Blockchain empowered federated learning for intelligent analytics in cattle livestock farming

1 Introduction

The global dairy industry faces significant challenges in ensuring consistent compliance with food safety and animal welfare standards. Current monitoring methods, which predominantly rely on manual inspections and periodic audits, are not only labour-intensive and costly but also susceptible to inconsistencies, delays, and potential record-keeping inaccuracies (Zhang X., 2024) ¹ . This underscores an urgent need for an automated, tamper-evident, and decentralized compliance mechanism that is accessible to farmers with limited technological resources while ensuring the integrity and privacy of sensitive operational data (World Bank Group, 2021).

Figure 1 compares bovine populations across countries, revealing a strong relationship between livestock availability and milk production capacity. Countries with large bovine populations, such as India and Brazil, maintain high production volumes through extensive systems (Backgrounders Press Information Bureau Government of India, 2026). Figure 2 presents a comparison of annual milk production among major producing countries, illustrating the global distribution of dairy output. India emerges as the largest milk producer, followed by the United States and China, indicating that both developing and developed economies play central roles in global dairy supply (Backgrounders Press Information Bureau Government of India, 2026; HLPE High Level Panel of Experts, 2022). The wide variation in production levels reflects differences in livestock population, productivity, farm structure, and technological adoption.

Figure 3 highlights the coexistence of organized and unorganized dairy production systems, using India as a representative case where the unorganized sector dominates (HLPE High Level Panel of Experts, 2022; Garro et al., 2025). Similar dual structures are observed in several developing and emerging economies, where smallholder farmers contribute a large share of milk output.

Federated Learning (FL) has recently gained prominence as a privacy enhancing distributed Machine Learning framework, enabling collaborative model training across decentralized data sources without centralizing raw data. However, conventional FL systems often depend on a central parameter server, creating a single point of failure and limiting transparency in the model aggregation process (Alghamdi et al., 2025). This architectural limitation can foster mistrust among participants and increase vulnerability to malicious activities such as model poisoning. To overcome these issues, the integration of Blockchain with FL has led to the development of Blockchain-empowered Federated Learning (BC-FL) systems. By capitalizing on Blockchain’s inherent decentralization, immutability, and transparency, BC-FL can securely record training contributions, manage participant reputations, and automate incentive distribution, thereby establishing a more accountable and trustworthy learning environment (Alghamdi et al., 2025; Zhang H. et al., 2024; Weng et al., 2019; Liu et al., 2023).

Despite these advances, many existing BC-FL implementations encounter practical barriers that restrict their adoption in agricultural and Internet of Things (IoT) based settings. These include high computational and communication overhead from traditional consensus protocols (Alghamdi et al., 2025; Xu and Chen, 2022), significant on-chain storage requirements (Mageswari, 2025), and a general lack of alignment with domain-specific regulatory workflows (Alghamdi et al., 2025; Jaberzadeh et al., 2023; Omheni et al., 2025). While several surveys provide comprehensive overviews of BC-FL architectures and challenges (Alghamdi et al., 2025; Myrzashova et al., 2023; Kang et al., 2022), few frameworks are designed to support end-to-end regulatory automation: particularly in sectors where real-time sensor data, edge intelligence and statutory compliance must converge seamlessly (Iqbal et al., 2025).

This research addresses this gap by introducing a tailored BC-FL framework for automated compliance monitoring in the dairy sector. The proposed system integrates several key components: (a) a lightweight edge-deployable machine learning model (CattleML) for real-time inference on animal health and milk quality; (b) a federated learning scheme enhanced with Graph Attention Network (GAT)-based dynamic clustering to handle data heterogeneity across farms (Li et al., 2022); and (c) a Layer-2 Blockchain (Polygon zkEVM) using Zero-Knowledge Succinct Non-Interactive Arguments of Knowledge (ZK-SNARKs) to generate and verify privacy-preserving compliance proofs. Smart contracts are deployed to automate identity management, audit logging, and interactions with regulatory authorities, thereby creating a closed-loop, self-regulating compliance ecosystem.

This research focuses on the core problem of the absence of a scalable, transparent, and resource-efficient system capable of automating regulatory oversight in agriculture while mitigating threats from unreliable or malicious participants. In response, the primary contributions are as follows:

• A lightweight BC-FL architecture optimized for deployment on resource-constrained IoT devices, centered on the CattleML runtime for edge inference.
• A transparent and secure FL aggregation mechanism that leverages Blockchain for reputation tracking and uses ZK-SNARKs to validate local updates without exposing raw data.
• A suite of regulatory smart contracts that autonomously execute compliance checks, license renewals, and audit triggers, bridging the gap between on-farm data and statutory requirements.

This architecture differentiates itself from prior BC-FL approaches through its vertical application focus and technical integration. Compared to general-purpose scalability solutions such as microchaining (Xu and Chen, 2022), the proposed framework adopts a Layer-2 zkEVM to minimize transaction costs and latency while preserving privacy through cryptographic proofs. Unlike reputation mechanisms that rely solely on historical performance (Jaberzadeh et al., 2023), the system employs GAT-driven clustering to form context-aware farm groups, enabling more equitable and personalized federated aggregation. Furthermore, while existing surveys often emphasize healthcare applications (Myrzashova et al., 2023), this design is grounded in agricultural IoT, combining multimodal sensing with automated regulatory enforcement to deliver a practical compliance instrument for the dairy industry.

Further, this article is organized as Section 2 represents the study about related work in literature, Section 3 presents the architecture of proposed framework, Section 4 elaborates the deployment procedure of smart contracts in the system, Section 5 discussed the implementation and results and finally the conclusion.

2 Related work

This section reviews the recent advances in Federated Learning (FL), BC-FL, trustworthy AI, and compliance automation. The review establishes trust deficits inherent in conventional FL, demonstrates how Blockchain integration addresses these shortcomings, and critically identifies a gap in research focused on end-to-end, automated regulatory compliance particularly for domain-specific applications like agriculture. A comparative analysis is provided in Tables 1, 2, mapping foundational works against the core features of transparency, trust, regulatory integration, and deployment practicality.

TABLE 1. Comparison of foundational surveys and frameworks in BC-FL.

TABLE 2. Alignment of proposed framework modules with related work and their role in automated compliance and license management.

2.1 Trust and security limitations in traditional FL

FL is designed to facilitate collaborative machine learning across distributed data sources while preserving data locality. However, its canonical architecture, which relies on a central server for model aggregation, introduces critical vulnerabilities that preclude its use in high-assurance, regulated environments (Zhang et al., 2025; Verma and Sharma, 2024). This server acts as a single point of failure and a central target for attacks, including model poisoning, Byzantine failures, and privacy inference attacks (Qammar et al., 2023). While algorithmic defences such as robust aggregation and differential privacy have been proposed to mitigate specific threats, they do not solve the underlying issue of centralized trust and the lack of verifiable auditability (Mothukuri et al., 2021).

A fundamental limitation of traditional FL is the absence of transparency and accountability in the training process. Participants cannot independently verify that their contributions have been incorporated faithfully or that the global model has not been manipulated. This “black-box” aggregation undermines trust in collaborative settings where participants may be competitors or operate under different regulatory jurisdictions. The need for an immutable, transparent record of all training transactions from local update submission to global model generation is therefore critical for applications requiring legal defensibility and compliance certification (Shi et al., 2026; Mageswari, 2025).

2.2 Blockchain-empowered federated learning (BC-FL)

To remediate the trust deficits of FL, the integration of Blockchain technology has emerged as a seminal research direction, giving rise to Blockchain-empowered Federated Learning (BC-FL). Blockchain provides a decentralized, immutable ledger capable of recording FL operations: such as model updates, aggregation results, and participant reputations, thereby establishing a verifiable and tamper-resistant audit trail (Alghamdi et al., 2025; Liu et al., 2023). Comprehensive surveys and literature reviews have systematically catalogued the architectures and benefits of BC-FL. These analyses consistently highlight Blockchain’s role in decentralizing the aggregation process to eliminate single points of failure, enabling transparent and fair incentive mechanisms through tokenized rewards, and creating a secure log for post-hoc security audits (Qu et al., 2022; Issa et al., 2023). For instance, survey works detail how smart contracts can automate client selection, model validation, and reward distribution, moving coordination logic from a proprietary server to transparent, code-based protocols (Alghamdi et al., 2025). Recent research advances the discourse by examining specific challenges such as scalability and resource efficiency. Proposals for lightweight BC-FL frameworks tailored for IoT environments emphasize the need for efficient consensus mechanisms and selective on-chain storage to manage the constraints of edge devices (Xu and Chen, 2022; Wankhede and Patel, 2025). Furthermore, the concept of trusted FL based on consortium Blockchains has been introduced, where a permissioned network of known entities (e.g., industry consortiums or regulatory bodies) provides a controlled yet transparent environment for collaborative learning, balancing performance with necessary oversight (Yin et al., 2025; Gupta and Tanwar, 2026).

2.3 BC-FL in healthcare and safety-critical domains

The healthcare sector, with its stringent data privacy regulations and distributed data landscape, has served as a primary testbed for BC-FL, providing a crucial precedent for compliance-sensitive applications. Research in this domain demonstrates how BC-FL enables multi-institutional training of diagnostic models without sharing sensitive patient data, while Blockchain maintains an immutable record of model provenance, data usage agreements, and audit events (Myrzashova et al., 2023; Bhasker et al., 2025). Frameworks integrating BC-FL with the Internet of Medical Things (IoMT) illustrate its utility in safety-critical monitoring. These systems use Blockchain to anchor hashes of sensor data and model inferences, ensuring data integrity and providing a verifiable chain of custody for clinical decisions (Almogadwy and Alqarafi, 2025; Swathi et al., 2025). Systematic reviews of intelligent frameworks for future healthcare systems confirm that the fusion of FL and Blockchain is considered foundational for building privacy-preserving, accountable, and collaborative AI ecosystems that can operate within complex regulatory frameworks like HIPAA and GDPR (Hassan et al., 2025). Authors in (Adebayo et al., 2025) proposed a Graph-Attention Federated Analytics method for dairy farm compliance prediction. While their work successfully applies graph attention to capture complex inter-farm dependencies, it remains reliant on centralized model aggregation for predictive analytics. The CattleML framework proposed in this research integrates Graph Attention Networks (GAT) with DBSCAN-based dynamic clustering directly at the edge.

2.4 Trustworthiness, transparency, and regulation

Beyond foundational security, the pursuit of trustworthy BC-FL encompasses broader requirements: algorithmic fairness, contributor accountability, and regulatory alignment. Blockchain’s immutable ledger provides a natural substrate for reputation systems, where smart contracts calculate node trust scores based on historical behaviour (e.g., model quality, latency, consistency). These scores transparently influence future participation and aggregation weights, promoting fair and reliable collaboration (Jaberzadeh et al., 2023; Javed et al., 2024). The integration of advanced cryptographic techniques with BC-FL enhances privacy-preserving verifiability, a cornerstone for regulatory compliance. Zero-Knowledge Proofs (ZKPs) and Secure Multi-Party Computation (SMPC) allow participants to prove the correctness of their local computations (e.g., that a model update correctly reflects their data) without revealing the underlying sensitive information. This enables external auditors or regulators to cryptographically verify process integrity without compromising data sovereignty (Bhutta et al., 2026; Nehal and Chinababu, 2025). Emerging research also addresses the challenge of machine unlearning and data-rights compliance within BC-FL. Solutions propose using Blockchain to immutably record data deletion requests and the subsequent model retraining steps, generating a verifiable certificate that “right to be forgotten” obligations have been fulfilled, a direct response to evolving data privacy laws (Zuo et al., 2025).

2.5 BC-FL for automated compliance monitoring and license renewal

Despite significant progress in using BC-FL to establish decentralized trust and security, a pronounced gap exists in its application for fully automated, domain-specific regulatory compliance and license management. Most existing frameworks employ Blockchain as a generic coordination or logging layer, lacking embedded, domain-aware logic to interpret real-time sensor data against regulatory statutes, trigger automated audits, or manage the dynamic lifecycle of licenses and certifications (Zhang X., 2024; Zhang H. et al., 2024). While some works point toward this potential such as those discussing Blockchain-based governance for collective decision-making in FL (Lee et al., 2023) or highlighting the importance of traceability for industrial audits (Al Asqah and Moulahi, 2023) they stop short of designing an integrated system where edge AI performs real-time compliance inference, FL collaboratively improves these models, cryptographic proofs validate outcomes, and smart contracts autonomously execute regulatory actions. A review of FL applications in livestock management notes the potential of these technologies but does not explore their synthesis for automated regulatory workflows (Garro et al., 2025). The present work directly addresses this gap. It synthesizes the decentralized trust of BC-FL, the verifiable privacy of ZKPs, and the automated execution of smart contracts into a cohesive framework explicitly designed for the regulatory workflow of the dairy industry. Addressing the critical need for data provenance, (van der Velden et al., 2025) introduced a verifiable dairy IoT compliance database tailored for privacy-preserving livestock analytics. While their architecture establishes a robust foundation for database-level verifiability, our proposed framework extends this concept by introducing an active enforcement layer. Specifically, our proposed research utilizes a closed-loop smart contract ecosystem where tokenized farm identities (Farm Non Fungible Tokens (FarmNFT)) and zero-knowledge proofs automatically trigger regulatory audits and license management.

2.6 Identified research gap from literature

The literature review demonstrates that while BC-FL research has matured in providing decentralized trust, transparency, and security (Alghamdi et al., 2025; Qammar et al., 2023; Qu et al., 2022), a significant gap persists in domain-specific, automated regulatory compliance. Existing systems offer foundational trust mechanisms but lack integration with real-time sensor inference, cryptographic proof generation, and smart contract-driven enforcement of statutory rules. No prior work (Qammar et al., 2023; Yin et al., 2025; Swathi et al., 2025; Zuo et al., 2025) combines lightweight edge AI, verifiable privacy-preserving proofs, and automated regulatory smart contracts into a single vertical solution. This work directly addresses this gap by proposing an integrated BC-FL framework for automated compliance monitoring and license management in the dairy industry.

3 Methodology

The methodology is explicitly designed to address three interrelated challenges: the absence of transparency and trust in collaborative learning environments, the difficulty of verifying regulatory compliance while preserving data privacy, and the lack of scalable mechanisms for enforcing compliance in resource-constrained agricultural IoT settings.

To this end, the proposed system integrates Federated Learning (FL) augmented with Graph Attention Network (GAT)–based dynamic clustering, a Layer-2 Blockchain infrastructure, and Zero-Knowledge Succinct Non-Interactive Arguments of Knowledge (ZK-SNARKs). This integration enables decentralized intelligence, cryptographically verifiable compliance, and automated regulatory enforcement without exposing sensitive farm-level data. The high-level architecture and workflow of the system are illustrated in Figure 4, while the internal learning and verification pipeline is detailed in Figure 5.

3.1 System architecture and entities

As shown in Figure 4, the proposed architecture follows a multi-layer decentralized design consisting of four interacting entity classes: (i) farm-level IoT nodes, (ii) federated aggregation nodes, (iii) a consortium Blockchain network, and (iv) external regulatory authorities. Each layer operates autonomously while interacting through cryptographically verifiable interfaces, ensuring scalability, fault tolerance, and regulatory transparency. At the farm layer, resource-constrained IoT devices continuously collect multimodal sensor data and perform local inference and training using a lightweight model. At the learning layer, federated learning enables collaborative model improvement across farms while preserving data locality. At the trust and governance layer, a Blockchain infrastructure records immutable compliance events and executes smart contracts that automate regulatory actions such as license renewal and audit triggering.

3.2 Edge layer: inference with cattleML

Farm IoT nodes are deployed on individual dairy farms and are typically implemented using low-power microcontroller ESP32. These nodes are responsible for continuous sensing, on-device intelligence, and cryptographic proof generation. Multimodal sensor data are collected from livestock and milking infrastructure, including body temperature, accelerometer-derived activity patterns, ambient ammonia concentration, and milk pH values. These signals provide complementary information regarding cattle health, hygiene conditions, and milk quality. Sensor readings are locally pre-processed to remove noise and compute statistical summaries before being consumed by the edge intelligence module.

Local inference is performed using CattleML, a lightweight machine learning runtime specifically optimized for microcontroller-class devices. The model performs real-time prediction tasks such as early disease detection, hygiene assessment, and milk quality compliance classification. Importantly, all inference and training operations are executed on-device, and raw sensor data are never transmitted off the farm, thereby preserving data sovereignty and privacy. In addition to inference, each node periodically updates its local model parameters using newly acquired data. These updates form the basis for federated learning and are later shared in encrypted form with aggregator nodes. The integration of multimodal data streams is essential for accurate compliance risk scoring, as demonstrated by Ramachandran et al. (2024), who effectively fused heterogeneous sensor data for aggregate risk assessment via federated analytics. Building upon similar data-driven principles, the CattleML framework differentiates itself by executing the multimodal fusion process (encompassing temperature, accelerometer features, ammonia, and pH) directly on resource-constrained microcontrollers at the edge.

3.3 Federated learning with GAT-based dynamic clustering

The system adopts a horizontal federated learning paradigm in which each farm i ∈ {1,…,} maintains a local dataset D i and trains a local model with parameters θ i. The objective is to learn a global model θ that minimizes the weighted empirical risk:where L i . denotes the local loss function at farm i. However, due to substantial heterogeneity in farm conditions, cattle breeds, and management practices, the data distributions D i are inherently non-IID. This significantly degrades the performance of standard FedAvg. To mitigate this issue, a GAT-based dynamic clustering mechanism, illustrated in Figure 2, is introduced.

min θ ∑ i = 1 N D i ∑ j D j L i θ ,

At the beginning of each FL round, each participating farm extracts a context feature vector c i, comprising non-sensitive metadata and statistical summaries such as average herd temperature, dominant activity patterns, farm size category, and historical compliance indicators. Raw sensor values are explicitly excluded.

An undirected graph G = (V, E) is constructed, where nodes V correspond to farms and edges E encode initial relationships based on geographical proximity or cooperative membership. Edge weights are initialized using similarity measures computed from the context vectors.

To capture latent similarities among farms, a Graph Attention Network is applied to refine node representations. For each node i, the refined embedding h i ′ is computed as:where a i j denotes the learned attention coefficient between nodes i and j, W is a trainable weight matrix, N(i) represents the neighbour set of node i, and σ(⋅) is a nonlinear activation function. This mechanism allows the system to prioritize updates from farms with similar yield changes and environmental codes. The refined embeddings are subsequently processed by a DBSCAN algorithm, which serves as a Robustness Guarantee by identifying and excluding outlier nodes (faulty sensors or malicious updates) from the aggregation process.

h i ′ = σ a i i W h i ∑ j ϵ N i ∪ i a i j W h j

Farms classified as outliers are temporarily excluded from aggregation, providing inherent resistance to noisy updates, faulty sensors, and model poisoning attacks. For each cluster C k, a cluster-specific federated averaging step is performed:where n i denotes the number of local samples at farm i and Δ θ i is the local model update. This clustered aggregation improves convergence and enables context-aware personalization.

Δ θ C k = ∑ i ∈ C k n i ∑ j ∈ C k n j Δ θ i

3.4 Privacy-preserving compliance verification using ZK-SNARKs

To ensure privacy-preserving verification without exposing sensitive farm vitals, the CattleML compliance logic is formalized as an arithmetic circuit using a Rank-1 Constraint System (R1CS). In this framework, the compliance verification task is decomposed into a set of quadratic constraints of the form L i · R i = O i, representing the mathematical gates of the circuit. The total count of 14,240 constraints is the result of a deliberate design trade-off between diagnostic depth and edge-level execution efficiency. The computational complexity of the circuit is categorized into three functional layers:

• Multidimensional Input Validation: The circuit validates the 16-parameter input vector from the Shahhet2812 ¹ dataset (including variables like rumination_time_hrs, milk_yield, and body_condition_score). Each parameter requires range-check constraints to ensure that the private witnesses are biologically plausible and fall within regulatory bounds before processing.
• Threshold-based Compliance Logic: This represents the core logic of the compute_compliance function. Each of the 8 regulatory rules (e.g., Temperature-Humidity Index, minimum water intake, and rumination duration) is implemented using bit-decomposition and comparison gadgets. Because ZK-circuits operate over large prime fields, simple inequality checks a b ≥ are computationally expensive, requiring multiple constraints per bit to prevent “overflow” and ensure proof reliability.
• Cryptographic Commitment and Hash Linking: To prevent “identity spoofing”, the circuit includes constraints for a Poseidon hash function. This links the private vitals to the public farm identity and the corresponding Federated Learning model version. This ensures that the generated proof $\pi$ is unique to a specific animal at a specific timestamp.

This high-density constraint optimization is specifically tuned for ARM-based gateways, the system achieves a Proof Generation Time of 1.25s, which is a critical benchmark for real-time scalability. This allows the Blockchain to verify compliance with O 1 complexity, meaning the on-chain cost remains constant regardless of the 16-parameter complexity of the underlying livestock analytics.

4 Blockchain-based compliance automation

Blockchain technology constitutes the trust, coordination, and enforcement layer of the proposed automated compliance monitoring framework. As illustrated in Figure 5, the Blockchain orchestrates the end-to-end lifecycle of compliance verification, violation detection, reputation update, and audit triggering in a fully deterministic and tamper-resistant manner. The framework is deployed on Polygon zkEVM, a Layer-2 Ethereum-compatible Blockchain, chosen for its low transaction cost, fast finality, and native support for zero-knowledge proof verification. Unlike conventional Blockchain-enabled FL systems that merely log model updates, the proposed design embeds regulatory compliance logic directly into smart contracts, enabling autonomous enforcement without reliance on centralized inspectors or manual audits.

4.1 Compliance proof submission and verification

As shown in Figure 5, the compliance workflow begins at the CattleML edge device, which performs on-device inference using multimodal sensor data. Once a compliance-relevant inference is produced (e.g., milk pH, hygiene score, or health indicator), the device generates a ZK-SNARK proof π attesting to the correctness of the inference with respect to predefined regulatory thresholds. In Step 1, the edge device submits the tuple t o k e n I d , y , π to the Blockchain, where t o k e n I d uniquely identifies the farm via its Farm NFT, and y denotes the declared compliance result. In Step 2, the validator contract verifies the cryptographic validity of the submitted proof. The verification process checks: (i) authenticity of the sender’s signature, (ii) integrity of the referenced model hash, and (iii) the correctness of the ZK-SNARK proof using an on-chain verification key. If verification succeeds, the system proceeds to Step 3, where the farm’s compliance status is updated on-chain by recording the compliance outcome along with a timestamp.

4.2 Violation detection and threshold evaluation

Following compliance status update, the system evaluates whether the submitted result violates regulatory constraints. In Step 4, the smart contract compares the reported outcome against predefined compliance thresholds τ. Formally, a violation is detected if:where M denotes the set of monitored compliance metrics. If no violation is detected, the workflow proceeds directly to reputation reinforcement. Otherwise, the system enters the audit-triggering phase.

∋ m ∈ M s . t . y m < τ m o r y m > τ m max

4.3 Audit triggering and external validation

When a violation threshold is exceeded, the smart contract automatically emits an AuditRequest event in Step 5, signaling the need for regulatory inspection. This event is observable by authorized auditors operating as permissioned Blockchain participants. In Step 6, the assigned auditor submits an audit result—either confirming or rejecting the violation, back to the Blockchain. The submission is cryptographically bound to the original Farm NFT, ensuring traceability and non-repudiation. Upon receipt of the audit result, the contract updates the on-chain audit log in Step 7, immutably recording the outcome of the inspection. This mechanism ensures that all regulatory actions are transparent, time-stamped, and verifiable.

4.4 Reputation update and incentive alignment

Reputation management is tightly integrated into the compliance workflow. In Step 8, if an audit confirms non-compliance, the farm’s reputation score is penalized according to predefined governance rules. Conversely, in Step 9, farms that consistently submit valid compliance proofs without violations receive positive reputation updates.

Let denote the reputation score at time t. The update rule can be expressed as:where α and β are system-defined reward and penalty coefficients. Reputation scores directly influence: (1) eligibility for future federated learning rounds, (2) weight assignment during model aggregation, and (3) automatic license renewal decisions.

R t + 1 = R t + α , if compliance verified and no violation R t + β , if violation confirmed

4.5 Automated license enforcement and regulatory oversight

License issuance and renewal are fully governed by smart contracts operating over the accumulated compliance and reputation history. When a farm maintains continuous compliance over a predefined window, its operational license is automatically renewed without manual intervention. Conversely, repeated violations or audit failures result in license suspension and mandatory inspection.

Regulatory authorities interact with the system as read-only or privileged participants, enabling them to query compliance histories, inspect audit trails, and update regulatory thresholds through governed on-chain transactions. This design ensures that policy changes remain transparent and auditable, preventing arbitrary enforcement.

5 Results

The CattleML framework was deployed and evaluated under realistic hardware and software configurations, reflecting the practical constraints found in rural dairy environments. The edge-level inference was implemented on two microcontroller platforms: the ESP32 (240 MHz, 520 KB SRAM) and the STM32H7 (550 MHz, 1 MB SRAM). Both platforms were integrated with sensors including a DS18B20 temperature sensor, an MPU6050 3-axis accelerometer operating at 100 Hz, an MQ-137 ammonia gas sensor for hygiene detection, and an Atlas Scientific pH probe for milk quality monitoring. These sensors were selected to simulate typical parameters monitored on dairy farms, with known tolerances and signal noise behaviors under environmental fluctuations.

On the software side, the inference model was a quantized TensorFlow Lite network, reduced to an 8-bit integer format occupying approximately 50 KB. The model was trained using FL with graph-based clustering (GAT) and DBSCAN for model fusion across farm participants. In addition to the inference system, a Blockchain layer was implemented using Polygon zkEVM as the Layer 2 infrastructure. Zero-knowledge proof integration was enabled using ZoKrates, allowing proof generation for compliance decisions made on-device.

To ensure the rigor and reproducibility of the experiments, the proposed framework was validated using the Shahhet2812 Cattle Health and Feeding dataset. The dataset comprises 54,200 samples distributed across diverse dairy cattle profiles, categorized into healthy, sub-clinical, and clinically non-compliant classes based on regulatory thresholds. For the federated simulation, data was partitioned across 5 simulated farm nodes, each configured with non-IID data distributions to mimic real-world operational heterogeneity. The multimodal fusion and compliance logic utilize 16 specific parameters, which are broadly grouped into four categories: 1) Environmental Metrics (ambient temperature, ambient humidity, temperature-humidity index, and ammonia concentration); 2) Physiological and Welfare Indicators (body temperature, body condition score, and hygiene score); 3) Activity and Behavioral Data (3-axis accelerometer data, daily steps, and rumination time); and 4) Production and Nutritional Metrics (feed and water intake, milk yield, and milk pH). These grouped parameters are directly ingested by the CattleML runtime for edge reasoning and ZK-SNARK cryptographic validation. The Architectural comparison of CattleML over other TinyML baselines are given in Table 3.

TABLE 3. Architectural and functional comparison.

The performance of CattleML was compared against three representative TinyML baselines: TensorFlow Lite Micro, Edge Impulse, and uTensor as presented in Table 4. Metrics assessed included inference accuracy, latency, memory usage, throughput, and power consumption per inference.

TABLE 4. Comparative results.

As depicted in Figure 6 CattleML achieved an average edge reasoning accuracy of 88.08%, exceeding the baseline frameworks by a margin of 7%–14%. TFLite Micro and Edge Impulse reached 78.5% and 82.1% respectively, while uTensor performed lowest at 75.3%. The improved accuracy of CattleML is attributed to the use of domain-specific input fusion, rather than general-purpose convolutional architectures.

Latency measurements illustrated in Figure 7 showed CattleML completing inference in approximately 35 milliseconds on the ESP32, with TFLite Micro and Edge Impulse requiring 58 and 62 milliseconds respectively. uTensor performed worst in this category, with latency averaging 70 milliseconds per inference. Memory usage in Figure 8 shows that CattleML, consuming 64 KB compared to 92 KB for TFLite Micro and 110 KB for Edge Impulse. Flash memory usage followed the same trend, with CattleML remaining under 100 kB, while other frameworks exceeded 120 KB.

Power consumption depicted in Figure 9 shows CattleML consumed approximately 80 milliwatts per inference, while Edge Impulse and TFLite Micro required 125 and 115 milliwatts respectively. This lower energy requirement makes CattleML more appropriate for battery-powered farm installations, particularly where continuous monitoring is required. CattleML was also the only framework among those evaluated to natively support FL and Blockchain-compatible output formatting. Its model output includes signed hashes and optional ZK-SNARK proofs, which were verified using smart contracts on the test Blockchain. Neither TFLite Micro nor Edge Impulse are designed to emit proof-generating inference outputs or operate in ZK-enabled systems without additional cryptographic wrapping.

The results suggest that CattleML maintains a consistent advantage across core metrics relevant to the use case. Its higher inference accuracy was not achieved at the cost of computational burden; rather, it demonstrated lower latency and energy usage than its general-purpose counterparts. While the advantage in memory and power efficiency may seem incremental, these differences are critical in long-term deployments where farm environments may lack stable power or high-capacity Micro Controller Units. In alignment with Sustainable Development Goal (SDG) 12 regarding sustainable production, evaluating the environmental footprint of IoT deployments is paramount. Wei et al. (2024) established a benchmark for green dairy compliance innovation, emphasizing energy minimization through TinyML. Compared to their baseline, the CattleML architecture exhibits highly competitive green computing characteristics. By compressing the communication payload to merely 4.25 KB per round and maintaining an average on-device power consumption of 82.659 mW, our framework proves that rigorous cryptographic compliance monitoring can be achieved without compromising the energy sustainability of rural IoT infrastructures.

5.1 Communication complexity system scalability and robustness

The proposed CattleML framework addresses the inherent constraints of rural IoT infrastructures by optimizing communication complexity to a marginal payload of 4.25 KB per update, achieving a 97.7% reduction in bandwidth overhead compared to standard TFLite Micro implementations and it is shown in Figure 10. This lean footprint is mathematically realized through a targeted 16-parameter multidimensional feature vector that minimizes gradient transmission requirements, enabling system scalability to over 500 concurrent nodes on a 1 Mbps gateway, a 40-fold improvement over conventional federated learning paradigms.

Scalability is further sustained at the architectural level by the integration of ZK-SNARKs, which utilize 14,240 R1CS constraints to generate succinct proofs that keep on-chain verification complexity constant at O 1, effectively decoupling the blockchain overhead from the increasing diagnostic depth of the livestock analytics. By maintaining a minimal 8.4 KB RAM footprint for the edge runtime, the system ensures high-performance intelligent analytics and Byzantine-resilient clustering on resource-constrained ARM Cortex-M microcontrollers without saturating the low-bandwidth NB-IoT or LoRaWAN networks common in decentralized dairy farming environments.

Robustness is analyzed by subjecting the CattleML edge nodes to varying intensities of Gaussian noise to simulate sensor drift and environmental interference common in rural dairy environments. The results demonstrate that the GAT-based dynamic clustering effectively serves as a Byzantine filter, maintaining high diagnostic accuracy by identifying and isolating statistical outliers before they can degrade the global federated model.

6 Conclusion

This article presents a novel framework that combines Federated Learning (FL), Blockchain technology, and Zero-Knowledge Proofs (ZKPs) to automate compliance monitoring and license renewal in the dairy industry. The system replaces traditional manual audits with real-time IoT data analysis (e.g., cattle health, milk quality) while ensuring privacy, scalability, and regulatory compliance. The framework introduces CattleML, a lightweight edge-deployable ML model trained via FL with GAT-based clustering and DBSCAN fusion, enabling collaborative learning without raw data sharing. It integrates a Layer 2 Blockchain (e.g., Polygon zkEVM) with ZK-SNARKs to validate compliance proofs privately. Smart contracts like FarmNFT contract and Validator contract automate license management, audit triggers, and reputation scoring, connecting farmers, veterinarians, and regulators (e.g., FSSAI) in a decentralized ecosystem. This research serves as an exploratory foundation for the future governance of food systems. This vision closely aligns with the paradigm of ‘zero-knowledge agricultural intelligence’ delineated by (Liu et al., 2025), who advocate for the convergence of TinyML, federated learning, and blockchain. The ZK-SNARK compliance certification scheme proposed in our framework acts as a practical realization of this paradigm. By mathematically proving that local edge inferences meet regulatory standards without exposing raw farm data, our scheme operationalizes the theoretical zero-knowledge concepts discussed by Liu et al., 2023 paving the way for scalable, privacy-preserving agricultural governance. This framework demonstrates how decentralized technologies can enhance agricultural compliance, offering a secure, efficient, and scalable solution that aligns with regulatory standards while preserving data privacy.

7 Future work

Future research will focus on advancing the framework through the integration of Asynchronous Federated Learning (AFL) to accommodate the intermittent connectivity inherent in remote dairy environments, ensuring model convergence is not hindered by individual node dropouts. To further harden the “trustworthy” ecosystem, we plan to implement hardware-rooted identities via Trusted Execution Environments (TEEs), providing a silicon-to-chain proof of integrity that prevents physical sensor tampering at the source. Additionally, the system will be expanded to include Local Differential Privacy (LDP) to provide a mathematically rigorous defence against membership inference attacks, effectively decoupling individual farm contributions from the global model gradients. These technical enhancements will be validated through a long-term field pilot in rural cooperatives to assess longitudinal scalability and the economic impact of automated compliance monitoring on small-scale dairy farmers.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

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.

Footnotes

Edited by: Yang Lu, Beijing Technology and Business University, China

Reviewed by: Jialin Liu, Beijing Technology and Business University, China

Dhiren Patel, Sardar Vallabhbhai National Institute of Technology Surat, India

Mai Hegazy, Mansoura Higher Institute for Engineering and Technology, Egypt

Footnotes

https://www.kaggle.com/datasets/shahhet2812/cattle-health-and-feeding-data