Early prediction of learner outcomes in online learning environments can support timely academic intervention and learner support. This study proposes DKBNE-AlexNet, a hybrid architecture combining a Deep Kronecker Network branch, a feature-token batch-normalized eLU AlexNet-inspired branch, gated residual fusion, and a wide linear residual path. Using the Open University Learning Analytics Dataset, 24,153 enrolments from 22 module presentations and 22,234 learners were analyzed. Eighty-four predictors were constructed from learner characteristics, course design, and engagement during the first 25% of each presentation. The target was the weighted mean of tutor-marked and computer-marked assessment scores. Models were evaluated using five-fold cross-validation grouped by module presentation and repeated three times. DKBNE-AlexNet achieved an MAE of 11.424 (95% group-bootstrap CI 10.389-12.560), RMSE of 14.843, and R 2 of 0.117. Random Forest and LightGBM achieved slightly lower MAEs of 10.916 and 10.918, respectively, but the presentation-level difference between DKBNE-AlexNet and Random Forest was not statistically significant (Wilcoxon p = 0.388; Holm-adjusted p = 1.000). Distinct learning materials accessed and active learning days were among the strongest predictive contributors. DKBNE-AlexNet provides a competitive and stable framework for early learner-outcome prediction, while strong tree-based models remain important baselines. The framework is best interpreted as an early-warning learning-analytics tool rather than a direct measure of teacher effectiveness.
University networks are increasingly vulnerable to Distributed Denial-of-Service (DDoS) and Address Resolution Protocol (ARP) spoofing attacks due to open-access architectures, Bring Your Own Device (BYOD) practices, Internet of Things (IoT) integration, and dynamic traffic patterns. This study systematically reviews machine learning (ML)-based approaches for detecting these threats in university and related network environments. Following PRISMA 2020 guidelines, peer-reviewed studies published from January 2021 to January 2026 were identified through IEEE Xplore, SpringerLink, ScienceDirect, Taylor & Francis Online, and Web of Science, with the final database search completed on 31 January 2026. Of 400 identified records, 200 studies met the predefined eligibility and quality criteria and were included in the synthesis. The review examined supervised, unsupervised, semi-supervised, deep learning, hybrid, and ensemble approaches with respect to detection performance, validation strategies, deployment readiness, computational efficiency, adaptability, and explainability. The findings indicate that many reviewed studies reported strong detection performance across supervised, unsupervised, deep learning, hybrid, and ensemble approaches. Hybrid and ensemble models frequently demonstrated strong reported results; however, these findings were not directly comparable across studies because of differences in datasets, attack scenarios, feature sets, validation procedures, and experimental environments. Benchmark datasets were used in 48% of the reviewed studies, whereas only 12% evaluated real university networks. Real-time detection was assessed in 40% of studies, while continuous learning, concept-drift detection, and model adaptation received comparatively limited attention. The evidence does not establish universal superiority of any particular learning paradigm and suggests that reported performance depends substantially on the evaluation context and methodological conditions. These findings highlight the need for adaptive, context-aware, cross-layer, and explainable intrusion detection approaches capable of addressing the operational characteristics of university networks.
Infrared small target detection is critical in applications such as missile guidance, security surveillance, and precision tracking. Conventional methods relying on manual feature engineering often fail under severe background clutter and noise, while deep learning-based methods tend to degrade small-target features in deep networks, especially in noisy conditions. We propose a novel dual-branch architecture that combines Transformers and CNNs in parallel for robust infrared small target detection, focusing on low-signature scenarios. The CNN branch uses large-kernel convolutions to expand the receptive field and enhance local feature discrimination, while the Transformer branch captures long-range global dependencies. An Adaptive Fusion (AF) module with dual attention (channel and spatial) dynamically optimizes feature representation by amplifying target-specific features and suppressing background noise. Experiments show that our method achieves competitive performance against state-of-the-art models, with higher Dice coefficient and mean Intersection-over-Union (mIoU) scores on the augmented NUAA-SIRST dataset, indicating more precise segmentation boundaries and improved robustness in cluttered environments. These advantages highlight the potential of our approach for general infrared small target detection tasks. The dual-branch design and adaptive fusion strategy make it well suited to mission-critical applications such as security surveillance, and precision tracking, and it may also extend to remote sensing and maritime surveillance.
Virtual Exchange (VE) has become a key approach in higher education for fostering intercultural competence through online collaboration. However, despite its inclusive framing, many VE initiatives continue to rely on English as a lingua franca or other shared-language arrangements, which may reproduce linguistic hierarchies and unequal participation. This paper examines the potential of AI-supported multilingualism to rethink predominantly monolingual approaches to VE. It argues that while generative AI and translation technologies may reduce communicative barriers and enable broader participation, they do not automatically transform underlying language ideologies or power relations. In doing so, it highlights a tension between accessibility and the interactional work through which intercultural learning occurs. Drawing on Critical Virtual Exchange, relational learning, translanguaging theory, and ethical interculturality, the paper proposes a critical perspective on AI-mediated VE. It further explores how AI may support linguistic equity, multilingual participation, and translanguaging-oriented meaning-making in Virtual Exchange. At the same time, it critically addresses risks such as technological solutionism, linguistic bias in AI systems, and the reinforcement of dominant language ideologies. It concludes that AI-supported VE should be grounded in critical pedagogy and linguistic justice. In this way, VE is reframed not as a site of frictionless communication, but as a relational and ethically accountable educational practice.
Large language models are increasingly embedded in agentic software systems that retrieve information, plan subtasks, call tools, modify code, execute tests, and adapt through memory or feedback. This transition changes the safety problem: when a model acts through software tools, failure can alter external state. This Perspective argues that agentic AI safety should move beyond exclusive reliance on model trust, alignment, or prompt-level defenses and should also be treated as a software-architecture problem. We propose software topology as a design lens for structurally governed autonomy and develop a five-plane reference topology comprising intent, orchestration, execution, oversight, and adaptation. Its core invariant is that components that decide should not directly act, and components that act should not act without supervision. The topology is realized in Polos, an open reference specification that encodes role contracts, message schemas, declared flow graphs, bounded loops, governed self-improvement, and validator-checked invariants. We argue that this approach complements model alignment and empirical red teaming by providing an inspectable substrate for evaluating, constraining, and governing tool-using AI agents.
