing services, particularly when sensitive imagery is involved. Privacy-preserving remote sensing image retrieval in cloud environments raises concerns related to retrieval accuracy and efficiency, image content security, process privacy, data distribution traceability, and retrieval result authenticity. In addressing these challenges, Hou et al. (2025) propose a secure and efficient retrieval framework that integrates CNN-based feature extraction with spectral hashing and clustering to improve retrieval accuracy and efficiency. Privacy protection is achieved through secure encrypted searchable indexes based on modified asymmetric scalar-product preserving encryption (ASPE), while image encryption is ensured via watermark embedding. The authenticity of retrieval results is further verified using a Merkle tree, and experimental evaluations demonstrate the effectiveness of the proposed method in achieving accurate, secure, and verifiable remote sensing image retrieval in cloud environments.

Distributed learning paradigms have recently been used in remote sensing, where data sharing is constrained by privacy and regulatory considerations. Zhang et al. (2023) examine privacy risks in federated learning for deep-learning-based remote sensing image recognition, particularly the vulnerability to white-box membership inference attacks that may expose sensitive information about model trainers. The authors propose Local Differential Privacy (LDP)-Fed, a federated learning framework that incorporates local differential privacy by pruning and perturbing uploaded model parameters. Different noise levels are applied across different layers to balance privacy preservation and model performance. Experiments on remote sensing and benchmark datasets confirm that the framework effectively defends against membership inference attacks while maintaining strong global model accuracy.

In addition to cryptographic protection, watermarking techniques are used for safeguarding the ownership and integrity of remote sensing imagery. Cao et al. (2025) propose a ring-template watermarking method for remote sensing images that improves robustness against geometric attacks and affine transformations. The method uses a convolutional neural network to automatically locate the annular watermark region, followed by ellipse fitting based on a least-squares approach to restore the watermark template after deformation. Watermark information is then analyzed from the binarized image by coordinate. Experimental results show that the approach maintains strong invisibility and robustness, achieving an Normalized Correlation (NC) mean value above 0.83 and Peak Signal-to-Noise Ratio (PSNR) values exceeding 35 under various geometric and affine attacks.

Selective encryption has emerged as an efficient alternative to full-image encryption, particularly for target-oriented remote sensing applications. Instead of encrypting entire remote sensing images, which increases computational complexity and reduces efficiency, Peng et al. (2024) introduce a selective encryption approach that focuses on encrypting only sensitive regions of an image. The method first applies an improved YOLOv8-based target detection algorithm to identify regions to be encrypted, followed by an encryption scheme combining Deoxyribonucleic Acid Sequence (DNA) coding, Logistic mapping, and the Chen system. Experimental results indicate strong encryption performance and resistance to attack, demonstrating suitability for target-detection-oriented remote sensing image protection.

Hyperspectral remote sensing introduces additional security challenges due to its high dimensionality and spectral redundancy. Hyperspectral images pose significant security challenges due to their large number of spectral bands, high variability, redundancy, and high encryption time. Bodke and Chaudhari (2025) present a selective hyperspectral image encryption scheme that integrates deep learning with hyperchaotic image encryption to enhance security and robustness. A deep neural network is used to increase the randomness of the hyperchaotic key sequence and improve encryption effectiveness. Performance evaluations on multiple hyperspectral datasets—including PAVIA, SALINAS, Indian Pines, Kennedy Space Center, and Cuprite—demonstrate strong encryption quality and resistance to brute-force and chosen-plaintext attacks.

Finally, covert data hiding techniques provide an additional layer of protection for secure remote sensing data dissemination. High-resolution remote sensing data require secure transmission while preserving image quality and robustness. Wang et al. (2025) propose a steganography network that embeds remote sensing images into natural images using an encoder–decoder architecture with independent encryption and decryption modules to reduce computational overhead and accelerate convergence. The model incorporates an enhanced Residual Convolutional Block Attention Module (Res-CBAM) and a Discrete Wavelet Transform (DWT)-based frequency-domain transformation to strengthen feature extraction and reduce the visibility of secret image contours. A dual-stage training strategy improves recovery accuracy under lossy conditions. Experiments on standard benchmarks and a custom dataset show superior PSNR and Structural similarity index measure (SSIM) compared with existing methods, while maintaining near-random steganalysis detection accuracy.

Overall, the reviewed remote sensing studies demonstrate the use of deployment-aware security architectures that jointly address onboard efficiency, cloud-based trust, and privacy-preserving collaboration across satellites, UAVs, and distributed platforms. By integrating lightweight deep learning with encryption, watermarking, federated learning, and covert data protection, modern remote sensing systems increasingly balance real-time operational constraints with stringent requirements for confidentiality, integrity, and verifiability.

6.4 Video analytics

Table 21 summarizes the video analytics–related studies across Tier 1 and Tier 2, relating deep learning architectures to security objectives, validation settings, data, and reported performance within their application contexts. The comparison covers edge-based video processing, facial privacy protection, traffic analytics, and forensic detection of video manipulation.

Table 21. Comparison of video analytics-related studies, combining tier classification, deep learning architectures, security objectives, validation settings, data, and performance outcomes.

Video analytics systems deployed in urban surveillance environments should be designed to satisfy real-time performance with stringent privacy requirements, particularly when sensitive demographic groups are involved. Privacy risks are particularly critical when closed-circuit television (CCTV) systems capture sensitive information about individuals, including minors. Yuan et al. (2020) propose a lightweight edge surveillance framework that enables real-time identification of minors in surveillance footage so that appropriate privacy-preserving measures can be applied. The approach modifies and cascades existing deep learning architectures to extract faces from video frames and classify them as adult or child using over 20,000 labeled samples. Designed for deployment within an edge–fog architecture, the system achieves near real-time performance on CPU devices. Experimental results demonstrate a classification accuracy of 92.1%, outperforming several face-recognition-based child detection methods.

Beyond privacy protection, ensuring the forensic integrity of surveillance video is important for legal, judicial, and evidentiary purposes. Sandhya and Kashyap (2024) focus on passive video forensics for detecting object-based forgery in surveillance footage. Their approach applies a motion residue windowing technique, enhanced with a max-averaging strategy, to highlight visual imprints of object-removal forgery. A lightweight depthwise-separable-layer-optimized CNN is then employed for efficient forgery detection, offering fast execution with reduced parameters. The model achieves 98.60% accuracy at the frame level and 99.01% accuracy at the video level, outperforming existing methods across compression rates in terms of detection accuracy, precision, recall, and F1-score.

Copy–move manipulation represents another common form of video tampering that threatens the reliability of surveillance footage in security-critical applications. Video fraud detection is a critical component in surveillance systems used for military, border security, and critical infrastructure protection. Eltoukhy et al. (2025) propose a convolutional neural network–based approach for detecting copy–move video forgeries, where segments are duplicated within the same video to conceal important content. The method decomposes videos into individual frames, extracts features using a pretrained CNN, and then trains a classification network to label frames as authentic or forged. Experimental evaluations on the Surrey University Library for Forensic Analysis (SULFA), Group on Research in Image Processing (GRIP), and Video Tampering Dataset (VTD) datasets demonstrate strong detection performance, achieving accuracies of 85.42%, 86.16%, and 81.87%, respectively, with relatively short processing time. The results show that the approach outperforms existing methods in both accuracy and efficiency, supporting its applicability for real-time video forgery detection.

In parallel with tamper detection, protecting facial identity within surveillance video streams has emerged as another privacy concern. Protecting facial privacy has become increasingly important as high-quality video surveillance has become widespread. Sivalakshmi et al. (2024) introduce a novel deep learning model (NDLM) designed to safeguard facial privacy in video surveillance. The framework operates in two stages, combining face detection with privacy protection, and integrates a Hybrid Convolutional Neural Network (HCNN) with the improved Fire Hawks algorithm (IFHA) for performance optimization. By embedding service quality preservation into the loss function, this method generates facial representations with controlled visual quality preservation. Evaluations demonstrate that the proposed approach is more effective than traditional privacy-preserving methods.

Bandwidth efficiency and data minimization are additional constraints in large-scale smart-city video analytics systems. Traffic surveillance generates large volumes of unstructured video data, thereby creating challenges related to bandwidth consumption and privacy protection. Skadins et al. (2020) address these issues by proposing an edge-based vehicle detection and tracking framework that performs real-time preprocessing directly on surveillance devices. The system transmits only a single high-resolution image per detected vehicle along its trajectory, significantly reducing bandwidth requirements while limiting the exposure of sensitive video data. Implemented and validated in a real city environment, the approach demonstrated up to a 24× speed improvement compared to state-of-the-art deep learning methods on a public benchmark dataset.

Taken together, recent advances in video analytics demonstrate the use of edge-centric and deployment-aware surveillance pipelines that jointly address real-time performance, privacy preservation, and forensic integrity. By combining lightweight deep learning models with privacy-by-design mechanisms and robust video forgery detection, modern systems increasingly enable scalable, trustworthy, and regulation-compliant surveillance across large urban camera networks.

6.5 Computer vision and image processing

Table 22 summarizes the computer vision and image processing–related studies, relating deep learning architectures to security objectives, validation settings, data, and reported performance within their application contexts. The comparison covers adversarial robustness, image watermarking, trusted data sharing, secure flood monitoring, and remote-sensing image classification.

Table 22. Comparison of computer vision and image processing-related studies, combining tier classification, deep learning architectures, security objectives, validation settings, data, and performance outcomes.

Advanced computer vision techniques are used in security-critical smart-city applications, particularly those requiring timely interpretation of large-scale satellite imagery. Urban flood monitoring demands rapid and reliable classification from satellite imagery, particularly for rescue missions. (Sathiyamoorthy and Subramanian, 2025) propose a secure deep learning framework based on DenseNet121 to distinguish urban flood regions using multispectral and SAR data. The model integrates adversarial training, homomorphic encryption, and federated learning, improving both data security and classification robustness. Experimental results show strong performance, achieving 96.8% accuracy with high precision, recall, and F1-score, outperforming existing deep learning architectures. The system is further deployed within a GIS web environment to support real-time monitoring and early warning, demonstrating its effectiveness for secure and efficient urban flood management.

Beyond classification tasks, ensuring the integrity and ownership of remote sensing imagery remains a challenge as image manipulation techniques become more sophisticated. Remote sensing images are increasingly exposed to tampering and unauthorized reuse, and conventional watermarking techniques are often weakened by deep learning–based removal methods. Xia et al. (2025) propose RSIAW, a robust watermarking framework that enhances robustness and authenticity by combining adversarial perturbations with style transfer. Adversarial perturbations are used to mislead watermark removal networks, while style transfer embeds the watermark in a visually coherent manner by aligning it with image textures. Experiments on the Northwestern Polytechnical University - REmote Sensing Image Scene Classification (NWPU-RESISC45) dataset show that Remote Sensing Image Adversarial Watermarking (RSIAW) maintains high visual quality and strong resistance to removal attacks, achieving superior PSNR, SSIM, and RMSE values compared with baseline methods.

Trustworthy perception is also essential in autonomous systems, where visual data often serve as forensic evidence in safety-critical scenarios. Autonomous driving systems require reliable supervision mechanisms to support accident liability determination and ensure mutual trust among stakeholders. Shi et al. (2024) propose a trusted supervision paradigm for autonomous driving (TSPAD) based on multimodal data authentication. The framework detects abnormal driving behavior using adaptive key-frame selection, enabling quantification of abnormal driving behavior. Selected frames are efficiently encoded and securely stored through a blockchain-based data on-chaining and certificate management strategy, ensuring tamper resistance and trustworthy data sharing among regulatory authorities, logistics platforms, and enterprises throughout the driving process.

Adversarial robustness has emerged as a concern in hyperspectral and remote sensing image analysis, where attacks may be difficult to detect visually. Deep neural networks used for hyperspectral image classification are highly susceptible to adversarial examples, particularly those generated by unknown attack strategies. Shi et al. (2023) introduce PD-Defense, a perturbation-disentanglement-based adversarial defense framework that improves robustness against unseen attacks. The method separates adversarial samples into attack-invariant features and perturbation features, performing defense on the attack-invariant component. Evaluations on the PaviaU and HoustonU 2018 datasets demonstrate that Perturbation-Disentanglement-Based Adversarial Defense (PD-Defense) consistently outperforms several state-of-the-art defense methods.

Complementary defense strategies have explored lightweight mechanisms that can be integrated into existing classification pipelines with minimal overhead. Adversarial attacks pose a serious threat to remote sensing image scene classification. Chen et al. (2021) observe that adversarial examples in Remote Sensing Image (RSI) classification exhibit attack selectivity, where misclassifications tend to occur toward specific classes. Building on this insight, they propose a soft-threshold defense strategy that distinguishes adversarial samples from genuine images based on class-specific confidence boundaries learned via logistic regression. The method operates without modifying the original model architecture and introduces minimal computational overhead. Experimental results show substantial reductions in fooling rates across multiple attack types, including FGSM, BIM, DeepFool, and C&W.

Generative modeling has also been leveraged to strengthen adversarial defenses by explicitly learning perturbation distributions. Deep convolutional neural networks used for remote sensing image scene classification remain highly vulnerable to adversarial examples, particularly those generated by unseen attacks. Cheng et al. (2021) propose Perturbation-Seeking Generative Adversarial Networks (PSGANs), a defense framework designed to enhance robustness against both known and unknown adversarial perturbations. The approach introduces a training strategy in which the classifier is trained on clean samples, adversarial samples, and reconstructed examples generated during perturbation modeling. By learning the distribution of adversarial perturbations and regulating perturbation diversity through a seeking radius parameter, the framework reduces defensive blind spots. Extensive experiments on multiple RSI scene classification benchmarks demonstrate that Perturbation-seeking Generative Adversarial Networks (PSGANs) significantly improves robustness under diverse attack scenarios.

While defense mechanisms aim to harden models, understanding the limits of robustness through adversarial attack modeling remains equally important for security evaluation. Multimodal remote sensing classification systems are increasingly exposed to adversarial threats, as deep neural networks remain vulnerable despite strong feature learning capabilities. Shi et al. (2022) examine the vulnerability of multimodal remote sensing classification networks and propose a Multifeature Collaborative Adversarial Network (MFCANet) to generate effective multimodal adversarial examples. The framework employs modality-specific generators to produce collaborative perturbations with strong attack capability, while discriminators guide generated samples to resemble real instances. Additional modality-specific generative and discriminative losses and an alternating optimization strategy further enhance attack effectiveness. Extensive experiments on the ISPRS Vaihingen 2D and Potsdam 2D datasets demonstrate that MFCANet achieves a stronger attack performance than classical methods such as Fast Gradient Sign Method (FGSM), Project Gradient Descent (PGD), and C&W.

Overall, recent work in computer vision and image processing demonstrates the use of accuracy-centric models as well as security-aware, robust, and trustworthy visual intelligence for smart-city and remote sensing applications. By integrating encryption, watermarking, federated learning, blockchain-backed authentication, and adversarial robustness into learning pipelines, these approaches help to improve the reliability, integrity, and accountability of vision systems operating in security-critical environments.

6.6 Detection

Table 23 summarizes the detection-related studies, relating deep learning architectures to security objectives, validation settings, data, and reported performance within their application contexts. The comparison covers privacy-preserving identity analysis, federated surveillance management, adversarial defense, UAV-oriented recognition, and LiDAR-based perception security.

Table 23. Comparison of detection-related studies for secure imaging and video surveillance, combining tier classification, deep learning architectures, security objectives, validation settings, data, and performance outcomes.

Detection tasks in surveillance and security systems must align with privacy-preserving requirements, particularly in biometric applications where sensitive personal data are processed. Privacy protection has become a major concern as facial recognition technologies are increasingly deployed in identity verification. Ghani et al. (2024) propose a privacy-enhanced detection framework that integrates generative modeling, clustering, and blockchain technologies to safeguard facial image data. This approach introduces a Privacy-Preserving Self-Attention GAN (PPSA-GAN) to generate realistic synthetic faces, which are subsequently anonymized using a mini-batch K-means clustering strategy to obscure identifiable attributes. Blockchain was incorporated to provide decentralized ledgers for secure data storage and auditing. Experimental evaluation on the CelebA dataset demonstrates strong generative quality and privacy guarantees with high Inception and Fréchet scores, whereas clustering and blockchain components further reinforce anonymization and privacy assurances.

Efficient deployment of detection models also requires efficient use of computational resources, particularly in continuous surveillance scenarios. Deep-learning-based video surveillance systems deliver high detection accuracy but impose substantial computational and memory demands, especially due to continuous GPU usage for object tracking and motion analysis. Ugli et al. (2024) present a cost-effective video surveillance management framework that combines edge computing, federated learning, and deep learning models to optimize resource utilization. The system adopts a two-tier edge architecture, where initial object detection is performed at the client-side edge close to the CCTV source, whereas a server-side edge dynamically manages GPU resources through a controlling threshold module. Federated learning is employed to train an LSTM model for predicting object appearances across multiple cameras while ensuring data privacy, and a Deep Q-Network dynamically adjusts threshold values to balance GPU memory usage and model reloading latency. The experimental results demonstrate improved efficiency in computational resource management for video surveillance systems.

In remote sensing detection, adversarial robustness has become a critical requirement as models are increasingly exposed to both digital and physical attack vectors. Adversarial examples pose a significant threat to detection models based on optical remote sensing imagery, affecting performance in both digital and physical attack scenarios. Xue et al. (2022) propose a cascaded adversarial defense framework specifically designed for aerial remote sensing detection. This method identifies adversarial patches by exploiting high-frequency and saliency information in the gradient domain and removes them directly. Image inpainting is then applied to recover the original semantic and texture information, while random erasing is incorporated to further enhance robustness. The experimental results demonstrate strong resistance to real-world adversarial attacks, with the AP₆₀ values of YOLOv3 and YOLOv4 experiencing only minor performance degradation under adversarial conditions.

UAV-based recognition systems face similar adversarial risks, particularly in safety-critical and mission-oriented deployments. Adversarial attacks pose serious risks to recognition systems deployed on UAVs, particularly in safety-related applications. Lu et al. (2023) address this challenge by proposing ensemble-based defenses that combine convolutional neural networks and transformers for both proactive and reactive adversarial protection. Proactive defense exploits weak adversarial transferability by combining output probability distributions from CNN and transformer models, whereas reactive defense integrates multiple adversarial detectors by fusing hidden features and averaged confidence scores from ResNet and Vision Transformer architectures. Experiments on optical and SAR remote sensing datasets show that the ensemble approaches outperform single models and detectors without requiring additional retraining. The study further introduces the adversarial robustness and performing defense on recognition models (AREP-RSIs) platform to support systematic evaluation of adversarial robustness and defense strategies in UAV-oriented remote sensing.

Finally, detection security in autonomous driving extends beyond cameras to LiDAR-based perception, where subtle perturbations can have severe safety implications. LiDAR-based 3D object detection is a core component of autonomous driving systems, yet it remains vulnerable to security risks that can compromise safety. Cai et al. (2025) introduce a LiDAR-based 3D adversarial attack that adaptively learns perturbations capable of imposing widespread hazards without targeting specific local point sets of genuine obstacles. The method maintains effectiveness, imperceptibility, and practicality in perturbed LiDAR point clouds. Extensive evaluations on the Waymo dataset demonstrate the effectiveness of the attack against multiple detection models, including CenterPoint, PV-RCNN, Dynamic Sparse Voxel Transformer (DSVT), and VoxelNext.

Overall, detection research shows the need for privacy-aware, resource-efficient, and adversarially robust frameworks that operate reliably across surveillance, remote sensing, and autonomous driving environments. By combining synthetic data generation, federated and distributed learning, edge-based resource management, and multi-level adversarial defenses, modern detection systems increasingly balance accuracy, privacy, and security under real-world deployment constraints.

7 Discussion

This section discusses design principles across application domains and derives implementation-level guidelines from the reviewed works in the previous section.

7.1 Design principles for secure visual intelligence under deployment constraints

Secure visual intelligence systems are designed under deployment constraints that directly determine model structure, data flow, and computation placement. Latency bounds, bandwidth limits, and data sensitivity motivate distributed architectures that reduce end-to-end delay while limiting raw data transmission, leading to edge, fog, and hybrid processing schemes (Amaithi Rajan et al., 2024; Ahmed et al., 2018; Miraftabzadeh et al., 2017). In these systems, computation is partitioned across nodes to satisfy real-time requirements while keeping data local. This is observed in edge-assisted secure retrieval (Amaithi Rajan et al., 2024), privacy-preserving inference on transformed inputs (Ahmed et al., 2018), and on-board FPGA-based processing for resource-constrained platforms (Fang et al., 2025). Deployment, therefore, serves as a primary design variable even before training.

Within these architectures, privacy is enforced at the level of representation and learning. Federated learning, local differential privacy, and encrypted embeddings enable collaborative training and inference without sharing raw data. Systems such as encrypted feature retrieval (Amaithi Rajan et al., 2024) and verifiable remote sensing retrieval using spectral hashing (Hou et al., 2025) show that feature-level protection can preserve utility while restricting data access. Edge-based anonymization methods, including autoencoder-based face obfuscation (Sivalakshmi et al., 2024), further reduce dependence on centralized data collection. As a result, privacy constraints directly influence feature design, model capacity, and training procedures.

Integrity is enforced through mechanisms that combine learning with verification. Watermarking, tamper detection, and traceable retrieval are integrated into processing pipelines to ensure authenticity and enable post hoc validation. Robust watermarking schemes (Cao et al., 2025; Xia et al., 2025) embed persistent identifiers, while forensic detection methods (Sandhya and Kashyap, 2024; Eltoukhy et al., 2025) identify manipulation at the frame level under operational conditions. These capabilities are complemented by verifiable retrieval frameworks (Hou et al., 2025) and trusted supervision systems (Shi et al., 2024), which ensure verifiability across distributed processing stages.

Robustness is an operational requirement as systems incorporate architectural and algorithmic defenses to maintain performance under environmental variation and adversarial inputs. Methods such as perturbation disentanglement (Shi et al., 2023), confidence-boundary filtering (Chen et al., 2021), and cascade defense modules (Xue et al., 2022) improve resilience with limited computational overhead. Evaluation increasingly includes physical threat models, such as LiDAR-based attacks (Cai et al., 2025), to capture deployment-relevant vulnerabilities. Resource-constrained implementations combine robustness with efficiency through edge inference and adaptive workload management, as demonstrated in UAV-based systems (Lu et al., 2023) and edge-deployed surveillance pipelines (Skadins et al., 2020).

In summary, these observations indicate that secure visual intelligence systems should consider multi-objective design considerations such as accuracy, latency, computational resources, and security risk. Systems combining federated learning, secure feature representations, and verification mechanisms demonstrate that performance and security can be balanced through coordinated design across sensing, representation, and inference stages. These design principles define the constraints under which practical systems operate. Their implementation-level requirements are provided next.

7.2 Implementation-level requirements

The reviewed systems show that deployment feasibility is determined by implementation constraints and does not depend on algorithmic performance alone. Practical designs must satisfy hardware limits, communication constraints, and security requirements simultaneously. The following implementation-level requirements are observed across multiple deployment settings:

• Hardware-aware design: Model architectures are adapted to platform constraints such as power, memory, and compute capability. On-board FPGA systems (Fang et al., 2025), two-tier edge architectures with dynamic GPU management (Ugli et al., 2024), and CPU-based edge–fog deployments (Yuan et al., 2020) demonstrate that inference pipelines must be co-designed with hardware to meet real-time and energy constraints.
• Optimized performance–security trade-offs: Systems balance accuracy, latency, and security overhead and often do not only optimize a single metric. FPGA-based implementations achieve 2.8 W average power with 67.8% mAP (Fang et al., 2025), while location-specific CNNs maintain comparable accuracy with 5× fewer parameters (Leroux et al., 2022).
• Hybrid resource management: Computation is distributed across edge, fog, and cloud layers to balance latency and security. Edge nodes perform preprocessing and inference to reduce data transmission, while cloud components support secure storage and retrieval. This is observed in edge-assisted retrieval with encryption (Amaithi Rajan et al., 2024), verifiable cloud retrieval using ASPE (Hou et al., 2025), and edge deployments achieving up to 24× speed improvement through reduced bandwidth usage (Skadins et al., 2020).
• Data minimization through local processing: Systems reduce transmission of raw visual data by performing feature extraction or inference at the edge. This is evident in edge-based traffic analytics that transmit only selected outputs instead of full video streams (Skadins et al., 2020), and in privacy-preserving retrieval frameworks that operate on encoded features and not on raw images (Amaithi Rajan et al., 2024).
• Model compression and lightweight design: Lightweight architectures are used to meet real-time and energy constraints without significant loss in accuracy. Examples include MobileNet-based edge retrieval systems (Amaithi Rajan et al., 2024), lightweight CNN deployments for traffic monitoring (Leroux et al., 2022), and FPGA-compatible detection models with reduced complexity (Fang et al., 2025).
• Distributed learning under communication constraints: Federated and distributed learning frameworks reduce the need for centralized data collection while introducing communication overhead that must be managed. Communication-efficient updates and aggregation strategies are demonstrated in federated surveillance systems (Gautam et al., 2024; Xie et al., 2025) and maritime monitoring frameworks (Tienin et al., 2024).
• Integrated security mechanisms: Security mechanisms are incorporated within processing or data representation stages. This includes encrypted feature representations in retrieval systems (Amaithi Rajan et al., 2024), verifiable indexing and traceability in cloud retrieval (Hou et al., 2025), and watermarking-based integrity protection for remote sensing imagery (Cao et al., 2025).
• Robustness to adversarial and environmental perturbations: Systems incorporate defenses against adversarial inputs while maintaining computational efficiency. This includes adversarially robust detection models (Ci et al., 2025), lightweight defense mechanisms for classification (Chen et al., 2021), and robustness-aware UAV recognition systems (Lu et al., 2023).

8 Future challenges and research directions

This section outlines open challenges and research directions for deploying deep learning in secure imaging and video surveillance systems within smart city environments. The discussion focuses on adversarial robustness, privacy-preserving computation, validation under real-world conditions, and the design of reliable learning-based systems under operational constraints.

8.1 Robustness and adversarial defense in urban sensing

Improving robustness in urban sensing requires addressing vulnerabilities of real-time detection systems to physical-world adversarial perturbations, particularly in traffic-related scenarios such as road signage. Detection architectures, including YOLO (v5/v8), remain susceptible to adversarial patches that can alter model predictions under varying environmental conditions.

Adversarial training strategies provide one approach to mitigating these effects. Mask-Reconstruction-assisted Adversarial Training (MRAT) introduces random masking during training to reduce overfitting to specific perturbation patterns (Yang H. et al., 2025). Similarly, Unified, a unified joint adversarial training framework incorporates variations in geometry, illumination, and viewing conditions to improve robustness under realistic scenarios (Huang et al., 2026). These approaches aim to improve generalization across perturbation types instead of optimizing for specific attack configurations.

Validation of such methods requires evaluation beyond static datasets. Simulation environments, such as the CARLA driving simulator, enable controlled variation of environmental factors including lighting, weather, and viewpoint (Nesti et al., 2022; Threet et al., 2021). Benchmark datasets such as APRICOT provide standardized evaluation protocols for physical adversarial attacks (Braunegg et al., 2020). However, the representativeness of simulated conditions relative to real-world deployment remains an open issue.

In addition to training-based defenses, modular approaches have been proposed. Separate detection modules can be used to identify adversarial patterns without modifying the primary detection pipeline (Jutras et al., 2022). Input-level transformations, such as Symmetric Image-Half Flip and Replace (SIHFR), attempt to reduce the effect of localized perturbations in recognition tasks (Siddiqui and Boukerche, 2021). Pre-processing techniques, including Fuzzy Image Transformation Pre-processing (FITP), aim to improve robustness at the input stage for video-based systems (Haque, 2024).

Architectural modifications also contribute to robustness. For example, models incorporating Channel Squeeze-and-Excitation mechanisms, such as YOLO-CSE, adjust feature weighting to emphasize consistent structural patterns in cluttered environments (Aiymbay et al., 2025). While these approaches show improvements under controlled conditions, their effectiveness under combined perturbations and real-world variability remains insufficiently characterized.

Overall, robustness in urban sensing systems depends on the interaction between training strategies, input processing, and model architecture. Future work should focus on evaluating these components under consistent and deployment-relevant conditions.

8.2 Privacy-preserving surveillance leveraging federated learning

Privacy preservation in surveillance systems requires limiting the exposure of raw visual data while maintaining analytical performance. FL, combined with edge computing, provides a distributed training paradigm in which model updates are shared instead of raw data. This approach reduces centralized data aggregation and aligns with constraints in surveillance environments where data sensitivity is high.

In human activity recognition (HAR) tasks, edge-based FL enables local processing of video streams, thereby reducing transmission of raw data. Additional mechanisms, such as Differential Privacy (DP), introduce controlled noise into model updates to limit information leakage during aggregation (Zhao, 2026; Zhang et al., 2025). Homomorphic encryption schemes, including Paillier-based methods, allow encrypted model parameters to be combined without exposing intermediate values (Tang and Wang, 2023). These techniques address different stages of the training pipeline but introduce trade-offs in computational overhead and convergence behavior.

Performance under privacy constraints remains dependent on system configuration. Reported results indicate that edge-based HAR systems can achieve high classification accuracy on benchmark datasets such as UCI HAR (Zhao, 2026; Mariappan et al., 2024). However, these results are typically obtained under controlled conditions with limited variability in sensing environments. The impact of non-IID data distributions, communication delays, and resource constraints on long-term performance remains an open issue.

From a systems perspective, edge-based processing reduces latency by performing inference close to the data source. Multi-tier architectures distribute computation across edge and cloud resources to balance responsiveness and computational load (Mariappan et al., 2024). Nevertheless, system scalability depends on coordination across heterogeneous devices and network conditions, which may affect model synchronization and update stability.

Additional mechanisms have been proposed to secure data exchange and storage. Blockchain-based frameworks provide a distributed ledger for recording transactions and model updates (Zhao, 2026; Jalali and Hongsong, 2025). While such approaches improve traceability, they introduce overhead in storage, communication, and consensus operations. The interaction between blockchain protocols and real-time constraints in surveillance systems requires further evaluation.

Lightweight encryption schemes have also been investigated for deployment on resource-constrained devices. Block ciphers such as SIMON, SPECK, and PRESENT provide reduced computational complexity suitable for embedded platforms (Zhao, 2026; Mariappan et al., 2024). Hardware implementations, including FPGA-based designs, enable real-time processing of encrypted data streams with improved energy efficiency compared to general-purpose processors (Bazarov Ravshan Ugli et al., 2024). However, integrating encryption, model training, and inference within a unified system introduces trade-offs among latency, throughput, computational overhead, and security guarantees.

Overall, privacy-preserving surveillance systems require coordinated design across learning algorithms, encryption mechanisms, and system architecture. Future research should quantify these trade-offs under realistic deployment conditions, including heterogeneous data distributions, communication constraints, and resource limitations.

8.3 Implementation using real-world datasets and simulations

8.3.1 Anomaly detection under real-world conditions

Anomaly detection in surveillance systems must account for variations in illumination, noise, and occlusion, which are common in real-world datasets such as UCF-Crime. Low-light conditions reduce contrast and degrade feature quality, affecting both spatial and temporal representations.

Preprocessing techniques, including pixel normalization and contrast enhancement methods such as Contrast Limited Adaptive Histogram Equalization (CLAHE), are used to improve visibility prior to feature extraction (Yuan et al., 2024; Bellegarda, 2023). These methods enhance local contrast but may also amplify noise, requiring careful parameter selection.

Spatio-temporal modeling approaches combine spatial feature extraction with temporal sequence learning. Hybrid architectures, such as 3D CNNs coupled with Bidirectional LSTMs, are used to capture both short-term spatial features and longer temporal dependencies (Yuan et al., 2024). Attention mechanisms are incorporated to prioritize regions with higher motion or semantic relevance (Zeb et al., 2025; Aranjuelo et al., 2021). However, their effectiveness depends on the quality of input features under degraded conditions.

Memory-based models, including ConvLSTM with spatiotemporal memory modules, retain feature representations over time and support differentiation between background variation and anomalous events (Wachter et al., 2026). Optical flow is often used alongside raw frames to provide motion-based features that remain informative when appearance cues are weak (Xiao et al., 2018). These approaches improve robustness to illumination changes but introduce additional computational overhead.

Reconstruction-based methods provide an alternative detection strategy. Autoencoder models are trained on normal patterns and detect anomalies based on reconstruction error (Aranjuelo et al., 2021; Xiao et al., 2018). This approach reduces dependence on labeled anomalies but is sensitive to distribution shifts between training and deployment data.

Deployment of these models requires balancing accuracy with computational constraints. Techniques such as knowledge distillation and model quantization reduce model size and inference latency (Xu, 2019). However, compression may affect sensitivity to subtle anomalies, particularly in low-quality video streams. Evaluation under realistic conditions, including variable lighting and camera noise, remains necessary for assessing deployment readiness.

8.3.2 Crowd monitoring with synthetic data

Crowd monitoring systems face limitations due to data scarcity and privacy constraints, particularly in aerial and thermal imaging scenarios. Synthetic data generation provides an alternative for training density estimation models by producing large-scale annotated datasets (Yuan et al., 2024; Bellegarda, 2023).

A key challenge is the domain gap between synthetic and real-world data. Differences in texture, noise, sensor characteristics, and viewpoint can reduce model performance when transferring to real environments. Domain adaptation techniques and mixed training strategies, where models are pre-trained on synthetic data and fine-tuned on real datasets, are commonly used to mitigate this issue (Xiao et al., 2018).

Thermal imaging offers advantages for aerial monitoring, as it reduces reliance on visible-spectrum features and enables detection under low-light conditions. However, thermal data introduces its own constraints, including lower spatial resolution and reduced texture information (Negied et al., 2016; Xu, 2019). Model performance depends on the alignment between training data and target sensor characteristics.

The effectiveness of synthetic training improves when simulation parameters match real deployment conditions, including camera perspective, resolution, and background distribution (Xiao et al., 2018; Delussu et al., 2020). Empirical results indicate that pretraining on synthetic datasets can support transfer to real-world scenarios, but performance depends on calibration and domain adaptation.

Overall, synthetic data provides a scalable training resource, but reliable deployment requires careful alignment between simulated and real-world conditions, as well as evaluation under domain shift.

8.4 Advanced architectures for trusted infrastructure

The design of reliable surveillance systems increasingly involves models that capture both local and global context, as well as the integration of multiple sensing modalities. This subsection examines Vision Transformers in maritime surveillance and multimodal sensor fusion combined with deep reinforcement learning (DRL) for autonomous navigation.

8.4.1 Vision transformers and convolutional models

In maritime surveillance, Vision Transformers (ViTs) and convolutional neural networks (CNNs) differ in how spatial information is represented. CNNs rely on local receptive fields and hierarchical feature extraction, while ViTs use self-attention mechanisms to model long-range dependencies and global context (Chami and Ajgou, 2025; Shahid and Hua, 2021).

ViTs are associated with improved modeling of spatial relationships across distant regions, which is relevant in scenarios where objects are sparsely distributed, such as maritime environments (Shianios et al., 2024; Jeeveswaran et al., 2022). In contrast, CNNs provide computational efficiency and are more commonly used in resource-constrained settings.

From an interpretability perspective, attention maps in ViTs provide a mechanism for visualizing the contribution of different regions to model outputs (Kasburg et al., 2025). However, such representations do not fully resolve issues related to model reliability or decision uncertainty. CNN-based models, while less directly interpretable, remain widely used due to their efficiency and established deployment in real-time systems (Ma et al., 2023; Korkmaz, 2025).

The selection between ViTs and CNNs is determined by trade-offs among computational cost, dataset scale, and the need for global context modeling. ViTs impose higher data and compute requirements, constraining their deployment in resource-limited edge environments. Hence, addressing these constraints requires efficient attention mechanisms, model compression, and data-efficient training.

8.4.2 Leveraging multimodal sensor fusion and reinforcement learning

Autonomous navigation in urban environments benefits from combining multiple sensing modalities. LiDAR provides geometric and depth information, while camera data contributes semantic and texture-based features (Lodhi et al., 2025; Carranza-García et al., 2022). Fusion of these modalities improves environmental representation compared to single-sensor approaches.

Challenges arise from incomplete or sparse sensor data. For example, LiDAR point clouds may contain gaps or low-density regions that reduce spatial completeness and degrade downstream perception tasks. Sparse-to-dense completion methods are commonly used to mitigate this limitation by estimating denser depth or point representations from sparse LiDAR measurements (Carranza-García et al., 2022). In addition, adaptive fusion frameworks, such as HydraFusion, adjust the contribution of each modality under varying environmental conditions (Malawade et al., 2022).

Deep reinforcement learning methods, including Deep Deterministic Policy Gradient (DDPG) and Deep Q-Networks (DQN), use sensor inputs to learn control policies under dynamic conditions (Khalil and Mouftah, 2023). The performance of these systems depends on the quality and consistency of fused inputs, as well as the stability of training.

Despite these advances, integrating perception outputs with decision-making and vehicle actuation remains complex. System performance depends on synchronization across sensors, robustness to environmental variability, and computational constraints. Evaluation under diverse traffic conditions, lighting variations, and sensor noise is required to assess reliability in real-world deployment.

8.5 Privacy-preserving and confidential visual analytics

Future surveillance systems should incorporate privacy protection throughout sensing, representation learning, storage, retrieval, and deployment rather than relying only on post-processing anonymization. Recent studies indicate that privacy risks can persist even when raw RGB images are not directly stored or transmitted. For example, event-camera streams remain vulnerable to reconstruction attacks, creating the need for event anonymization methods for privacy-preserving person re-identification and human pose estimation (Ahmad et al., 2024). Similarly, person re-identification features can be inverted to recover sensitive identity information, highlighting the need for feature-level privacy protection (Li et al., 2025a). Future systems should therefore evaluate privacy leakage across multiple layers and representations rather than assuming that privacy is guaranteed once raw images or videos are discarded.

Several recent directions attempt to reduce these risks through alternative sensing modalities and privacy-preserving visual transformations. LiDAR-based monitoring systems demonstrate that selected healthcare monitoring tasks can be performed while reducing exposure of identifiable visual information, since LiDAR captures geometric point-cloud data rather than conventional RGB imagery (Ju Y. et al., 2025). Seismometer-based pedestrian monitoring further shows the potential of non-visual sensing for privacy-aware urban mobility (Yoon et al., 2026). In camera-based systems, privacy preservation can also be integrated directly into the analytics pipeline through real-time anonymization (Asres et al., 2026). Variational autoencoder-based de-identification methods further indicate that privacy protection can be achieved while maintaining acceptable task utility in surveillance applications (Sivalakshmi et al., 2025). Future studies should quantify the trade-off among anonymization strength, task utility, latency, and resource cost for operational tasks such as crowd monitoring and traffic analysis.

Privacy-preserving visual analytics also requires secure storage, retrieval, and cloud-assisted inference mechanisms. Recent encrypted-domain approaches demonstrate that moving-object detection can be performed directly on encrypted compressed video streams without exposing raw content (Tian et al., 2024). Similarly, confidentiality-preserving video search frameworks based on protected Bloom-filter representations show potential for reducing the computational and communication overhead associated with secure retrieval systems (Yang X. et al., 2025). As cloud-assisted analytics expands, future research should emphasize scalable cryptographic integration, lightweight secure inference, and distributed privacy-preserving analytics frameworks that can be evaluated under real-time edge–cloud constraints.

8.6 Federated and adaptive surveillance under distributed data

Federated learning is emerging as a practical solution for privacy-aware surveillance because modern smart-city cameras, UAVs, and industrial IoVT nodes continuously generate distributed visual data that are often difficult or unsafe to centralize. Recent studies have explored federated and hierarchical federated learning for small-object detection, pedestrian fall recognition, person re-identification, and anomaly surveillance without direct raw-data sharing (Wu et al., 2024; Kim et al., 2025; Qamar et al., 2026; Chen X. et al., 2025; Alnajjar et al., 2024). Although these approaches reduce privacy risk and communication load, several deployment challenges remain unresolved, including non-IID data distributions, local dataset imbalance, annotation cost, concept drift, unstable client participation, and limited edge-device resources.

Future research should therefore prioritize adaptive and resource-efficient surveillance frameworks. Recent UAV-edge surveillance studies show that motion blur, occlusion, small targets, and strict real-time constraints continue to limit performance in practical deployment environments (Wu Y. et al., 2025). Similarly, recent research on real-world video anomaly detection highlights the need for periodic adaptation using streaming unlabeled data and evaluation under multicamera, multiday deployment conditions, rather than relying solely on static offline benchmark settings (Yao et al., 2026). Multi-view multi-human tracking systems also face persistent challenges related to cross-camera consistency and complex annotation requirements, increasing the importance of self-supervised learning and large-scale public benchmarks (Feng et al., 2025). Future federated surveillance systems should therefore be evaluated not only using task-specific metrics such as accuracy and mAP, but also using latency, communication cost, privacy leakage, robustness to domain shift, and long-term deployment stability.

8.7 Adversarial robustness under digital, physical, and multimodal attacks

Future secure visual-intelligence systems must evaluate adversarial robustness across threat models that extend beyond conventional image-level perturbations. Object detectors used in surveillance and smart-city systems remain vulnerable to white-box and hard-label black-box attacks, including attacks that suppress detections, reduce detector recall, or induce misclassification (Nguyen et al., 2025; Li J. et al., 2025). Recent work further shows that query-efficient black-box attacks can exploit surrogate-model features or sparse spatio-temporal perturbations to attack image classification, object detection, face recognition, and video recognition under limited model access (Li J. et al., 2025; Chen J. et al., 2025). These findings indicate that deployed surveillance systems should report robustness under multiple access assumptions, including white-box, black-box, closed-box, transfer-based, and query-limited settings.

Physical adversarial attacks require separate consideration because they operate under environmental and sensor constraints that are not represented by purely digital perturbations. Recent studies demonstrate that adversarial patches, camouflage, laser-based perturbations, and dynamic scene-aware attacks can significantly degrade object detection in autonomous-driving, UAV, and unstructured-road environments (Duan et al., 2025; Zhao et al., 2025; Hu et al., 2026; Xi et al., 2025). These attacks are particularly relevant to smart-city surveillance, where cameras, UAVs, and roadside sensors operate under varying viewpoints, illumination, altitude, occlusion, and weather conditions. Future evaluations should therefore include physical-world attack simulations and, where possible, controlled field testing rather than relying only on clean benchmark images or synthetic perturbations.

Multimodal and remote-sensing systems introduce additional robustness challenges. Point-cloud detectors, monocular 3D detectors, SAR classifiers, hyperspectral models, and segmentation networks are vulnerable to sparse voxel attacks, depth-aware perturbations, transferable ViT-based attacks, patch-size sensitivity, and annotation noise (Wu J. et al., 2025; Abraha et al., 2025; Ju X. et al., 2025; Zheng et al., 2025; Chandrappa et al., 2026; Kothari et al., 2026). These findings show that adversarial robustness must be evaluated at the data, representation, and annotation levels, especially in UAV and remote-sensing applications where small objects, domain shift, and multimodal fusion errors can directly influence operational decisions.

Defense mechanisms should therefore be treated as system-level components rather than isolated model improvements. Robust object detection must account for adversarial intensity variation and cross-model transfer (Cheng et al., 2025; Long et al., 2026). Transformer-based and remote-sensing models further require architecture-aware defenses, including virtual adversarial training and mutual-information-based representation learning (Shi and Li, 2025; Li L. et al., 2025). Deployment-oriented defenses also include low-latency physical-attack detection and patch-repair pipelines using segmentation and diffusion-based reconstruction (Chi and Msahli, 2025; Yi et al., 2025). Accordingly, future benchmarks should report not only clean and robust accuracy, but also attack success rate, latency, transferability, repair overhead, and performance under physical and multimodal perturbations.

8.8 Real-world object and event detection bottlenecks

Although recent surveillance frameworks have achieved strong detection accuracy, several challenges still limit reliable real-world deployment. Small-object detection remains a major issue in UAV, CCTV, and traffic-monitoring applications, particularly under adverse weather, poor illumination, dense backgrounds, occlusion, and motion blur. Recent studies address these limitations through multi-scale attention mechanisms, adaptive feature fusion, and thermal–visual integration (Dasu et al., 2026; Wang C. et al., 2026; Li et al., 2026; Tang et al., 2026). Nevertheless, robustness across heterogeneous environmental conditions remains inconsistent, indicating the need for evaluation strategies that extend beyond controlled benchmark datasets and report performance under operationally relevant sensing conditions.

Applications such as traffic accident detection, abandoned-object monitoring, wildfire smoke detection, near-miss assessment, and crowd anomaly recognition also require temporal reasoning beyond frame-level object detection (Nitilappool et al., 2026; Shah et al., 2026; Yang and Hong, 2026; An et al., 2026; Jung et al., 2026). Recent approaches increasingly incorporate trajectory analysis, persistence modeling, optical flow, and spatiotemporal learning to improve event understanding. However, reliable long-term temporal consistency, duplicate-alert suppression, severity estimation, and false-alarm control remain challenging in unconstrained surveillance environments. Future evaluations should therefore distinguish object-level detection accuracy from event-level validity and response timeliness.

Cross-scene generalization represents another persistent bottleneck because many models experience substantial performance degradation under unseen camera views, day–night transitions, dynamic backgrounds, and changing environmental conditions. To improve robustness, recent works explore domain adaptation, scene-invariant anomaly scoring, depth-aware modeling, and unsupervised background subtraction (Yang et al., 2026; Alawode and Javed, 2026; Mishra et al., 2026). Even so, current evaluations remain heavily dataset-dependent, highlighting the need for more realistic cross-camera, cross-weather, cross-site, and long-duration testing protocols.

Real-time deployment constraints further complicate large-scale surveillance adoption. Systems operating on UAVs, IoT devices, and edge platforms must satisfy strict requirements involving latency, bandwidth, storage, model size, memory, and power consumption. Recent studies indicate that lightweight detection architectures, pruning-based compression, efficient attention modules, and bitrate-aware video processing can reduce inference cost, processing time, or transmission bitrate while maintaining competitive detection performance or incurring only limited accuracy loss (Tang et al., 2026; Belememis et al., 2026; Wang Z. et al., 2026). Future benchmarks should therefore evaluate operational efficiency alongside detection accuracy, including inference speed, detection-to-alert delay, energy consumption, memory footprint, and communication overhead.

Finally, future intelligent surveillance systems must become more adaptive, scalable, and secure. Recent studies introduce class-incremental learning, reinforcement-guided adaptation, multimodal reasoning, and decentralized blockchain-based surveillance frameworks for auditability, data integrity, and access control (Raja et al., 2026; Sarar et al., 2026). However, maintaining stable adaptation under evolving environments while preserving privacy, scalability, and operational reliability remains an open research challenge.

Collectively, the reviewed challenges indicate that the development of secure visual intelligence systems depends on coordinated advances in robustness, privacy preservation, real-world validation, and system-level design. While individual techniques demonstrate effectiveness under controlled conditions, their integration under deployment constraints remains insufficiently characterized. Progress requires evaluation frameworks that jointly account for environmental variability, adversarial conditions, resource limitations, and data governance requirements. These factors determine the reliability of surveillance and sensing systems in real-world smart-city environments.

9 Conclusion

This study presented a dual-database bibliometric and science-mapping analysis of deep learning research for secure imaging and video surveillance in smart-city and remote sensing environments. Using independently constructed Scopus (𝒟_*S*) and Web of Science (𝒟_*W*) corpora covering 2010–early 2026, the analysis characterized relationships among computer vision, surveillance analytics, remote sensing, urban sensing systems, and security-oriented visual intelligence. Annual publication trend interpretation was restricted to complete calendar years from 2010–2025 to avoid bias from partial 2026 retrieval.

The Scopus keyword co-occurrence analysis retained 322 keywords at the minimum occurrence threshold of ten and produced 12 thematic clusters. These clusters separate the literature into remote sensing and geospatial image enhancement; video surveillance and anomaly analysis; CNN-based biometric recognition; real-time object detection and traffic surveillance; security surveillance, UAV sensing, and sensor fusion; edge/cloud smart-city security and privacy-preserving analytics; crowd analytics and public-health monitoring; re-identification and scene segmentation; computer vision, gait recognition, and event/attention modeling; tracking and surveillance calibration; multi-object/pedestrian tracking; and foreground/motion detection with human tracking and thermal imaging.

The results show that deep learning is used for automated visual tasks, including object detection, activity recognition, anomaly detection, re-identification, biometric recognition, semantic segmentation, and scene understanding. Bibliometric mapping indicates a transition from algorithm-focused development toward deployment-oriented secure visual intelligence systems, where models are integrated with distributed sensing, edge computing, UAV platforms, satellite imagery, smart cameras, and multimodal data sources to support scalable and low-latency operation.

Security, privacy, and trustworthiness are distributed across the keyword co-occurence network rather than confined to a single cluster. Privacy-preserving and infrastructure-oriented terms are concentrated around edge/cloud computing, IoT, blockchain, encryption, cybersecurity, federated learning, and smart-city monitoring. Model-level reliability is reflected through adversarial robustness, adversarial defense, robust learning, image restoration, and reconstruction-related methods. System-level security appears through surveillance, UAV sensing, radar, sensor fusion, public-safety monitoring, and edge implementation contexts.

The identified keyword clusters show the use of multiple imaging platforms, including fixed surveillance cameras, UAVs, satellite systems, edge devices, smart cameras, and distributed camera networks. These platforms produce heterogeneous visual data requiring scalable learning approaches for applications such as public safety, traffic monitoring, infrastructure inspection, maritime surveillance, disaster response, precision agriculture, and crowd monitoring.

Deep learning models remain susceptible to adversarial perturbations, domain shift, degraded imaging conditions, sensor heterogeneity, and data variability, which affect reliability in safety-critical applications. The expansion of sensing infrastructures also requires privacy-aware system design, secure data governance, trustworthy model validation, and practical deployment mechanisms that account for computational, communication, and regulatory constraints.

The cross-database validation indicates that the dominant bibliometric patterns identified from the Scopus dataset are broadly consistent with those observed in the independent WoS dataset. The publication trends, main thematic concentrations, and ranking patterns are therefore not dependent on a single bibliographic source. This consistency supports the reliability of the main findings while accounting for database-specific differences in indexing coverage, keyword extraction, and term weighting.

Future work requires coordinated development of distributed learning systems, privacy-preserving training methods, adversarially robust model designs, multimodal sensing architectures, and deployment-aware validation protocols. These directions are necessary to support reliable, privacy-aware, and trustworthy surveillance and imaging systems in smart-city and remote sensing environments.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript as an auxiliary tool to support the initial screening of large bibliographic datasets, assist in brainstorming possible keyword-normalization decisions, support the preliminary interpretation of patterns in bibliometric keyword analyses, improve the readability of initial drafts, and refine language. For tasks requiring structural alignment, such as the parallel presentation of Scopus and Web of Science (WoS) results, AI assisted with initial draft preparation, after which the text was reviewed, corrected, and refined by the author(s) based on the underlying data. Final paper selection, study validation, full-text review, thesaurus construction, keyword normalization, final normalization decisions, bibliometric interpretation, and all reported analyses were conducted and verified by the author(s), who take full responsibility for the content.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fimag.2026.1854187/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Hao Wang, Laoshan National Laboratory, China

Reviewed by: Xiaotong Li, Memorial University of Newfoundland, Canada

Weibo Zhang, China University of Petroleum, China

Huanyu Li, China University of Petroleum, China