A federated blockchain framework for secure and intelligent smart farming in sustainable industrial agriculture Authors: Amar Y Jaffar Published: 2026 ABSTRACT AbstractThe fast growth of smart farming technologies and Internet of Things (IoT) sensors has transformed the agricultural sector, but has brought some fundamental problems, such as data privacy, scalability, and trust in distributed systems. In this paper, AgriChain-FL, a federated blockchain architecture, is introduced and guarantees privacy-preserving, transparent, and energy-efficient collaboration in agricultural ecosystems. The framework integrates federated learning (FL) of decentralized model training with a hybrid Practical Byzantine Fault Tolerance-Proof of Authority (PBFT-PoA) blockchain to allow aggregating data safely, auditing, and participation through incentives. The proposed system was developed based on the SmartFarm Sensor Dataset. Models are trained individually at each local farm node. The hybrid consensus protocol is more scalable, latency-reduced, and uses less energy. The experimental results indicated that AgriChain-FL was more accurate (92.4%), had a throughput of 2350 TPS, and a block latency of 1.0 s, which is also more favorable to analogous frameworks in learning and blockchain performance. Privacy leakage and energy cost were minimized by 73% and 45% respectively. AgriChain-FL is a successful balance between federated intelligence and blockchain trust that provides a privacy-conscious and sustainable smart farming platform. These findings confirm its possibilities as the basis of secure, decentralized, and energy-aware digital agricultural systems.Keywords: Federated Learning, Blockchain, Smart Agriculture, Privacy Preservation, Differential Privacy, Homomorphic Encryption, Energy Efficiency, Decentralized IntelligenceSubject terms: Engineering, Mathematics and computing Introduction The Internet of Things (IoT), artificial intelligence (AI), and advanced analytics have been instrumental in transforming the digitalization of agriculture and have greatly changed the current food production and supply-chain systems. AI-based predictive analytics applied to smart farming solutions can help farmers to optimize their irrigation, fertilizer use, and pest management, and boost their yield and sustainability with less environmental impact1,2. Nevertheless, these solutions rely very much on the further gathering and sharing of heterogeneous sensor, drone, and machinery information across distributed networks. The issue of data confidentiality, ownership, and traceability becomes an issue in such systems. The decentralized and immutable ledger of blockchain technology has reached the agriculture sector to secure the integrity and transparency of data and track the provenance of products throughout the production process3. With blockchain integrated with AI-powered analytics, the stakeholders will have an opportunity to verify agricultural dealings, certify sustainability assertions, and allow responsible data sharing without centralized intermediaries4. The recent industrial projects, such as traceable agri-food supply chains and cooperative farming models based on smart contracts, demonstrate the promise of blockchain to increase the degree of data integrity as well as sustainable resource management5. However, the majority of applications are still siloed, and the traceability of the supply chain, instead of the integration of farm-level intelligence, is pursued, and the full potential of blockchain-enabled AI is not fully exploited. Despite the considerable progress made by smart agriculture platforms in the areas of automation and analytics, there are still significant challenges to ensuring multi-stakeholder cooperation. The farm data (soil conditions, weather parameters, and crop health indicators) are frequently shared among the self-sufficient sensors and personal data. The cloud-based models are centralized, and this creates vulnerabilities such as leakage of data, unauthorized access, and single-point failures, which complicate the large-scale collaborative optimization6. Besides, mistrust between farmers, cooperatives, and regulatory organizations will inhibit the exchange of data to support AI-based decision support, which will limit the generalization and scalability of the models in the long run. Existing blockchain solutions in agriculture are mainly focused on transactional transparency or product traceability, and they do not provide much support in constant data-driven intelligence. At the same time, federated learning (FL) has become a distributed paradigm allowing model training on the nodes that are decentralized without transmitting raw data7. Nevertheless, the conventional FL models do not have the trust validation, auditability, and incentive alignment that are paramount in inter-farm cooperation in an ecosystem of industrial scale. Integrating blockchain and FL will eliminate these limitations by uniting decentralized training with unalterable traceability, but robust structures designed to address the specifics of agricultural data, including unbalanced sampling of data, sensor noise, and uneven farm conditions, are not abundant8. Thus, there is a significant research gap in creating a federated blockchain system ensuring privacy-conserving, auditable, and incentive-congruent collaboration to achieve sustainable agricultural intelligence. To overcome these shortcomings, this paper develops a Federated Blockchain Framework of Smart Agriculture (AgriChain-FL) that will merge the decentralization of blockchain and privacy-sensitive smarts of federated learning. Such integration is driven by the necessity of facilitating secure data collaboration without infringing ownership, as well as keeping the models in real-time as distributed agricultural nodes9. Blockchain is used as a trust and coordination layer, writing model updates and rewards in smart contracts, while FL makes sure that farm data is not stored in a central database, but instead is localized in each node. With this architecture, participation may be achieved in a transparent manner, hotels may not be compromised, and aggregation and incentive-based cooperation among farms, cooperatives, and agribusiness companies are possible. Federated blockchain mechanisms are directly incorporated in industries, which directly translate into the objectives of sustainable and transparent production of agriculture, and the vision of which is Industry 5.0 and SDG 9 (Industry, Innovation, and Infrastructure)10. Together with the increase in accountability of the data, AgriChain-FL supports the increase in the reproducibility of AI models within the context of vast farming networks and increases its effectiveness, as well as offering an audit trail, energy-saving, and immutable audit trails. It is this feature of blockchain and federated learning that provides the basis for the potential of safe, decentralized, and smart agricultural systems, which may prompt the development of ethical data administration, functional efficiency, and endurance. The primary goal of this work is to design and implement AgriChain-FL, a federated blockchain system that will allow secure, transparent, and intelligent coordination of distributed agricultural settings. The model covers the most important concerns of privacy, scalability, and management of trust of smart farming networks. The following are some of the contributions made by this paper to research: A novel hybrid PBFT–PoA consensus protocol is developed to enhance scalability, reduce latency, and maintain fault-tolerant trust in decentralized agricultural networks. A novel hybrid PBFT–PoA consensus protocol is developed to enhance scalability, reduce latency, and maintain fault-tolerant trust in decentralized agricultural networks. An integrated differential and homomorphic encryption scheme is proposed to protect sensitive farm data during model training and aggregation without sacrificing computational efficiency. An integrated differential and homomorphic encryption scheme is proposed to protect sensitive farm data during model training and aggregation without sacrificing computational efficiency. A smart-contract logic is introduced for secure data validation, transparent audit trails, and automated reward allocation among participating farm nodes. A smart-contract logic is introduced for secure data validation, transparent audit trails, and automated reward allocation among participating farm nodes. The SmartFarm Sensor Dataset is utilized to evaluate the proposed framework’s effectiveness in improving model accuracy, privacy preservation, and energy efficiency under realistic agricultural workloads. The SmartFarm Sensor Dataset is utilized to evaluate the proposed framework’s effectiveness in improving model accuracy, privacy preservation, and energy efficiency under realistic agricultural workloads. Collectively, these contributions render AgriChain-FL a scalable, long-term model for privacy-conscious, decentralized agricultural intelligence and enable the development of Industry 5.0-based smart farming environments. The rest of the paper has been structured as follows. Section 2 provides the literature review of smart farming, federated learning, and blockchain. Section 3 outlines the system model, mathematical model, and security architecture of AgriChain-FL. The architecture, algorithms, and smart-contract mechanisms are presented in Sect. 4. Section 5 contains research results and performance assessment. Finally, Sect. 6 presents the research conclusion and outlines future research directions. Literature review and theoretical context Smart farming is a synthesis of intensive IoT sensing and AI pipelines to maximize irrigation, fertilization, and pest management, and reduce environmental externalities. Early syntheses identified the role of big data in agriculture in revolutionizing decision support and value chains, but also pointed to fragmentation, data silos, and governance gaps as the limiting factors to collaborative analytics across farms11,12. Recent surveys focus on secure telemetry, provenance, and information maintenance for long-term assets and harvests, enabling verifiable information exchange across multiple enterprise structures13,14. These findings support architectures that protect farm information at the source, ensure traceability, and even enable multi-party model improvement. Federated learning (FL) is an algorithm-based approach that trains models at nodes and stores raw data locally to minimize the risks of exposure associated with centralized aggregation15–18. Convergence and fairness are problematic in agriculture due to non-IID sampling, intermittent connectivity, and heterogeneous hardware. The field of practical deployments, accordingly, will need strong aggregation, adaptive client selection, and compression to constrained communication, besides maintaining accuracy in skewed sensor distributions9. Surveys emphasize the three strategies of differential privacy (DP), secure aggregation, and homomorphic encryption (HE) to counter membership inference and gradient leakage in a farm environment where sensitive operational information is available19. Blockchain offers agri-food systems the services of provenance, certification, and incentive design on tamper-evident ledgers and smart contract design as programmable15,20–23. Controlled access and deterministic finality can be implemented in controlled supply chains using authorized networks, such as Hyperledger Fabric24. The recent breakthrough in smart agriculture AI- and IoT-based cybersecurity has also highlighted the importance of ensuring that devices are properly authenticated and anomalies are identified in the distributed farming ecosystem. Alyahya et al.25 have discussed a tamper-resistant authentication architecture for IoT-enabled smart agriculture, focusing on lightweight cryptographic schemes for resource-constrained environments. The system described in26 was a hybrid of the DTLS-based secure communication system for the smart container environment and was more resilient to network-layer attacks. In the same vein, Sarwar et al.27 proposed a more efficient dynamic SBPSO-based anomaly detection system for IoT networks, and Sarwar et al.28 developed an intrusion detection system that employs AI-based threat modeling of IoT infrastructure. These studies justify the introduction of effective cybersecurity measures in AI-based agricultural machinery. To enable data-driven agriculture, ledgers can notarize model updates, automate rules of participation and incentive alignment, and maintain auditability. Integrating blockchain and FL shifts trust from a centralized coordinator to verifiable protocols, at the cost of increased message latency and complexity, which must be engineered to support field-scale applications9,14. Consensus design determines throughput, latency, and energy footprint. PBFT ensures finality with bounded Byzantine faults but incurs quadratic overhead in the number of validator messages; PoA reduces overhead by validating a fixed set of validators; hybrid PBFT-PoA pipelines attempt to trade off safety and performance24,29. According to industrial research, multi-chain partitioning, credit-grouped PBFT, and policy-aware endorsement can alleviate bottlenecks without compromising auditability14,19. In agricultural FL, optimal patterns are lightweight consensus in steady-state learning, with stronger finality of milestone checkpoints, and combining DP/HE at the update layer to contain leakage without derailing convergence9,15,19. Recent literature also shows the increased convergence of blockchain, federated learning, and secure distributed intelligence in new IoT ecosystems. Ann et al.30 proposed a safe and effective renewable energy sharing model among distributed prosumers, demonstrating how blockchain can organize decentralized agents and maintain trust and transactional integrity. Dong et al.31 proposed using blockchain to protect distributed edge operations and to achieve a privacy-aware, blockchain-based task-distribution architecture in UAV communication systems. BDFL was introduced by Dong et al.32, a framework for federated learning built on blockchain to deliver smart UAVs at the edge, enabling secure model aggregation and decentralized trust management. Table 1 present the comparative summary of blockchain federated learning frameworks in agriculture and industrial IoT. Table 1Comparative summary of blockchain–federated learning frameworks in agriculture and industrial IoT domains.Study [Ref]Domain/DatasetLedger & ConsensusFL/Privacy MechanismNotable FindingsHyperledger Fabric Design24Industrial traceability (testbeds)Fabric (PBFT-style ordering)FL-agnostic; chaincode policiesDeterministic finality; performance sensitive to endorsement policy.PBFT Original (OSDI)29Generic BFTPBFT—Foundational BFT with quadratic messages; basis for permissioned ledgers.Blockchain-Enabled Smart Ag Review33Agriculture (survey)Permissioned variantsDP, access controlArticulates on-farm governance and audit needs for data sharing.Blockchain + Industrial IoT (TII)14IIoT/agri use-casesFabric/consortiumSecure agg., DPLedgered updates improve trust with modest overhead.FL Privacy Survey (IoT-J)19Cross-domain—DP, HE, secure aggregationTaxonomy of privacy–utility trade-offs for FL deployments.FL System at Scale (MLSys)17Cross-domain—FedAvg, secure agg.Practical FL orchestration, compression, and client selection.Big-Data Smart Farming Review11Agriculture——Identifies integration gaps across farm data silos.Agri Traceability (Food Chain)21Agri-foodPublic/perm.RFID + ledgerEnd-to-end provenance; limited ML integration.IoT + Smart Contracts (Access)23IoT/AgriPoA/PoS variantsAccess control, tokensProgrammable incentives; energy/latency depend on the validator set.FL for IIoT (IoT-Mag)12IIoT—Robust agg., client samplingPractical recipes for non-IID and intermittent clients. Comparative summary of blockchain–federated learning frameworks in agriculture and industrial IoT domains. Conceptual and theoretical foundations System model and operational design of agriChain-FL The conceptual design of AgriChain-FL centers on the interaction between distributed farm nodes, a blockchain coordination layer, and a federated learning server. The farm nodes will be mini data silos, with their own IoT sensor networks to gather soil moisture, humidity, temperature, and crop production data. Each node also does not transmit raw data; instead, it trains a local model using a local dataset, and only encrypted model parameters are sent to the blockchain layer. The blockchain serves as a trust and audit layer that maintains an immutable record of transactions, model updates, version control, and rewards. The hybrid model is a PBFT-PoA consensus system used to validate updates in practice, thereby maintaining high block finality speed and low-latency performance even under high network workload. Smart contracts help regulate node admission, verify updates, and reward participants, while providing automated control over the federated learning process. The global aggregator is the federated server: it receives validated changes to the blockchain, secures the aggregation, and re-notifies the participating farm nodes of the recalculated global model. The design ensures that the collaboration among heterogeneous agricultural entities is privacy-preserving while also guaranteeing complete transparency and traceability of all learning processes. Figure 1 represents the conceptual model of the system as decentralized farm nodes at the network edge, the blockchain ledger as the coordination backbone, and a federated aggregator in the center to carry out secure, consensus-driven model synthesis. System Model of AgriChain-FL showing decentralized farm nodes, blockchain ledger, and federated server interaction. The figure was manually created by the author using Wondershare EdrawMax, Version 12.6.1. Available at: https://www.edrawsoft.com/edraw-max/. Mathematical formulation of federated learning for crop data optimization The AgriChain-FL framework employs a decentralized federated optimization process in which each farm node trains a local model \:wusing its private dataset . The objective is to minimize the global empirical risk across all clients without exchanging raw data: where and each local objective is defined as: Each node performs local stochastic gradient descent (SGD) as: After each round \:t, the aggregator updates the global model using size-weighted FedAvg aggregation over participating clients : In which is the training sample size of the client \:i, and is the client model after \:E local epochs (or local update steps) initialized by . The local optimization process of each client takes place over \:E epochs of SGD with mini-batch size \:B, is a global learning rate (constant between clients) in most cases (where heterogeneous hardware may be supported by client-specific learning rates). To give the learning process theoretical rigor, we formulate typical assumptions under which federated optimization is performed, and then provide a convergence guarantee under these assumptions. Assumption A1 (Smoothness): Each local objective is \:L-smooth, i.e., Assumption A2 (Unbiased stochastic gradients): Each client computes an unbiased stochastic gradient with bounded variance: Assumption A3 (Bounded client drift/heterogeneity): The statistical heterogeneity across clients is bounded as: where and captures the degree of non-IID data. Theorem 1 (Convergence with privacy perturbations) Suppose that the AgriChain-FL is run with weighted aggregation (FedAvg) and \:E local steps per round and a global learning rate of . Assume that A1-A3 holds, injected differential privacy noise is zero-mean, and its per-coordinate variance is . Then, with the right selection of eta, the expected overall utility decreases, and the iterates approach a neighborhood of a stationary point that expands with respect to , , and contracts with the number of active clients per round. Proof sketch Within the parameter range A1-A3, the standard FedAvg descent inequality is represented by an upper bound on the smoothness-based one-step progress, and decomposing the error into (i) stochastic gradient variance , (ii) client heterogeneity drift , (iii) and privacy perturbation . The DP noise is zero-mean; thus, it does not alter the update, only improves the second moment; therefore, the method approaches an area that is proportional to . The more client involvement per round, the more stable it is, since it helps average out noise, which is consistent with our empirical results in Sect. 5 on privacy and accuracy. Corollary 1 (Formal convergence rate under non-IID data) Under Assumptions A1–A3, with a global learning rate , where denotes the number of clients selected per round and \:T the total number of communication rounds, AgriChain-FL satisfies: In which ) is the original objective, is the global optimum, \:E is the number of local SGD steps per round, is the client-heterogeneity (non-IID) parameter of Assumption A3, is the stochastic gradient variance of Assumption A2, and is the per-coordinate differential is the second term that explicitly measures the interaction between non-IID heterogeneity and local update depth: a larger (skewer client data) or larger local steps \:E each increases client drift and slows convergence. This finding indicates that convergence decreases quadratically with \:E when heterogeneity is present, which is why we have selected \:E=5 (see Sect. 4.6.3) and the deviation-score screening mechanism. Remark 1 (Impact of local epochs on convergence and client drift) The client-drift error bound demonstrates that the client-drift error term is of the order of , which means that the convergence error increases quadratically with the local training steps \:E when the data are non-IID. Naively, every local SGD step will bring the model of a client to its local optimum, as opposed to the global one; with Esteps, this drift will accumulate, creating aggregated updates that are biased off . In order to describe this empirically, we ran AgriChain-FL with under non-IID partitions and measured the convergence rounds and the ultimate accuracy. Table 4 indicates that \:E=5 is the most accurate with the least convergence trade-off: after \:E=10, the variance between the nodes grows exponentially (from 2.10% to 4.83%), and the number of convergence rounds increases by up to 28, as expected by theory. In the case of , the communication overhead becomes dominant, and there are no significant improvements in accuracy. As such, \:E=5 is held at the operating point in all experiments. Table 4Summary of training performance metrics (Accuracy, F1-score, Convergence rate, and Stability index) across consensus mechanisms.Consensus typeModel performance metricsTraining dynamics and system metricsGlobal Accuracy (%)F1-ScorePrecisionRecallAUCConvergence roundsMean variance across nodes (%)Training loss (Final)Validation loss (Final)Stability index (SI)Throughput (TPS)Latency (ms)PoA (Proof-of-Authority)89.730.8870.8740.9010.936643.420.1840.2120.902225098PBFT (Practical Byzantine Fault Tolerance)92.150.9110.9250.8970.948582.870.1690.1910.9311820143Hybrid PBFT–PoA (Proposed)95.840.9430.9510.9360.973452.100.1420.1560.948238591PoS (Proof-of-Stake)88.920.8720.8680.8770.925663.950.1960.2190.8932105106RAFT (Leader-based)91.340.9030.8960.9110.944522.740.1720.1880.927243585Centralized Baseline (No Blockchain)96.100.9510.9560.9470.978421.880.1390.1510.958N/AN/A Summary of training performance metrics (Accuracy, F1-score, Convergence rate, and Stability index) across consensus mechanisms. Remark 2 (Gradient variance: theoretical justification and empirical validation) Assumption A2 assumes that the mini-batch stochastic gradient of every client is unbiased and has a finite variance . This is the standard assumption of federated optimization, which is explained by two observations here. In the first step, every farm node takes independent mini-batches of size \:B=32 of its local dataset by the central limit theory, the sample mean of the gradient is asymptotically unbiased of the full gradient. Second, following Min-max normalization and PCA compression (Sect. 4.6.2), the range of features is bounded by , so the second moments of the loss gradient are also finite. We empirically estimated the roundwise by computing on a held-out batch and found a constant per-round variance of as expected by the convergence bound. Moreover, the Corollary 1 formula demonstrates that the gradient variance term is suppressed by the number of participating clients \:K and by the local dataset size \:N, which accounts for the better aggregation stability achieved by increasing the client participation in Fig. 4(e) and the low inter-node variance in Table 4 (2.10% with the proposed Hybrid PBFT-PoA scheme). Training and validation convergence curves of AgriChain-FL under different consensus settings. Remark 3 (Robustness of FedAvg aggregation to outliers and comparison with robust alternatives) One of those is plain FedAvg, a weighted averaging technique, which is said to be prone to outliers or to adversarial, skewed client contributions because a single large-norm update may prevail in the aggregate. AgriChain-FL manages this using the median-absolute-deviation (MAD) deviation-score filter of Sect. 4.3 (Eq. 21), which rejects updates that meet . We used three aggregation strategies on the non-IID SmartFarm partition: (ii) plain FedAvg, (iii) Krum aggregation [selecting the client update with the fewest sum-of-squared distances to its \:K-f-2 nearest neighbors], and (iii) coordinate-wise trimmed-mean [removing the top and bottom 10% of values in each parameter dimension and averaging them]. To determine residual, with a 20% Byzantine-client injection and sign-flip attacks, plain FedAvg accuracy reduced to 81.3% compared to 89.7% in trimmed-mean, 90.2% in Krum, and 91.8% in AgriChain-FL with MAD screening. The findings affirm that the MAD-based filter achieves an equivalent level of robustness to specialized robust aggregators while retaining the computational simplicity of weighted averaging in benign settings. The difference between the residual gap and Krum’s (1.6% points) is due to Krum maximizing the single most representative update, which is overly conservative in high-heterogeneity regimes; the MAD method is more accurate and robust in the agricultural non-IID regime. Remark 4 (Influence of noise injection on the convergence rate and model accuracy): Corollary 1 gives us the additional irreducible term to the convergence residual due to the DP noise. In particular, the additive Gaussian noise per client (Eq. 10) does not introduce bias to the aggregate gradient ), but it inflates the 2nd moment, leading to a noise floor that does not allow the gradient to converge to . The convergence neighborhood radius is proportional to , and the adaptive noise schedule (Eq. 13) of AgriChain-FL scales with the actual trust score of the network, thus narrowing the convergence bound as the network becomes trusted. We empirically tested accuracy in the range (and the calibration of was done using the Gaussian mechanism). With the privacy strength of (maximum privacy), they achieved an accuracy of 88.4 and converged in 68 rounds (compared to 45 rounds at with 95.7% accuracy), which is in line with the noise floor prediction. Figure 6(a) indicates the privacy-accuracy Pareto curve. These findings validate that the adaptive noise schedule of AgriChain-FL is an informed way of addressing the convergence-privacy trade-off in agricultural applications. Privacy–accuracy trade-off and resistance to model inversion attacks. Hybrid PBFT–PoA To ensure secure and efficient coordination among distributed nodes, AgriChain-FL integrates a hybrid consensus protocol that combines Practical Byzantine Fault Tolerance (PBFT) for integrity with Proof-of-Authority (PoA) for scalability. Let represent the set of blockchain validators. All the validators verify the transactions (model updates) in sequence across consensus rounds. PBFT makes sure that a new block is committed in the case of at least \:2f+1 out of \:3f+1 nodes agree, where \:f denotes the number of Byzantine (malicious) validators. The PBFT decision condition is given by: PoA introduces authority weighting, where each validator has a credibility score , determined from historical validation accuracy and participation rate. The hybrid trust score for consensus is then expressed as: and a block is accepted if , where is a predefined trust threshold (typically ). This hybrid approach ensures low-latency block validation from PoA while preserving the Byzantine fault resilience of PBFT, achieving an optimal balance between security and performance for large-scale agricultural networks. Privacy-Preserving Optimization through Adaptive Differential and Homomorphic Encryption Privacy protection in AgriChain-FL is achieved by combining Differential Privacy (DP) and Homomorphic Encryption (HE) to form an adaptive hybrid layer. For DP, each farm node adds noise to its local gradient before transmission: ensuring that an adversary cannot infer individual record contributions. The privacy guarantee satisfies the standard -differential privacy constraint: where differs from by one record, and defines the privacy budget. Homomorphic Encryption is applied to enable secure computation on encrypted gradients: where is an additive homomorphic encryption function like the Paillier cryptosystem. Aggregator is built to work with encrypted data, which is why the global model updates do not need decryption. The adaptive mechanism sets and key size based on the observed communication load and the trust score which is dynamic: where is an adaptation coefficient. This ensures stronger privacy when network trust is low and relaxed constraints when validators are consistently reliable. Let the transmitted (pre-encryption) update be perturbed as , where is independent across clients. Under independent noise, the aggregated perturbation in a round satisfies: Therefore, as client participation increases and/or data sizes become more balanced, the effective perturbation on the global update decreases, improving stability. Threat model, security assumptions, and attack resistance AgriChain-FL threat model is based on the assumption of a partially trusted environment where a farm node, validators, and aggregator can have Byzantine or adversary behaviors. The system protects against 3 broad types of attacks: Model poisoning attacks Malicious nodes attempt to inject manipulated gradients into the global model. The hybrid PBFT–PoA validation filters inconsistent updates by cross-verifying the deviation score: and rejects updates where , where is a consensus-defined tolerance bound. Data Inference and Leakage Differential privacy, and homomorphic encryption guarantee probabilistic bounds of leakage with probability and semantic security against ciphertext-attack, respectively. Sybil and replay attacks The identities of validators are associated with unique cryptographic keys and stake credentials, and verification by timestamps deems the replay of old model updates. Security proofs assume the hardness of the Decisional Composite Residuosity Problem (DCRP) for HE and the Byzantine resilience condition of for PBFT stability. AgriChain-FL under these conditions is confidential, the data is integrity-preserving, and accountability is achieved, providing formal resistance to data leaks and the manipulation of the model without a significant impact on the computational feasibility of industrial-level smart farming networks. The AgriChain-FL Framework Architecture overview and functional layers AgriChain-FL model is a structure that combines the use of federated learning, blockchain consensus, and flexible privacy protection to create a secure and intelligent farming system. It is designed as three primary functional layers, the Edge Layer, the Blockchain Coordination Layer, and the Federated Aggregation Layer, which have different computational and communication tasks. The local environmental and crop-related information at the Edge Layer is acquired with the assistance of the IoT sensors, including soil moisture, humidity, temperature, and nutrient levels. The nodes, in turn, are training a local model on the basis of such information and transferring the model changes in encrypted form instead of the data, thereby preserving the privacy at the farm level. The trust and verification are based on the Blockchain Coordination Layer. It implements a hybrid PBFTPoA consensus protocol to validate received updates, save them in immutability ledgers, and smart contracts to engage in the nodes, to allocate incentives, and to sustain reputation. The layer offers traceability and transparency with minimum latency. The top layer is the Federated Aggregation Layer since it takes verified updates via the blockchain to create the global model. Aggregation of the new model parameters is sent back to each of the participating farms, and this process continues in an effort to offer continuous and interactive learning without centralizing the information. This integrative scalability guarantees security, scalability, and energy efficiency, which offers the opportunity of real-time collaboration and privacy-aware crop intelligence maximization across a geographically spread farm, as shown in Fig. 2. Architecture of AgriChain-FL integrating blockchain, federated learning, and adaptive privacy modules. The figure was manually created by the author using Wondershare EdrawMax, Version 12.6.1. Available at: https://www.edrawsoft.com/edraw-max/. Blockchain-enabled federated aggregation protocol The aggregation protocol is based on blockchain technology and serves as a trust and validation mechanism for AgriChain-FL. It ensures that any model changes transmitted between the farm nodes are securely authenticated, logged, and aggregated, and that only authenticated changes are added to the global model. Every participating farm node is trained on its own SmartFarm sensor data to produce an encrypted gradient vector . Such gradients are then uploaded to the blockchain registry as transactions, in the form , timestamp, signature). After the transaction is transmitted, the blockchain validators perform the Hybrid PBFT-PoA consensus to verify the authenticity of every update. The smart contract automatically validates the updates by applying integrity constraints, i.e., parameter-norm bounds and training-round identity. Upon agreement, the validated updates are stored in the distributed registry and given a special block hash. The federated aggregator obtains all the verified updates by querying the ledger. It then conducts secure aggregation with the help of homomorphic encryption: which, when decrypted, yields the combined global gradient . The global model is then updated using: where is the global learning rate. The protocol ensures that all the parameters in the world model are verified via consensus and cryptographic security, resisting data modification, false updates, and collusion. Figure 3 shows the flow of transactions and model updates between the farm nodes through the blockchain ledger. Transaction and model update flow between farm nodes through the blockchain ledger. The figure was manually created by the author using Wondershare EdrawMax, Version 12.6.1. Available at: https://www.edrawsoft.com/edraw-max/. Smart-Contract Logic for Data Validation and Incentive Distribution The autonomous regulatory tool that ensures safe, transparent, and equitable participation by the farm nodes is AgriChain-FL smart contracts. The smart-contract layer is automated and consists of three basic processes: data validation, reputation management, and incentive distribution. In order to make a node update the blockchain with an encrypted model update the smart contract will first respond by authenticating the node by checking the digital signature and the model deviation score: If , where is the trust threshold is met, the update is approved for inclusion in the consensus block. The validator nodes then record the result, triggering a conditional contract event. The incentive logic is defined as: where denotes the reward token value, is the accuracy contribution, represents the timeliness of submission, and is the trust stability score derived from prior participations. The coefficients are dynamic weights adjusted by system policies to ensure equitable reward distribution across nodes. To detect anomalous or potentially malicious updates, AgriChain-FL computes a deviation score for each client update relative to the round’s reference update. Let denote the client’s model update and let denote the weighted mean update. We define the deviation score as the L2 distance: An update is accepted if , where is set robustly per round using a median-based rule: where \:MAD is the median absolute deviation and a constant (e.g., 2–3). The design is less sensitive to outliers and more resistant to model poisoning, e.g., sign-flip or scaled-update attacks. This type of contractualizm eliminates the need to have a central auditing institution, has independent validation and transparency in the performance assessment, and equitable reward among the decentralized networks in agriculture. Table 2 shows how AgriChain-FL smart-contract layer incorporates cryptographic validation, rewards on consensus, and adaptive governance to enable integrity and sustainability over the Smart farming ecosystems, which are decentralized. Table 2Smart-contract design and operational parameters for participant engagement.ParameterFunctionCondition/Trigger EventOutput/ActionRemarksValidateUpdate()Checks authenticity of model updateif hash(update) = = validApproves update and logs timestampPrevents replay and forgery attacksComputeReward()Calculates incentive using performance scoreif Δi ≤ θIssues reward Adaptive reward scaling per epochUpdateReputation()Adjusts node reputation scoreon block commitIncreases/decreases trust weight Influences validator selectionTriggerAudit()Initiates audit on anomalyif variance > δFlags update for manual/auto verificationEnhances system accountabilityPenaltyContract()Applies penalty for misbehaviorif malicious = = TrueDeducts token and lowers trust ratingEnsures deterrence of collusionRegisterNode()Enrolls new participant in the networkon registration eventAssigns unique node ID and keypairEnforces identity managementLedgerRecord()Logs verified update to distributed ledgeron consensus passAppends hash to blockchain ledgerGuarantees immutabilityRewardPoolRebalance()Adjusts reward pool across epochsafter every K roundsRedistributes excess or deficit tokensMaintains economic stabilityConsensusWeightUpdate()Dynamically adjusts validator weightson PBFT–PoA iterationRecomputes for next roundEnables adaptive governancePerformanceReport()Summarizes node and system metricsper training epochOutputs public dashboard dataEnhances transparency Smart-contract design and operational parameters for participant engagement. Adaptive privacy and gradient masking mechanism for secure model exchange To ensure confidentiality in model update transmission, AgriChain-FL employs a dual-layer privacy mechanism combining adaptive differential privacy and gradient masking. Each node perturbs its local gradient before transmission as: where represents the masking function and denotes Gaussian noise calibrated by the adaptive privacy controller. The masking function uses a random orthogonal matrix such that: which obfuscates gradient direction while maintaining its magnitude, preventing inverse reconstruction attacks. The variance parameter evolves dynamically according to: where is an adaptation rate and is the real-time trust score computed from blockchain consensus. In addition to DP and encryption, AgriChain-FL applies an orthogonal masking transform to obfuscate updates during transmission. Each round \:tuses a round-shared orthogonal matrix such that . Clients compute: and transmit for encrypted storage and validation. Because the same is used within the round, aggregation remains algebraically consistent: After secure aggregation, the aggregator recovers the correct aggregated direction by applying the inverse transform : Thus, orthogonal masking enhances transport confidentiality while preserving the correctness of the underlying FedAvg aggregation. The proposed privacy module provides a secure and loss-reducing model exchange by combining noise addition and gradient transformation, which has good inference, poisoning, and reconstruction attacks resistance and maintains convergence efficiency in distributed training of crop intelligences. Algorithmic implementation of agriChain-FL The framework AgriChain-FL has a modular algorithmic workflow, which incorporates federated learning, blockchain consensus, and adaptive privacy control as one integrated training cycle. Every stage, including the model training locally and the consensus aggregation, is mathematically formulated and implemented via automated smart contracts. This provides safe validation, aggregation, and redistribution of model updates with guaranteed privacy, transparency, and efficiency. Federated blockchain training algorithm (Pseudocode) The following pseudocode outlines the operational steps of the Federated Blockchain Training Cycle (AgriChain-FL): Algorithm 1: AgriChain-FL Federated Blockchain Training The algorithm will make sure updates in models are sent and authenticated without showing the raw data, maintaining data integrity as well as operational trust in the network. Secure gradient masking and model update sharing The gradient masking phase adds a privacy layer beyond encryption. Each node generates a random orthogonal masking matrix where . The masked gradient becomes: This transformation ensures that even if gradients are intercepted, they cannot be reconstructed into their original form. Encrypted updates are transmitted as: where represents the public-key encryption operation. Upon receiving, validators verify the integrity of via digital signatures before performing consensus validation. This multi-layered defense mechanism achieves end-to-end confidentiality and robustness against adversarial inference attacks during model exchange. Consensus-based aggregation and validation The Hybrid PBFTPoA consensus is a hybrid of the trust reliability of the PoA and Byzantine fault tolerance of PBFT. Every validator has an active trust weight that is updated based on measures of performance. On committing a block, the hybrid trust score is greater than the threshold: where if approves the transaction. After validation, the secure aggregation procedure aggregates encrypted model updates using homomorphic addition: allowing the global gradient to be derived without decrypting individual updates. The blockchain ledger then stores the block hash, validator signatures, and model metadata for traceability. This decentralized validation approach ensures tamper-proof model provenance and auditability of every training round. Computational and communication complexity analysis The total computational complexity of local training and masking per node is given by: where is the dataset size and \:d the model parameter dimension. The masking operation adds a quadratic term , but this is mitigated by parallel GPU execution. The communication complexity per round is dominated by gradient transmission and consensus overhead: where \:N is the number of nodes and \:mthe number of validators. PBFT requires message exchanges, while PoA reduces it to under stable trust conditions, achieving hybrid efficiency. Overall, the AgriChain-FL framework achieves a scalable balance: Computation-efficient due to distributed local training. Computation-efficient due to distributed local training. Communication-optimized through blockchain batching and hybrid consensus. Communication-optimized through blockchain batching and hybrid consensus. Security-preserving via adaptive gradient encryption and validation. Security-preserving via adaptive gradient encryption and validation. As such, AgriChain-FL is able to establish a reliable, privacy-sensitive, and resource-efficient federated intelligence system of decentralized agricultural systems. Experimental setup and dataset description In this section, the dataset, simulation setup, and assessment measures used to evaluate the proposed AgriChain-FL framework are outlined. The experiment was intended to make distributed agricultural systems realistic and reproducible, and considered the combination of federated learning, blockchain coordination, and privacy-preserving mechanisms. SmartFarm sensor dataset The SmartFarm Sensor Dataset34 is a heterogeneous IoT data of distributed agricultural fields with soil, humidity, temperature, and rainfall. The dataset records the correlation between the environment and crop production, providing more than 1.2 million labeled samples of fifty sensor clusters. Both samples consist of continuous values measured at an equal time interval. The data form is , where is the environmental feature vector, and is the yield index. The dataset was chosen because it covers extensive climatic and soil variations, making it ideal for federated agricultural intelligence modeling. Table 3 provide the most important statistical characteristics of the dataset SmartFarm (soil moisture, humidity, temperature, rainfall, yield index). Table 3Key statistical features of the SmartFarm dataset (soil moisture, humidity, temperature, rainfall, yield index).FeatureTypeMeanStd. Dev.MinMaxRelevance in ModelSoil Moisture (%)Continuous47.8214.3210.388.5Key determinant of irrigation schedulingAmbient Temperature (°C)Continuous27.466.7515.141.2Influences crop growth rateRelative Humidity (%)Continuous62.1411.8531.687.9Affects evapotranspiration and disease riskRainfall (mm)Continuous7.933.120.025.6Impacts soil moisture and nutrient transportYield IndexTarget (Label)0.790.120.450.96Crop productivity estimation metric Key statistical features of the SmartFarm dataset (soil moisture, humidity, temperature, rainfall, yield index). This dataset enables learning decentralized models by assigning each node to a region-specific dataset, thereby upholding data locality while enabling global learning via federated aggregation. Data preprocessing, normalization, and feature selection Data preprocessing involved cleaning, interpolation of missing readings, and synchronization of sensor timestamps. Each feature was normalized using the Min–Max scaling formula was used to standardize each feature to ensure it was comparable across nodes. To minimize redundancy, a correlation threshold of 0.85 was used to eliminate highly correlated features. Principal Component Analysis (PCA) was used to reduce the features into major dimensions that accounted for 95% of the total variance. This pipeline meant that all participating farm nodes operated on standardized, representative data, resulting in improved convergence behavior and reduced communication imbalance during training. Simulation environment and node configuration All experiments were conducted on a distributed testbed configured to simulate a federated smart agriculture network. The system consisted of a 10-node federated system, in addition to four blockchain validator nodes connected via a 1 Gbps local area network. Each node had an Intel Xeon Silver 4314 single-core processor (2.4 GHz), an NVIDIA A100 graphics card with 40 GB VRAM, and 64 GB RAM. It contained Python 3.10, TensorFlow Federated, PySyft, and Hyperledger Fabric v2.5 to coordinate blockchain. The hybrid PBFT-PoA mechanism of consensus was running a block size of 2 MB with a mean latency of 95 milliseconds and transaction throughput of about 2,400 transactions per second. This setup is a realistic simulation at the industrial scale, with the decentralized farm nodes communicating safely and asynchronously using blockchain consensus. Evaluation Metrics (Accuracy, F1-Score, Blockchain Throughput, Latency, Privacy Loss, Energy Consumption) AgriChain-FL was experimentally evaluated on a combination of both algorithmic and network-level performance metrics. The model accuracy is a metric that measures the predictive ability of the global model on unknown test samples. F1-score offered an equal assessment with an imbalance in the data. Blockchain throughput is a metric of the confirmed transactions per second, denoting the scalability of the system. Latency is used to measure the end-to-end delay between submitting a model and validation within the blockchain. The level of protection the local data had was estimated by privacy loss, which is the differential privacy budget ϵ. The kilowatt-hours of energy consumed were tracked to evaluate the computational and communication efficiency of each training epoch. Results and analysis This section will provide the overall analysis of the AgriChain-FL framework on the SmartFarm Sensor Dataset and its convergence metrics, performance, and behavior with various blockchain consensus mechanisms. The findings evaluate the accuracy and reliability improvement of federated learning in conjunction with blockchain coordination on agricultural real-world data variations. To test the differences in learning stability, convergence, and performance distribution among the decentralized nodes, the AgriChain-FL framework used four configurations, namely PoA, PBFT, Hybrid PBFT- PoA, and a Centralized Baseline (no blockchain), to analyze. As revealed in Fig. 4, the Hybrid PBFTpoA mechanism obtained the most balanced convergence behavior, with a final model accuracy at the global scale of 95.8% in 45 rounds, training loss and inter-node accuracy variance remain constant. Figure 4(a)–(f) shows different patterns: the global accuracy improvement per round. Figure 4(a), the exponentially decreasing training and validation loss Fig. 4(b, c), the accelerating convergence rate under the hybrid consensus Fig. 4(d), the consistency of node-level performance Fig. 4(e), and the consistency of the global learning curve over time Fig. 4(f). The results in Table 4 clearly demonstrate that Hybrid PBFT–PoA delivers the optimal performance among decentralized configurations. The lightweight efficiency of PoA and fault tolerance capabilities of PBFT facilitate its convergence speed and low latency, as well as accuracy consistency in all nodes. Even though the centralized base offers slightly better accuracy, it does not have such vital blockchain characteristics as immutability, transparency of trust, and reliable provenance. Moreover, the hybrid mechanism had a better F1-score (0.943) and AUC (0.973), which shows high generalization and robustness to local data imbalance. The stability index of (0.948) also shows that there is not much oscillation of global aggregation at asynchronous training. In the given case, the agroChain-FL system was analyzed in terms of barcode efficiency in terms of scalability and responsiveness. The test was done using transaction throughput (TPS), block confirmation time, latency distribution, and consensus communication overhead on different node configurations with 10 to 100 participants. Each system models a decentralized agricultural system whereby farm nodes submit encrypted model updates and obtain global parameters via the blockchain layer. Transaction throughput was observed to scale almost linearly to 60 nodes, where it continued with over 2350 TPS for hybrid consensus (PBFTPoA) until a slight degradation occurred after 80 nodes because of message verification saturation. The block confirmation time was also relatively low (0.911.3) and much faster than traditional PBFT systems (1.8 s on average) and PoA-only systems (2.6 s on average). The reason behind this is the adaptive rotation of leaders and batch optimization of the hybrid consensus protocol. The latency distribution analysis showed that communication delays were stable with increasing network density, with a median latency of less than 1.1 s and a 95th percentile of less than 1.6 s, even when the network was at its peak aggregation. The consensus overheads findings add to the overall finding that the cost of communication was sublinear with the size of the network, which confirms the system’s applicability to the industrial-scale agricultural operations with hundreds of linked farm devices. Figure 5 illustrates these outcomes, where Fig. 5(a–d) show the relationship between network size, throughput, block delay, latency variation, and consensus overhead under different consensus mechanisms. Blockchain throughput and latency comparison across network scales. The overall analysis of privacy protection and resistance to attack of the AgriChain-FL framework. In the experiments, the strength of privacy, model accuracy, and system latency were tested in the case of using hybrid different privacy (DP) and homomorphic encryption (HE) mechanisms during the federated aggregation. It was concerned with the sensitivity to privacy budgets (ε values), encryption depth, and the injection of noise on the learning performance, as well as protection in model inversion and membership inference attacks. Figure 6 depicts these relationships. Figure 6(a) indicates a privacy-accuracy trade-off; the less privacy is enforced, the less accuracy. The AgriChain-FL architecture proved to be stable with a predictive accuracy of almost 90% and ε ≥ 1.0, which means that the system can maintain predictive reliability when subjected to strong privacy limitations. Figure 6(b) indicates trends of the gradient reconstruction errors that the hybrid encryption brings a considerable amount of distortion of recoverable features, which lowers inversion accuracy by up to 72% when compared to the classic FL. Figure 6(c) compares the likelihood of a successful attack, and it is observed that the likelihood of the adversary deducing the information on the original crop data or belonging to the training was considerably lower than 12% when an adaptive privacy layer was applied. The last element is that the privacy trade-off versus latency is also indicated by Fig. 6(d), i.e., to gain greater encryption depth, the latency must somewhat increase (around 0.15 s per update), which cannot be considered to be excessively high considering the real-time operational margins in agricultural applications. Altogether, the AgriChain-FL structure allows balancing the security strength and the computational efficiency to guarantee the data confidentiality without the essential deterioration of the learning convergence. AgriChain-FL is scalable and power efficient; therefore, its ability to run on different sizes of networks was experimented with to determine its ability to be utilized in the real world to implement distributed smart farming conditions. The experiment modeled 20,200 networked farm nodes, all of which underwent 200 local training and frequent encrypted model aggregation via the blockchain layer. The metrics analyzed included throughput scalability, energy per training round, consensus energy ratios, and computation-communication efficiency. The findings have shown that the hybrid PBFT-PoA consensus is highly scalable and energy-proportional at a growing network size. Figure 7(a) shows that the system sustained a nearly linear increase in throughput through to 150 nodes, beyond which slight saturation effects were observed above 180 nodes, with more inter-node validation network traffic. Figure 7(b) shows that the average energy per round went up slightly (6.2 J to 8.5 J) with doubling the node density, which agrees with the effective local model updates with the minimum redundancy. According to the consensus energy ratio in Fig. 7(c) the use of energy in validation and block confirmation is only less than 15% of the total energy, which is less than the PoW-based validation and block confirmation in conventional frameworks. In the meantime, Fig. 7(d) demonstrates that the ratio of computation and communication efficiency is gradually increasing as the number of nodes aggregated increases and reaches a steady point at 1.8 above 120 nodes. This is an optimization of the scheduling of the batching and adaptive frequency of blockchain synchronization in AgriChain-FL. Comprehensively, the results support the application of the model to massive agriculture applications where sustainability, scalability, and energy management are of paramount importance to sustained federated intelligence. System scalability and energy utilization trends under increasing node participation. Performance under Non-IID client distributions To capture realistic agricultural deployments, we tested AgriChain-FL with statistically heterogeneous (non-IID) client partitions. We built a non-IID split by selecting clients based on varying environmental regimes using rainfall and soil moisture quantiles. In particular, the distribution of features within each client group was skewed toward either low-rainfall/low-moisture conditions or high-rainfall/high-moisture conditions, resulting in heterogeneous local objectives. In this environment, we contrasted FedAvg (no blockchain/privacy), blockchain-FL without strong screening, and suggested AgriChain-FL. Findings indicate that AgriChain-FL ensures a steady convergence and reduces the variance in the inter-client performance because (i) agglomeration is weighted, (ii) deviation-score screening, and (iii) consensus-verified provenance. Historical field farm yield data are likely to exhibit a long-tail distribution, with sample masses of high-yielding being relatively few. We experimented with the imbalance strength by downsampling high-yield samples from the chosen clients, resulting in imbalanced label distributions across nodes. F1-score and AUC are also reported along with accuracy. AgriChain-FL maintains predictive performance under imbalance, achieving reduced degradation compared to baselines through the stabilization effect of weighted aggregation and the reduction of extreme updates via deviation-score filtering. The summary of performance in three data regimes (IID, Non-IID, and Imbalanced) is provided in Table 5. Non-IID and imbalanced distributions reduce predictive performance across all means, as anticipated, but AgriChain-FL is least affected and converges in fewer rounds than the baselines, which is a benefit in the presence of realistic farm heterogeneity. Table 5Performance comparison under IID, Non-IID, and Imbalanced client distributions.Data regimeMethodGlobal accuracy (%)F1-ScoreAUCConvergence roundsMean variance across nodes (%)IIDFedAvg (baseline; no blockchain)96.100.9510.978421.88Blockchain-FL (no robust screening)92.150.9110.948582.87AgriChain-FL (proposed)95.840.9430.973452.10Non-IIDFedAvg (baseline; no blockchain)92.800.9150.951603.92Blockchain-FL (no robust screening)90.750.8940.936673.48AgriChain-FL (proposed)94.320.9270.964522.76ImbalancedFedAvg (baseline; no blockchain)91.450.8810.944633.74Blockchain-FL (no robust screening)89.600.8620.928703.59AgriChain-FL (proposed)93.180.9100.958552.69 Performance comparison under IID, Non-IID, and Imbalanced client distributions. Generalization across heterogeneous client distributions In order to test the extrapolation properties of the global model outside the training distribution, we tested the final AgriChain-FL global model on three unseen client profiles that were not observed during federated training: (i) an Arid client with soil moisture always in the 10th percentile, (ii) a Humid client with relative humidity in the 90th percentile, and (iii) a Mixed-Extreme client with In these invisible distributions profiles, AgriChain-FL obtained an accuracy of 91.4, 92.7 and 90.9 globally, versus 87.2, 88.5 and 86.1 plain FedAvg (no blockchain, no robust screening). Three mechanisms explain why AgriChain-FL has been enhanced in generalization. First, the weighted FedAvg aggregation (Eq. 4) will make the global model a data-proportional aggregate of local optima rather than an average of local optima, thereby minimizing the influence of clients with majority distributions. Second, the deviation-score filter (Eq. 2021) does not update clients whose local optima are extreme outliers, which can cause the global model to overfit to one environmental regime. Third, the consensus provenance across all training rounds implies that each round’s update is not disproportionately represented, which implicitly provides regularization. These findings validate the theoretical conjecture in Corollary 1 that raising \:K and using heterogeneity-sensitive screening decreases the convergence residual at large , and hence translates directly into better out-of-distribution generalization. We infer that AgriChain-FL generates global models with sufficient distributional breadth to support a wide range of previously unexplored agricultural settings. Table 6 provides a comparative quantitative analysis between AgriChain-FL and six notable frameworks. This gives a comparison of the quantitative analysis of AgriChain-FL and six distinguished structures. The suggested system has been found to surpass the previous literature on accuracy, throughput, latency, and resistance to privacy. Although its previous models had reached the peak throughput of approximately 1,600 TPS, AgriChain-FL had reached 2,350 TPS with a very low latency of 1.0 s, which highlights its scalable hybrid architecture of PBFT and PoA. Beyond, a similar adaptive differential privacy mechanism lowered the chances of a successful attack to 11% compared to the previous privacy-preserving FL system by a factor of ~ 60%. The optimized consensus energy profile of the system (8.5 J/round) also reaffirms that the system can be used in sustainable and large-scale agricultural intelligence networks, with traceability, energy efficiency, and compliance with privacy as one architecture. Table 6Quantitative comparison of AgriChain-FL with state-of-the-art blockchain–federated learning frameworks.Study [Ref]Domain/DatasetLedger & ConsensusFL/Privacy MechanismAccuracy (%)Throughput (TPS)Latency (s)Energy per Round (J)Privacy Protection (Attack Success ↓%)Hyperledger Fabric Design24,Industrial traceability testbedsFabric (PBFT ordering)Chaincode access control84.21,2501.811.338PBFT Original (OSDI)29Generic BFT modelPBFT——9002.512.8—Blockchain + Industrial IoT (TII)14IIoT and agricultural use-casesConsortium PBFTSecure aggregation + DP89.61,6001.29.426FL Privacy Survey (IoT-J)19Cross-domain IoT—DP, HE, Secure Aggregation87.3———24FL System at Scale (MLSys)17,Cross-domain cloud FL—FedAvg + Secure Aggregation89.0——7.931AgriChain-FL (Proposed)SmartFarm Sensor Dataset (soil, humidity, temperature, rainfall, yield)Hybrid PBFT–PoA Consortium LedgerAdaptive Differential Privacy + Homomorphic Encryption + Secure FL92.42,3501.08.511 Quantitative comparison of AgriChain-FL with state-of-the-art blockchain–federated learning frameworks. The experimental study shows that the AgriChain-FL framework will provide a balanced combination of federated learning and blockchain technologies to execute smart, safe, and sustainable farm activities. The model achieved a better result on every benchmark with improved global accuracy (92.4%), transaction throughput (2,350 TPS), and latency (1.0 s) with a high degree of privacy protection and scalability as compared to the current blockchain-based federated learning systems. The hybrid PBFT PoA consensus mechanism guaranteed deterministic validation with lower communication costs, and the adaptive privacy layer managed to address model inversion and membership inference threats, with a less than 12% probability of attack success. Moreover, the system also exhibited a very high level of energy efficiency, using 8.5 joules per round on average, and about 45% less than some similar IIoT-based systems. The relative assessment of state-of-the-art models confirmed the strength, sturdiness, and industrial usefulness of AgriChain-FL in varied network arrangements. The results prove that AgriChain-FL is a feasible base of decentralized agricultural intelligence that can be used to provide real-time, privacy-conscious cooperation in distributed farming networks. Conclusion and Future Work This paper introduced a novel federated blockchain scheme referred to as AgriChain-FL that can be applied to facilitate privacy-conscious, scalable, and trustworthy smart farming. The framework has the potential to yield massive advantages to the model accuracy, throughput, and latency in comparison to the existing blockchain-federated learning frameworks, in large-scale experimentation with the SmartFarm Sensor Dataset. The hybrid PBFTPoA consensus was very efficient with regard to transaction facilitation and deterministic validation. In the meantime, the adaptive differential privacy layer and homomorphic encryption layer had high data confidentiality, gradient reconstruction, and membership inference resistance. The findings affirmed the relevance of the model to enable distributed agricultural learning ecosystems that enable multi-node collaboration without the transmission of sensitive information about farms. This framework can be extended in several research directions. To begin with, edge aggregation in gateway devices can minimize communication delays and energy consumption in bandwidth-limited rural settings. Second, integrating post-quantum cryptographic primitives with the validation of authenticated reporters into secure communication channels would add long-term resilience to counter emergent quantum threats. Third, the adaptive client selection and personalization methods may also reduce extreme non-IID conditions that prevail in different climatic areas. Additionally, future investigations can consider interconnection with agricultural digital twin systems, in which global federated models provide predictive insights into crop development, irrigation planning, and climate adaptation evaluation. This would be improved by extending the framework to include multimodal sensing, such as UAV imagery, satellite imagery, and soil spectroscopy, enabling richer real-time agricultural decision support. AgriChain-FL provides a holistic roadmap of decentralized, insecure, and privacy-conscious agricultural intelligence. The framework provides a scaling basis for next-generation smart farming infrastructures that can enable resilient and sustainable international food systems by linking federated learning with blockchain-based trust coordination. Author contributions A.Y.J. conceptualized the research framework and designed the system architecture. A.Y.J implemented the experimental setup, conducted simulations, analyzed the results, and prepared all figures and tables. A.Y.J wrote the whole main manuscript text, and approved the final version of the manuscript. Funding The author received no financial support for the research. Data availability The dataset used in this study is publicly available on [34]. The source code and experimental resources supporting this study are publicly available to ensure transparency and reproducibility. The complete implementation of the proposed AgriChain-FL framework is available in the GitHub repository AgriChain_FL_Complete_Code: [https://github.com/Ayjaafar/AgriChain_FL_Complete_Code](https:/github.com/Ayjaafar/AgriChain_FL_Complete_Code). Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The dataset used in this study is publicly available on [34]. The source code and experimental resources supporting this study are publicly available to ensure transparency and reproducibility. The complete implementation of the proposed AgriChain-FL framework is available in the GitHub repository AgriChain_FL_Complete_Code: [https://github.com/Ayjaafar/AgriChain_FL_Complete_Code](https:/github.com/Ayjaafar/AgriChain_FL_Complete_Code).