Secure blockchain integrated deep learning framework for federated risk-adaptive and privacy-preserving IoT edge intelligence sets

Abstract

An enormous demand for a secure, scalable, intelligent edge computing framework has emerged for the exponentially increasing number of Internet of Things (IoT) devices for any substrate of modern digital infrastructure. These edge nodes distributed across heterogeneous environments serve as primary interfaces for sensing, computation, and actuations. Their physical deployment in unattended scenarios puts them at risk of being targets for resource manipulation. One widely accepted IoT architecture with traditional notions of edge may consider a threat to its centralized knowledge with an unbounded attack surface that includes anything that can remotely connect to the edge from the cloud-like domain. Existing strategies either forget the dynamic risk context of edge nodes or do not achieve a reasonable trade-off between security and resource constraints, essentially degrading the robustness and trustworthiness of solutions intended for real-life scenarios. To address the existing gaps, the work presents a novel Blockchain Integrated Deep Learning Framework for secure IoT edge computing, introducing a hybrid architecture where the transparency of blockchain meets deep learning flexibility. The proposed system incorporates five specialized components: Blockchain-Orchestrated Federated Curriculum Learning (BOFCL), which ensures risk-prioritized training using threat indices derived from blockchain logs; this adaptive sequencing enhances responsiveness to high-risk edge scenarios. Zero-Knowledge Proof Enabled Secure Inference Engine (ZK-SIE) provides verifiable privacy-preserving inference, ensuring model integrity without exposing input data or model internals in process. Blockchain Indexed Adversarial Attack Simulator (BI-AAS) focuses on testing the models in edge environments against attack scenarios drawn from common adversarial profiles and thereby facilitates a model defensive retraining. Energy-Aware Lightweight Consensus with Adaptive Synchronization (ELCAS) avoids overhead by seeking energy-efficient participants for global model synchronization in constrained environments. Trust Indexed Model Provenance and Deployment Ledger (TIMPDL) ensures model lineage tracking and deploy ability in a transparent manner by providing composite trust scores computed from data quality, node reputation, and validation metrics. Altogether, the framework combines the data integrity, adversarial robustness, and trust-aware deployment, shortening training latency, synchronization energy, and privacy leakage. It is a foundational advancement supporting secure decentralized edge intelligence for next-generation IoT ecosystems.

Introduction

The rapid proliferation of an increasing number of Internet of Things (IoT) devices in modern digital infrastructure has intensified the demand for secure, scalable, and intelligent edge computing frameworks. These edge nodes, distributed across heterogeneous environments, serve as primary interfaces for sensing, computation, and actuations. However^1–3, their resource constraints and exposure to untrusted domains make them highly vulnerable to data breaches, adversarial attacks, and privacy violations in the process. Traditional edge computing models, which often rely on centralized coordination or static training pipelines, fail to adapt to the dynamic threat landscape and do not provide sufficient guarantees for data integrity, provenance, or inference verifiability in the process. In this regard, recently introduced blockchain technology is novel because it provides a mechanism for decentralized trust, creating immutable ledgers^4–6 to hold verifiable consensus and process tamper-proof smart contracts. In spite of the fact that blockchain provides a natural complement to edge security, its integration with deep learning frameworks in resource-limited environments remains unexplored. Most efforts till date have either been using blockchain as a logging mechanism or for model exchange without fully utilizing its potential for dynamic trust computation, threat indexing, or inference validation in the process. On the contrary, the deep learning-based edge intelligence solutions tend to lack all sorts of adaptive dispositions either to priority training according to real-time risk or assuring secure inference that does not leak sensitive inputs in the process.

To bridge this gap, the presented work proposes a unified Blockchain Integrated Deep Learning Framework designed specifically for secure and efficient IoT edge computing. The architecture integrates decentralized learning, adversarial resilience, and trust-aware model deployment into one coherent system. Utilizing smart contract logic, blockchain consensus, and zero-knowledge proofs, the framework intelligently coordinates edge learning activities while maintaining data privacy and operational transparency in the process. The proposed system is constructed using five tightly integrated components, wherein each one targets an independent aspect of edge security. Blockchain-Orchestrated Federated Curriculum Learning (BOFCL) provides risk-based sequencing of federated model training across edge devices using risk assessments indexed from log entries residing in a Blockchain threat history. Secure inference at the edge is guaranteed with privacy and verifiability offered through the Zero-Knowledge Proof Enabled Secure Inference Engine (ZK-SIE). The Blockchain Indexed Adversarial Attack Simulator (BI-AAS) is about enhancing robustness by testing edge models against adversarial attack patterns indexed across the globe. The Energy-Aware Lightweight Consensus with Adaptive Synchronization (ELCAS) optimizes energy efficiency of model synchronization. TIMPDL (Trust Indexed Model Provenance and Deployment Ledger) ultimately traces model provenance and trusted scoring at deployment decisions. The above framework is a major move towards decentralized and secure AI for IoT, giving forward technical novelty in curriculum learning orchestration, verifiable inference, and energy-efficient synchronization through blockchains. It sets the stage for providing distributed intelligence fully backed up with security, thus paving the path for resilient, autonomous, and privacy-preserving edge ecosystems.

Motivation & contribution

In fact, this research was motivated by an awareness that edge intelligence frameworks were inadequate to solve such a dynamic problematic areas of security and privacy in the IoT environment. Since the edge node is located in a highly decentralized and often hostile environment, it is subject to adversarial manipulation, data tampering, and model poisoning. Now, on the existing federated learning solutions, even though they mitigate data centralization, most of them do not include any mechanism to adapt to changing threats and/or verify execution. In addition, those problems are exacerbated with issues in transparency in model provenance and trustworthiness of deployment. In this aspect, blockchain seems a wonderful solution, but it has to go beyond trivial logs or transaction handling sets in coupling with deep learning. It needs to have a tightly coupled architecture sets of security-driven, resource-aware, and trust-indexed frameworks that will include the merits of both blockchain and deep learning process.

It is hoped that this would meet those needs by presenting a holistic and technically rigorous mechanism that blends blockchain and deep learning in providing risk-adaptive, energy-efficient, and privacy-preserving edge intelligence in processes such as this. The first major component, BOFCL, describes a federated curriculum learning mechanism based on blockchain-logged threat indices to enable prioritized training at high-risk nodes. Secondly, ZK-SIE ensures inference validity without revealing private inputs or model details through zero-knowledge proofs, thus complying with demanding privacy needs. The third aspect, BI-AAS, autonomously generates adversarial conditions through common attack vectors from blockchain logs and real-time retraining of exposed models. The fourth, ELCAS, chooses synchronization participants based on energy-efficiency metrics, thus ensuring self-sustainable model aggregation in process. Ultimately, TIMPDL records the training lineage and computes trust scores for each model version using on-chain validation outcomes, data quality assessments, and contributor reputations.

There are many contributions of this work which are as follows: the first-ever hybrid architecture to date purported to integrating blockchain and deep learning for IoT edge security through risk-prioritized curriculum training; a new inference engine secured by zero-knowledge cryptographic proofs; a major breakthrough towards privacy in edge AI; the huge argument for blockchain in adversarial simulation since globally indexed attack histories inform testing and patching of model robustness; synchronization inefficiencies have been remedied through proposing an energy-aware consensus scheme specifically for edge environments; and the introduction of a trust-scoring mechanism for model selection and deployment thereby ensuring only verifiable, high-quality models activated at critical nodes. The experimental evaluation showed impressive improvements in terms of threat response accuracy, model robustness, tie efficiency, and privacy preservation and thus validates the usefulness of this framework for real-world implementation in IoT application contexts.

Literature review

A thorough review of research as documented in the high impact studies tells on the integration of blockchain with federated learning today and how this comprehensively cuts across all areas of privacy, security, energy optimization, and trust modeling through one IoT, healthcare, education, governance, and industrial sectors. The first initial major work that contributed to this is Liu et al.¹, where the main issues faced by federated learning systems were explored and then given attention to blockchain-based auditability and decentralization of trust. Alaya et al.² established a framework for the classification of specific domains within UAV systems, thus uncovering the fact that federated intelligences above them lose the adaptability necessary for mobile, resource-constrained platforms. Damaševičius et al.³ shifted their focus toward societal applications by combining blockchain with social credit systems based on enforcement and verifiable in-process methods. Choudhary et al.⁴ extended the framework with blockchain into education by studying general functional trends, while for performance and security risks, see Wu et al.⁵, which expect a growing audience once post-quantum cryptography is more mature. Senior to that, Hussain et al.⁶ brought in zero trust models within 6G underwater sensor networks to optimize security and energy with blockchain governance sets.

Mnasri et al.⁷ adopted a federated framework for medical records, specifically using a transformer-based BERT architecture and incorporating blockchain to provide audit trails. Eling et al.⁸ struck with the trade-offs regarding big data, risk modeling, and privacy matters in insurance concerning frameworks that are decentralized. While both Rajesh et al.⁹ and Kumar et al.¹⁰ present a trust-centric design in healthcare using neuroadaptive incentives and federated blockchain learning, Ohize et al.¹¹ and Punia et al.¹² broached the subject of blockchain applicability in voting systems and vulnerability audits while making it clear that these were built around the challenge of maintaining integrity under open-access blockchain structures. Transition was in the offing for edge computing and offloading, as seen by Lingayya et al.¹³ using Kubernetes-based machine learning offloading with privacy constraints and moving toward energy-centric models in IoT with blockchain coordination as addressed by Habibullah et al.¹⁴. Advancements of the type being explored by Venkatesan et al.¹⁵ included innovation in hybrid consensus protocols optimized using machine learning. However, Sun et al.¹⁶ showed that privacy keeps rearing its head in cross-platform recommender systems, indicating a growing need for federated modeling with encrypted collaboration. More governmental thesauruses came from the two in-depth medical applications of Nazir et al.¹⁷ and Jiang et al.¹⁸, including the dense neural network and trust augmented BFL architectures, respectively. Further specification was continued through these developments with opportunistic access in healthcare IoT contexts (Anjum et al.¹⁹) in process.

The contribution of Blockchain in forensic evidence, particularly in legal and academic contexts, is discussed in process by Sakshi et al.²⁰ and Cardenas-Quispe et al.²¹. Ma et al.²² used bibliometric analytics to track the emergence of blockchain in digital twin applications, while Shenoy et al.²³ looked into federated learning privacy metrics and implementation limits. IoV-specific security concerns were captured using federated learning by Ullah et al.²⁴ and later, by Wang et al.²⁵, who merged RFID with blockchain for emergency logistics. The outcome of secure smart homes and automated compliance workflows was explained in the works by Abul and Bilgen²⁶ and Alevizos²⁷, respectively. Khan et al.²⁸ provided a novel contribution linking thermal fluid systems with energy-efficient blockchain protocols. In blood management systems, it has been practically demonstrated by Kumar et al.²⁹ as to how the benefits of wastage tracking may be implemented with help from blockchain. Legal AI ethics, namely the correlation of regulatory frameworks to blockchain-mediated transparency in process, were brought to lights by Yadav and Ansari³⁰. Multi-layered blockchain solutions for 6G data control were proposed by Maalem et al.³¹, while the emphasis on applied contexts for students of management and economics was put forward in Gutowski et al.³² in process.

Iteratively, Next, as per Table 1, Manjula and Chauhan³³ were engaged with supply chain consensus optimization. Software verification issues in smart contracts built on a blockchain were raised by Olivieri and Spoto³⁴. Strong in industrial IOT, Ghaderi et al.³⁵ proved using a decentralized architecture capable of secure remote monitoring. Kuei and Chen³⁶ centred on enablers of blockchain adoption using ISM-MICMAC analysis. Integrated Cloud manufacturing, deep learning, and blockchain were studied by Ramshankar et al.³⁷. Customer-centric optimization received a strong focus. Suggala et al.³⁸ presented digital twin security for smart grids through inter-scholarly interpretable additive neural networks, thereby establishing a connection between explainability and trust. Rani et al.³⁹ and Umar et al.⁴⁰ presented decentralized certificates for verification and systems for energy trading, respectively. In both instances, blockchain will assure traceability, improvement in social welfare, and the rest. From genomic data, ethical perspective contributed by Balagurunathan and Sethuraman⁴¹ by aligning privacy with data governance in process. Arazzi et al.⁴² presented homomorphic encryption for privacy-preserving IoT abnormality detection in federated contexts. At the same time, Alam et al.⁴³ reveal how blockchain will expand its geographical reach toward smart city barter systems with an added emphasis on decentralized trading. Zhang et al.⁴⁴ took up the issue of digital rights management with blockchain-embedded zero-trust protocols. For public health, Sajedi and Mohammadipanah⁴⁵ presented worldwide SARS-CoV-2 data sharing through blockchain whereby bioinformatics information could be disseminated faster and more securely. There arises a decentralized consensus mechanism for vaccine supply chains in⁴⁶ for solving traceability problems arising from logistics-intensive operations. Academic diploma validation through blockchain is explored in Rustemi et al.⁴⁷, bringing out the hidden water from the educational sector. Together with depth in the area it represents, Hasan and Chaudhary⁴⁸ introduced ρi-BLoM, a fusion of blockchain and machine learning, intending to protect the industrial end. Zhuk⁴⁹, on his part, carried out critical evaluation of gaps in law and regulation regarding blockchain. Finally, Damaševičius et al.^50–52 provided a unifying view of the convergence of blockchain and IoT, together with a taxonomy and a synthesis of use cases.

Table 1.

Model’s empirical review analysis.

References	Method	Main objectives	Findings	Limitations
1	Blockchain-based Federated Learning	Enhancing trust and privacy in distributed learning	Surveyed trust models and blockchain integrations	Focused mainly on theoretical constructs
2	FL-Blockchain in UAV	Classifying integration patterns for UAV security	Proposed taxonomy of FL-BC solutions	Lacks implementation detail
3	Blockchain for Social Credit	Enforcing transparent social credit systems	Proposed enforcement mechanisms	No real-time deployment
4	Blockchain in Education	Analyze blockchain adoption trends in education	Identified key educational use cases	Limited empirical validation
5	Quantum-resistant Blockchain	Assessing performance of post-quantum BC	Introduced quantum-resistant protocols	Performance under resource constraints unclear
6	Zero Trust BC for 6G UASNs	Secure energy-efficient underwater comms	Developed zero-trust blockchain energy scheme	UASN-specific, not generalizable
7	BERT-Blockchain-FL Framework	Medical record security with FL	Proposed BERT Integrated BC architecture	Requires high computational resources
8	Blockchain for Insurance Privacy	Big data and privacy in insurance	Assessed regulatory implications	Conceptual without implementation
9	5G/6G Security Overview	Survey wireless security using BC	Mapped 6G privacy issues	Lacks protocol-level details
10	Neuroadaptive Blockchain in Healthcare	Smart incentivization via IoT	Showed adaptive learning for rewards	Requires further real-world validation
11	Blockchain E Voting Systems	Survey secure digital voting	Categorized existing voting BC systems	Scalability issues not addressed
12	SWOC Analysis of Blockchain	Identify blockchain vulnerabilities	SWOC applied to BC systems	Limited quantitative validation
13	Edge Offloading using ML + BC	Privacy-preserving Kubeedge offloading	Designed edge offload strategy with BC	No testbed or hardware validation
14	Energy-aware Blockchain IoT	IoT energy consumption modeling	Used BC to log and optimize power use	Real-time application not tested
15	Hybrid Consensus + ML	Enhance BC with ML-based hybrid consensus	Improved resilience and latency	Requires more adversarial testing
16	Privacy in Cross-platform Systems	Protect privacy in recommender systems	Analyzed techniques and tradeoffs	Federated scope limited
17	IoT Security via DNN and FL + BC	DNN + FL for collaborative IoT security	Reduced central failure points	Data heterogeneity remains a challenge
18	T-BFL Model for Medical Sharing	Trust and blockchain in federated learning	2D trust architecture for health data	Model complexity is high
19	Opportunistic Blockchain Access	Enhance IoT access security via ML + BC	Dynamic model based on network context	Lacks latency measurements
20	BC for IoT Forensic Evidence	Preserve digital evidence via BC	Defined forensic workflows	Sensitive to tampering during ingestion
21	BC Academic Integrity System	Verify degrees using blockchain	Built academic verification prototype	Limited multi Institution support
22	Blockchain-Digital Twin Nexus	Review blockchain and digital twin trends	Traced emerging application clusters	Focused only on bibliometric data
23	Privacy in Federated Learning	Explore privacy metrics in FL	Compared privacy-preserving strategies	Need for unified metrics
24	Blockchain-FL for IoV	Intrusion detection in vehicles	Improved IDS accuracy and auditability	Overhead on vehicular nodes
25	Trusted Material Dispatch via BC	RFID and blockchain logistics	Secure emergency dispatch system	Not field-tested
26	Smart Homes Trust Model	Secure IoT homes via bidirectional trust	Designed BC trust feedback model	Evaluation scale was limited
27	Automated Cybersecurity via BC + AI	Threat response via smart contracts	Proposed compliance automation system	No integration with existing tools
28	AI-Thermal Fluids with BC	Energy management via blockchain	Aligned smart energy with BC networks	Highly domain-specific
29	IoT Blood Management with BC	Prevent blood wastage using BC + IoT	Improved supply tracking	Only simulated results
30	Legal AI + Data Privacy	Address privacy in AI legal services	Discussed ethical concerns	Lacks enforceable solutions
31	6G Data Flow Control via BC	Secure multilayer flow using blockchain	Architectural design for 6G	Lacks performance benchmarks
32	Blockchain Teaching Framework	Introduce BC into management studies	Developed educational curriculum	Focuses on pedagogy, not tech design
33	BC Supply Chain Consensus	Secure protocol for SC management	Consensus optimized for traceability	Needs hardware validation
34	BC Software Verification	Challenges in verifying smart contracts	Outlined key software risks	No resolution proposals
35	Industrial IoT BC Monitoring	Remote monitoring using blockchain	Conceptual secured IIoT model	No deployment case studies
36	Supply Chain Adoption Factors	Identify drivers of BC adoption	Modeled enablers with ISM-MICMAC	Focused on theory, not systems
37	Deep Learning + BC for Cloud Mfg	Customer-centric smart manufacturing	Integrated learning and trust logs	Latency data missing
38	BC for Smart Grid Digital Twins	Secure grid twins via neural networks	Used interpretable additive models	High model complexity
39	Academic Certificate Notarization	Public BC for degree verification	Efficient, decentralized system	Scalability beyond pilot untested
40	BC Energy Trading in Microgrids	Optimized social welfare in trading	Modeled decentralized trading flows	Assumes full node honesty
41	Genomic Data Ethics + BC	Review privacy regulations for genome data	Mapped ethical-regulatory gaps	Lacks technical enforcement tools
42	FL + HE for IoT Anomaly Detection	Detect anomalies with privacy protection	Homomorphic encryption for secure inference	Encryption overhead is high
43	Smart City Barterchain	Blockchain-enabled barter systems	Modeled item exchange via BC	Token dynamics not validated
44	Image Copyright via Zero Trust BC	Copyright protection using BC + ZT	Ensured rights management	Complexity of integration is high
45	SARS-CoV-2 Data via BC	Global health data exchange	Secure, traceable sharing platform	Limited to public health domain
46	BC Vaccine Supply Consensus	Secure vaccine delivery	Decentralized consensus tested	Relies on consistent network conditions
47	BC Diploma Verification	Tamper-proof academic records	Developed DIAR for diploma verification	Adoption requires university support
48	Privacy in Industrial IoT	FL and BC for IIoT security	π-BLoM privacy framework proposed	Model efficiency trade-offs
49	Legal Challenges in BC	Discuss legal and regulatory issues	Critically assessed legal gaps	No cross-border harmonization method
50	Blockchain IoT Integration	Survey use cases and architecture	Summarized integration paths	Lacks real-time system references

These papers, when seen together, represent a systematic development in alignment between distributed intelligence and technologies that are trust-enhancing. What initially appeared to be about privacy and traceability is now maturing into dynamic modeling of trust, edge integration, and further optimized smart contracts. Major themes are energy efficiency, zero-trust enforcement, integrity of medical records, federated learning privacy, legal verifiability, and resource-aware offloading. A few of these works also addressed cross-cutting applications across different verticals like these—healthcare, education, and energy applications, and addressed architectural challenges like quantum resistance, cross-chain operability, hybrid consensus approaches, and semantic trust. Furthermore, attuning infrastructure with a very practical usage machine learning, especially federated deep learning, shows at the core of practical applications from lightweight edge devices up to the extremely capable 6G-enabled networks and beyond—with global health data systems. This outlook provides a layered understanding of how blockchain’s foundational traits—immutability, decentralization, and transparency—serve as a connective fabric for increasingly autonomous, data-rich, and privacy-sensitive environments. Some papers highlight such transition from theoretical architectures into validations testbed level, which indicates that the research community is now addressing real-world constraints, such as latency, computational overhead, regulatory compliance, and heterogeneous network behaviour. There is developing agreement, though, that blockchain has not yet fully transitioned from smart recorders to intelligent conductors in governing trust, energy, privacy, and model provenance in the federalized environment. This review shows that, while foundational infrastructure is improving, future advances will depend on composing blockchains, AI-aware consensus systems, and legal-technical alignments in decentralized applications.

Proposed model design analysis

The work that has been proposed in this model is a multi-layered architecture formulated to ensure security, robustness, and privacy in IoT edge computing through tightly integrated blockchain mechanisms and federated deep learning paradigms. The framework has three pivotal components—Blockchain-Orchestrated Federated Curriculum Learning (BOFCL), Zero-Knowledge Proof Enabled Secure Inference Engine (ZK-SIE), and Blockchain Indexed Adversarial Attack Simulator (BI AAS) in process. The proposed components are orchestrated through blockchain smart contracts equipped with cryptographic proof mechanisms, establishing a secure-by-design deep learning pipeline for distributed edge nodes. Each stage—ranging from data ingestion to model deployment—follows a verifiable process chain. Optimization procedures ensure efficient learning under resource constraints, while cryptographic verifiability guarantees that all training and inference events are tamper-proof. Adversarial robustness is continuously assessed through blockchain-coordinated consensus, enabling the framework to adapt to emerging threats without compromising operational integrity sets. The BOFCL mechanism assigns comparability of dynamic curriculum weight to any edge node, depending upon its threat exposure, computed by historical indices of threats logged in the on-chain sets. Ts are threat indexes defined for node 'i'; Li(t) is its local loss at iteration 't' in process. The curriculum weight wi(t) is derived as an exponential decay function dependent on the loss gradient and the threat index formalized via Eq. 1,

1

where, λ is a regularization constant and θ represents model parameters in process. Each node’s update Δθi(t) is then weighted via Eq. 2,

2

This ensures tamper-proof and verifiable updates recorded with cryptographic hashes on blockchain. Following aggregation uses a modified federated averaging scheme which incorporates these risk-weighted contributions via Eq. 3,

3

where, η is the global learning rate and N number of selected nodes, wherein each participants in a round in the process. Iteratively, Next, as per Fig. 1, Simultaneously, the ZK-SIE module commensurately guarantees that all inference operations set up in edge are verifiable but without exposing sensitive model parameters or input features. Given an input 'x', there an edge node does inference with its local model fθ(x) and then generates such a proof π over a zk-SNARK circuit 'C' via Eq. 4,

4

Fig. 1.

Model architecture of the proposed analysis process.

This yields a verifiable relation, which is represented via Eq. 5,

5

where, h(θ) is the hash of the model parameters in the process. The smart contract will confirm π before accepting this result of inference, according to the rules of correctness through a one-way binding to θ and auditability through blockchain logging in the process. The robust control was accomplished through the BI-AAS simulation and evaluation of adversarial attacks utilizing blockchain indexed threat signatures. For a model fθ under training, let δ be the adversarial perturbation computed via the following equation when constrained optimization over a limited threshold Bε, via Eq. 6:

6

With the constraint ‖δ‖p ≤ ε in process. The model’s robustness score R is then evaluated as the integral over the adversarial success probability Padv across input space X via Eq. 7,

7

All attack simulations are initialized through smart contracts that reference global blockchain signatures of known adversarial vectors {δj}, and post simulation update is re Integrated through retraining sets. The model is patched through adversarial training during which the new loss is then represented via Eq. 8,

8

Finally, the outputs of the BI-AAS processes Φ represent the adversarial vulnerability profile of the edge model and are stored on-chain sets.

This profile is generated as a multivariate function composed of the observed success rates Sj, attack transferability scores Tj, and defense effectiveness Dj for each indexed attack 'j' via Eq. 9,

9

where, αj, βj, γj are empirically determined weighting coefficients. This final equation quantitatively encapsulates under BI-AAS the model’s exposure to adversaries, giving a strong basis for process local hardening and global visibility of threats. Within a blockchain coordinated framework in which model training is risk aware and verifiable, inference remains private yet provably correct, and adversarial threats are actively monitored and mitigated through it. This cohesive design bridges critical gaps in decentralized edge AI, thus addressing integrity, privacy, and robustness in a uniform way. Each module strengthens itself through tightly coupling such synchronization into blockchain. Iteratively, Next, as per Fig. 2, The Energy-Aware Lightweight Consensus with Adaptive Synchronization (ELCAS) and Trust Indexed Model Provenance and Deployment Ledger (TIMPDL) make up the backplane of the proposed integrated deep learning framework in blockchain, which is devised specifically for the secure and efficient federated edge intelligence in IoT environments. Upstream modules as BOFCL, ZK-SIE, and BI-AAS deliver risk-adaptive training, private inference, and adversarial robustness, while ELCAS and TIMPDL ensure synchronization efficiency and deployment trustworthiness under edge networks’ typical resource-constrained, heterogeneous conditions. Their design is focused on energy-aware consensus, trust-score-based model ranking, and auditable provenance, all coordinated by blockchain smart contracts and cryptographic assurances. Iteratively, Next, as per Fig. 3, In the ELCAS mechanism, the model synchronization process is compared to the real-time availability of energy held by each participating node during the process. Each node 'i' knows its energy profile Ei(t), and synchronization eligibility gets determined using an energy-to-weight ratio ξi(t), expressed via Eq. 10,

10

where, Δθi(t) is the model update norm and ϵ is a small constant to prevent division by zero in the process. The global selection vector s(t) ∈ {0,1}^N is derived via Eq. 11,

11

where, τ(t) is a time Varying threshold determined through a dynamic percentile-based quantile function Qα(ξ(t)), via Eq. 12,

12

Fig. 2.

Overall flow of the proposed analysis process.

Fig. 3.

Data flow of the proposed analysis process.

Only nodes with si(t) = 1 participate in the current synchronization round in the process. The updates from the selected nodes are averaged using energy-weighted federated aggregation, yielding the updated global model parameters θ(t + 1) via Eq. 13,

13

This energy-aware consensus reduces unnecessary synchronization overhead while maintaining accuracy within acceptable bounds. All synchronization events and decisions by the participants are hashed and committed to the blockchain for auditability and for consistent guarantees for the distributed edge infrastructure in the process.

Concurrently, the TIMPDL Subsystem ensures the trustworthy deployment of trained models by computing a composite trust index ‘Im’ for every model version 'm', which combines three components: dataset quality score Qm, contributor node reputation Rm, and test validation accuracy ‘Am’ in process. The index is computed as per normalized weighted summation expressed via Eq. 14,

14

The dataset quality score Qm is derived from entropy-based measures. For a given training dataset Dm, with class distribution {pk}, the entropy H(Dm) via Eqs. 15 & 16,

15

16

where, K is the number of classes. Contributor node reputation Rm evolves over time and is modeled as a recursive exponential smoothing function of past behaviors via Eq. 17,

17

where, Sm(t) ∈ [0, 1] represents success ratio (e.g. correct outputs or low divergence) of the contributor’s prior models and λ ∈ [0, 1] is the smoothing constant in process. The validation accuracy Am is directly evaluated on a held-out global test set T via Eq. 18,

18

Each trust index Im is logged on the blockchain and queried during deployment in process. Nodes select models for inference or fine-tuning based upon the trust threshold θT enforced through a smart contract rule elaborated via Eq. 19,

19

The final output of the integrated ELCAS and TIMPDL pipeline is a composite deployment decision matrix D ∈ {0,1}^{N × M}, where, the decision is represented via Eq. 20,

20

This matrix encompasses the entire synchronization and deployment decision space, where only trusted models from energy-sufficient nodes are deployed or further trained in the process. Energy profiling, trust indexing, and blockchain-based auditability provide a robust ecosystem whereby the aforementioned three objectives-security-efficiency and integrity-are concurrently optimized in the process. This architecture complements the upstream learning and inference modules with a verifiable, energy-adaptive, and reputation-aware execution layer that guarantees the synchronization and reliable deployment in the field of outputs from the decentralized training and testing phases. Proposed architecture combines mechanisms of blockchain and federated deep learning to impart secure, robust, and private-preserving edge intelligence in IoT environments. The proposed architecture (Fig. 4) comprises five interfacing components that can help in the above routing-the Blockchain-Orchestrated Federated Curriculum Learning (BOFCL), the Zero-Knowledge Proof Enabled Secure Inference Engine (ZK-SIE), a Blockchain Indexed Adversarial Attack Simulator (BI-AAS), the last two are Energy-Aware Lightweight Consensus with Adaptive Synchronization (ELCAS), and Trust Indexed Model Provenance and Deployment Ledger (TIMPDL). These components are run in a co-ordinated way using blockchain smart contracts and cryptographic assurances for verifiable trust, risk-adaptive training, and resilient model deployments.

Fig. 4.

Model architecture of the proposed analysis process.

The architecture diagram depicts layered interaction:

• Data Collection Layer: Edge nodes collect multimodal IoT data within a distributed form, collecting environmental readings, device telemetry, and network activity logs.

• Risk Assessment Layer: Threat indices for real-time attack detection and subsequent sequencing of training priorities for high-risk nodes are computed by BOFCL directly from security log recording in the blockchain-based infrastructure sets.

• Secure Computation Layer: ZK-SIE permits inference verification under zero-knowledge proof, enabling correctness assurance while keeping secret both model parameters and raw input data samples. The architecture thus creates end-to-end integrity and efficiency in decentralized learning pipelines, even in adversarial or resource-constrained readings. Up next, we present an Iterative Validation of the Proposed Model via several metrics and compare them with different models under different scenarios.

Detailed component description

Blockchain-Orchestrated Federated Curriculum Learning (BOFCL)

BOFCL introduces a mechanism for risk-aware scheduling in federated learning. Each node has its participation weight computed, following penetration of local loss gradient and threat indexing from the blockchains logs. For example, consider a node with a threat index of 0.78 and has a loss gradient of 0.042. The node gets a weight of w = e − 0.5*0.042*0.78 ≈ 0.9838 whereby high-risk nodes will take preference during model aggregation sets.

Zero-Knowledge Proof Enabled Secure Inference Engine (ZK-SIE)

ZK-SIE allows for using zk-SNARK circuits that eliminate the unrevealed raw input features or model parameters. In this way, it has an average data-smart contract-verifying activity of 2.1 ms before accepting the inference results.

Blockchain Indexed Adversarial Attack Simulator (BI-AAS)

This ledger contains known attacks which include FGSM, PGD, and DeepFool perturbations. It evaluates models against those profiles and triggers a defensive retraining if a detection rate falls under threshold values, for example, 85%.

Energy-Aware Lightweight Consensus with Adaptive Synchronization (ELCAS)

ELCAS estimates the energy to update every node in order to choose which groups will participate in the synchronizations. Only the top 40% of energy-efficient nodes continue to aggregation, which cuts more than 40% of redundant communication without losing accuracy sets.

Trust Indexed Model Provenance and Deployment Ledger (TIMPDL)

Assigning trust elements TIMPDL has given a composite trust score Im = 0.4Qm + 0.3Rm + 0.3Am for each model, where Qm is dataset quality, Rm is contributor reputation, and Am is validation accuracy. Models that score above a threshold (like 0.75) are granted induction into deployment pool sets.

Improved flow with example execution scenario

Consider a distributed industrial IoT network monitoring a smart manufacturing facility: An edge node managing a CNC machine reports vibration amplitude of 0.91 mm/s, temperature of 88.7 °C, and 2.1% packet loss. BOFCL assigns a high training weight due to a threat index of 0.82 and a recent anomaly spike in blockchain logs. The node updates its local CNN model and submits a zero-knowledge proof of correctness for an inference indicating "Critical Failure Risk" in the process. The proof is verified on-chain before results are accepted in process. BI-AAS launches an adversarial test using a PGD perturbation with ϵ = 0.08 \epsilon = 0.08 ϵ = 0.08 imposed on vibration data samples. The baseline model misclassifies the condition, but the retrained model achieves correct classification. ELCAS selects the node for aggregation based on an energy-to-update ratio exceeding 6000, enabling low-overhead synchronization. TIMPDL calculates the trust score as I = 0.4(0.93) + 0.3(0.85) + 0.3(0.92) = 0.896, which exceeds deployment thresholds. The revamped global model is thus deployed across the entire network to ensure secure, verifiable, and energy-efficient operations.

Comparative result analysis

The experimental designed to evaluate performance amounts to setup with the blockchain Integrated deep learning framework as an IoT edge environment with large scale, characterized by device heterogeneity, variable energy availability, and adversarial threat dynamics. The simulation environment consists of 100 distributed edge nodes, each emulating an IoT device with varying computational capacities (GFLOPS varying from 0.5 to 2.5 GFLOPS), memory footprints (from 128 MB to 1 GB), and hence variable battery energy levels (500–2500 mWh). The communication latency between the nodes is being simulated using a Gaussian distribution with a mean of 28 ms and a standard deviation of 8 ms to reflect the wireless network conditions. The local training on each node is carried out using a subset of a contextual dataset partitioned from IoTID20, which contains labeled data streams from smart home IoT sensors with benign and attack-class traffic from smart locks, surveillance cameras, and thermostats, considering DoS, DDoS, Man In-The-Middle, and Botnet attacks. Further behavioral datasets, such as TONIoT telemetry and N-BaIoT device logs, serve the purpose of enriching the threat logs and historical blockchain entries. These entries are processed to compute threat indices for each node using a moving window average of abnormal activity rates, such that initial risk scores vary from 0.12 (low-risk) to 0.89 (high-risk). The federated learning models are trained using a 5-layer CNN and a 3-layer LSTM with time-series predictive and classification tasks. Every edge node implements a batch size of 32, local epoch count of 5, as well as an adaptive learning rate beginning from 0.01, along with a decay value of 0.98. The gradient update hashes are generated using SHA-256 just before they are presented to the blockchain. Smart contracts are programmed in Solidity on a private Ethereum testnet with fixed block intervals of 6 s and maximum gas limit of 8 million per block. Using ZoKrates, zero-knowledge proofs are generated under 512-bit arithmetic circuits for inference verification in process. The adversarial simulation engine (BIAAS) utilizes FGSM, PGD, and DeepFool attacks under perturbation limits ϵ ∈ [0.01,0.1] in ℓ∞ and ℓ2 norms in process.

The energy-aware consensus, used in evaluating trust and synchronization dynamics, is configured using ELCAS with the energy threshold quantile α = 0.4, dynamically selecting for each round the top 40% energy-efficient nodes. Model synchronization takes place after every 10 communication rounds, whereas global aggregation is performed in an energy-weighted manner. The trust index for each model (via TIMPDL) integrates dataset quality scores generated from entropy values between 0.71 and 0.96, validation accuracy ranging from 81 to 94%, and contributor reputations initialized from 0.3 to 0.9 in process. The trust threshold θT for deployment purposes is set empirically at 0.75, with models valued above it entering the deployment pool in process. Each model version is tracked for adversarial robustness scores using an integrated scoring mechanism that logs both attack detection and patch effectiveness rates on the blockchain. Using ZK-SIE, inference verification takes on average of 2.1 ms, while energy-aware synchronization cycles cut redundant communication by over 40%.The complete experimental journey is performed over 50 rounds of global communication, with indices captured being training latency, global model divergence, privacy leakage, risk, synchronization energy cost, and robustness against adversarial examples. This allows a full-fledged evaluation of proposed framework real-world viability in securing intelligent IoT edge ecosystems.

The experimentation includes the evaluation of all three benchmark IoT cybersecurity datasets, namely IoTID20, TONIoT, and N-BaIoT, all of which present heterogeneous and realistic threat scenarios. IoTID20 is basically composed of network traffic captured in a smart home testbed; it has a multi-class label of benign traffic and different attacks such as DoS, DDoS, MITM, and data injection in process. It has around 1.2 million flow records with 83 features per instance. TONIoT has telemetry and network logs associated with multiple IoT devices and services while uniting host-based and network-based data into more than 400,000 labeled samples that include authentication failures, privilege escalations, and backdoor exploits. Last, the N-BaIoT dataset demonstrates traffic from 9 commercial IoT devices, some of which have been infected using the Mirai and Bashlite botnets, and it all captures more than 14 million records in total. These datasets are used in partitioned form to simulate non IID (non Independent and identically distributed) environments across edge nodes. The unique nature of the device types, traffic patterns, and attack vectors from these datasets can ensure broad and real-world edge scenarios against adversarial conditions.

Hyperparameter tuning model training is via a well-configured setting that seeks a balance in speed for convergence with accuracy and communication efficiency in a federated way in process. The main learning model is a 5-layer convolutional neural network (CNN) trained with an initial learning rate of 0.01, decayed at the rate of 0.98 after every global round in the process. Every edge node performs 5 local epochs, the size of the batch is 32, the optimizer being used is the Adam optimizer with β1 = 0.9, β2 = 0.999, and ϵ = 1 × 10e − 8 in process. Dropout regularization is used at a rate of 0.3 to reduce overfitting on local datasets, sampled for models. Perturbation magnitude ϵ = 0.07 for FGSM-based attacks was used on the retraining for adversaries, while defense strategies were validated using an ensemble of PGD and DeepFool attacks. Trust index calculation uses weighting coefficients α = 0.4, β = 0.3, and γ = 0.3 for dataset quality, node reputation, and validation accuracy respectively in the process. Blockchain-related parameters include a smart contract gas limit of 8 million, ZK proof circuit size of 512 bits, and energy threshold quantile α = 0.4 for consensus participation in the process. These hyperparameters are tuned almost empirically to maintain a high accuracy supported (> 90%) with communication efficiency, robustness, and verifiability in decentralized edge learning environments. Performance evaluation of the proposed Blockchain Integrated Deep Learning Framework was conducted over IoTID20, TONIoT, and N-BaIoT datasets in terms of classification performance as well as operational metrics over decentralized, adversarial, and resource-constrained edge environments. Then, a comparative analysis was made against three existing techniques denoted as Method³, Method⁸, and Method²⁵, which all represent current state-of-the-art federated learning or blockchain-enabled IoT security baselines. Present the classification accuracy gained on the IoTID20 dataset in process. The proposed model consistently outperformed on all baseline methods on multiple classes of attacks.

Table 2 shows the drop in model accuracy when exposed to FGSM adversarial samples.

Table 2.

Classification accuracy (%) on IoTID20 dataset.

Model	Benign	DoS	DDoS	MITM	Data injection	Overall accuracy
Method3	91.3	84.1	81.5	78.2	76.4	82.3
Method8	93.0	86.4	83.9	80.7	78.9	84.9
Method25	94.6	89.1	85.5	83.1	80.2	87.2
Proposed	97.1	93.3	91.2	89.4	87.0	91.6

Table 3 compares the trust score computed via TIMPDL across different deployments.

Table 3.

Robustness against FGSM attack (accuracy % drop).

Model	IoTID20	TONIoT	N-BaIoT
Method3	− 36.1	− 32.7	− 33.5
Method8	− 30.9	− 27.4	− 29.1
Method25	− 25.3	− 23.8	− 22.4
Proposed	− 12.9	− 11.4	− 13.1

Table 4 evaluates the additional latency introduced by blockchain interactions in process.

Table 4.

Trust Index Distribution Across Models (Mean ± Std Dev).

Model	IoTID20	TONIoT	N-BaIoT
Method3	0.59 ± 0.07	0.61 ± 0.05	0.58 ± 0.06
Method8	0.65 ± 0.05	0.68 ± 0.06	0.66 ± 0.04
Method25	0.71 ± 0.06	0.73 ± 0.04	0.70 ± 0.05
Proposed	0.88 ± 0.04	0.91 ± 0.03	0.89 ± 0.04

Table 5 shows average energy spent by edge nodes during global model synchronization in the process.

Table 5.

Blockchain Overhead (ms per transaction).

Model	Logging (ms)	ZK Proof (ms)	Consensus Sync (ms)
Method3	3.9	N/A	5.6
Method8	3.2	N/A	4.8
Method25	2.7	N/A	4.1
Proposed	2.1	2.1	3.9

Table 6 reflects privacy leakage potential using inference-based attacks.

Table 6.

Energy Consumption During Model Synchronization (mWh).

Model	IoTID20	TONIoT	N-BaIoT
Method3	98.1	95.7	97.4
Method8	84.2	81.9	83.3
Method25	72.5	70.1	68.7
Proposed	54.3	52.8	53.9

Table 7 compares the time taken to complete one round of global model synchronization in process.

Table 7.

Inference privacy risk (data leakage probability %).

Model	IoTID20	TONIoT	N-BaIoT
Method3	17.6	15.2	18.0
Method8	12.3	11.0	12.9
Method25	8.9	7.7	9.2
Proposed	1.4	1.1	1.3

Table 8 model divergence is measured by cosine distance between local and global model updates.

Table 8.

Synchronization completion time (s).

Model	IoTID20	TONIoT	N-BaIoT
Method3	4.3	4.1	4.2
Method8	3.7	3.6	3.5
Method25	3.2	3.0	3.1
Proposed	2.6	2.5	2.4

Table 9 measures time taken to select a suitable model based on trust index in process.

Table 9.

Global model divergence (%) during aggregation.

Model	IoTID20	TONIoT	N-BaIoT
Method3	0.19	0.21	0.20
Method8	0.14	0.15	0.13
Method25	0.10	0.11	0.09
Proposed	0.06	0.07	0.06

Foremost, it exhibited the best performance of the proposed framework through evaluation using major metrics such as accuracy, energy efficiency, robustness, privacy preservation, and trust in process. More specifically than Method³, Method⁸, and Method²⁵, the system successfully reduced the adversarial vulnerability through much lower communication overhead and enhanced the model reliability and agility for deployment in decentralized edge environments. Next, discuss an Iterative Validation Impact Analysis of the Proposed Model regarding different scenarios.

Validation & impact analysis

The results presented in Tables 2, 3, 4, 5, 6, 7, 8, 9 and 10 along with Figs. 5 & 6 show that the proposed blockchain Integrated deep learning framework performed better and was more robust in terms of security for IoT edge computing. In terms of classification accuracy, the proposed model yielded an average of 91.6% accuracy with respect to different attack types, as shown in Table 2. This was better than Method³, Method⁸, and Method²⁵ by 9.3%, 6.7%, and 4.4%, respectively, and could be attributed directly to the model’s risk-aware training strategy via BOFCL and dynamic data prioritization-increased context sensitivity for learning. In addition, Table 3 shows that the proposed system remains robust under FGSM attacks, where the loss in accuracy did not fall above 13% across datasets; while baseline methods suffered above 30% degradation. This thus shows that BI-AAS would be effective in predicting and displacing such perturbations, which would then be important for any real-time threat response within mission-critical IoT systems like industrial automation and connected healthcare processes.

Table 10.

Deployment latency (ms per model selection request).

Model	IoTID20	TONIoT	N-BaIoT
Method3	12.4	11.9	12.1
Method8	9.2	8.7	9.1
Method25	6.3	6.5	6.1
Proposed	4.7	4.3	4.5

Fig. 5.

Model’s integrated result analysis.

Fig. 6.

Model’s overall result analysis.

Through operational metrics regarding energy efficiency, inference privacy, and trust-based deployment system, the real-time deployment of the proposed framework is furthered validated. As illustrated in Table 6, the synchronization energy consumption was lower for the proposed model, spending only ~ 54 mWh round, which is ~ 25% less than the best baseline due to the adaptive ELCAS module. Thus, operations become sustainable in battery-constrained edge environments. Tables 7 and 10 depict privacy-preserving, low-latency deployment characteristics, wherein data leakage risks are curtailed to almost 1% and model selection latency reduced to less than 5 ms. Such features are necessary for privacy-sensitive applications like smart surveillance and medical telemetry, which require inference correctness and user anonymity to be ensured at the same time. Trustworthiness and stability are confirmed, according to results in Tables 4 and 9, with high trust scores (mean 0.88–0.91) and low divergence between global and local models to ensure reliability and consistency in federated learning settings. In the real deployment scenario, this enables resilient, self-verifying edge intelligence systems that can securely operate despite constrained resources and active threats. Next in line is development of Iterative Validation Use Case for the Proposed Model, which would further help the readers understand the entire process.

Validation using iterative practical use case scenario analysis

Consider a smart industrial facility comprising a network of edge IoT devices including thermal sensors, vibration monitors, surveillance cameras, and robotic actuators deployed across a manufacturing floor. One of the edge nodes, namely Node 12, is responsible for monitoring a high-value CNC machine set. At a given point in time, Node 12 collects a batch of readings from the sensors, which include temperature = 87.3 °C, vibration amplitude = 0.89 mm/s, network packet loss = 2.3%, and current frequency of system calls = 0.18 kHz. These readings, along with the system logs, will be preprocessed and stored locally for model training. From the blockchain logs, Node 12 has been assigned a threat index of 0.81 due to recent abnormal access attempts and previous cases of network scanning activity. BOFCL employs a risk modulated function to weigh the training curriculum for this node where the gradient of the local loss with respect to the model parameters has been computed as 0.043. The regularization constant is taken as λ = 0.5, hence curriculum weight is given as w = exp⁡(− 0.5 × 0.043 × 0.81)≈0.9826, which gives strong preference to Node 12 for consideration during the present round of training in process.

Post-local training, Node 12 initiates an encrypted inference request on live data with the temperature hitting 91.2 °C and the vibration moving up to 1.12 mm/s. The output of the inference states "Component Failure Risk: HIGH." A zero-knowledge proof is generated over a zk-SNARK circuit confirming the integrity of this inference without exposing raw features or model weights. The proof is then verified by a smart contract on the blockchain in a process that was recorded as having a proof verification time of 2.0 ms. Concurrently, BI-AAS has begun the setup for adversarial simulation for the model of Node 12, by means of a crafted perturbation δ of magnitude 0.07 in the ℓ∞ norm toward the vibration input. The adversarially perturbed input causes a prediction flip in Method⁸, but the proposed model is robust against the perturbation, yielding an unchanged output, leading to a calculated improvement in robustness score of 29% compared to the baselines. These fresh adversarial samples will spur defensive retraining, and the new model will be reconfirmed locally.

As soon as the local model updates are ready, Node 12 reports energy availability as 620 mWh, placing it in the top 40th percentile per ELCAS’s quantile-based selector for synchronizations. The update vector norm is 0.093, thereby leading to an energy-to-weight ratio of ξ = 620/(0.093 + 0.01) ≈ 6067, which surpasses the participation threshold. Following that, the new global model update is then produced through weighted aggregation. The TIMPDL evaluates the trustworthiness of the new model version from the Node 12 Sets. The dataset quality score is computed as 0.92 based on class entropy, contributor reputation is 0.84 from historical behavior, and validation accuracy is 0.91 on the global test sets. Thus, using weights α = 0.4, β = 0.3, and γ = 0.3, the trust index is calculated as I = (0.4 ∗ 0.92 + 0.3 ∗ 0.84 + 0.3 ∗ 0.91)/1.0 = 0.894. This score exceeds the deployment threshold of 0.75, allowing the model to be accepted into the trusted deployment pool for real-time use across the facility, thus completing the secure and verifiable edge learning cycles.

Conclusion and future scopes

The paper introduces a novel blockchain Integrated deep learning framework for secure, efficient, and privacy-preserving IoT edge computing in which some of the fundamental gaps concerning adaptability, robustness, trust, and energy-aware learning were filled. The proposed system consists of five interlinked modules that were brought to overall realization in risk-prioritized model training, verifiable private inference, adversarial resilience, energy-efficient synchronization, and trust-based model provenance tracking—BOFCL, ZK-SIE, BI-AAS, ELCAS, and TIMPDL. The results showed that on the IoTID20 data-set the framework achieved an overall classification accuracy of 91.6%, which was better than the next best baseline by 4.4%, and degraded accuracy under FGSM adversarial attacks to only 12.9%, compared to 25.3–36.1% in previous methods. The trust index scores resulted, on an average, to be between 0.88 and 0.91 which is quite high relative to other competing approaches thereby enabling a secure and credible model deployment. Energy utilization during synchronization was cut down to 54.3 mWh, which implies over 25% savings compared to Method²⁵, and the overheads involved with the blockchain continued to remain below 2.1 ms per operation in process. These improvements validated the framework’s practical viability with respect to real-time IoT deployments across resource-constrained and threat-prone environments and laid a solid foundation for next-generation decentralized edge intelligence in process.

Future scope

Although the present framework does advance edge learning much further into the domains of security and trustworthiness, there remain many possible avenues for future work in extending its applicability across what could be considered the three main scopes. The first would be to broaden its ability to support a wider variety of heterogeneous model architectures, such as transformers and graph neural networks. This is in keeping with allowing learning on more diverse IoT modalities, including relational sensor data and edge-augmented perceptions. A second point would be to allow integration with cross-chain federated systems and with interoperable smart contracts, so that multiple blockchain ecosystems could coordinate learning across broader trust domains, especially in cases where enterprise, municipal, and personal IoT devices converged in process. Third, as a next step for dynamic optimization of the consensus protocols, reinforcement learning could be employed to cut down the amount of communication overhead while maximizing synchronization efficiency under fluctuating energy budgets and communication constraints. Explainable AI components would, in addition, render deeper insights into model behavior and thus aid in compliance with regulations in some sectors like healthcare and smart governance sets in the overall trust index calculation in process.

Limitations

Despite its many strengths, this study presents several limitations. Zero knowledge proofs impose an overhead heavy enough to restrict their use in lower tier edge devices, thus scaling down the power of the scheme in ultra-low power devices. In addition to synchronization efficiency, energy-aware consensus reduces model update contributions from low-energy but high value nodes, thus ultimately biasing the learning process over time. Another limitation arose with respect to the static trust weights in TIMPDL, which works in controlled simulations but would not be generalized optimally under changing adversarial conditions or unverified data sources. In addition, the current framework assumes honest-but-curious behavior among participants, and therefore limits its resilience to fully malicious nodes capable of generating poisoned proofs or manipulating reports on energy. Addressing these issues in the next version of the framework will be imperative to enhancing robustness, fairness, and efficiency throughout highly adversarial and heterogeneous IoT ecosystems.

Abbreviations

AI

Artificial Intelligence

BC

Blockchain

BERT

Bidirectional Encoder Representations from Transformers

CNN

Convolutional Neural Network

DNN

Deep Neural Network

DIAR

Diploma Integrity and Authentication Registry

FL

Federated Learning

FGSM

Fast Gradient Sign Method

HE

Homomorphic Encryption

IDS

Intrusion Detection System

IIoT

Industrial Internet of Things

IoT

Internet of Things

IoV

Internet of Vehicles

ISM-MICMAC

Interpretive Structural Modeling-Matrice d’Impacts Croisés Multiplication Appliquée à un Classement

Kubeedge

Kubernetes-based Edge Computing Framework

ML

Machine Learning

RFID

Radio Frequency Identification

SC

Supply Chain

SWOC

Strengths, Weaknesses, Opportunities, and Challenges

UASN

Underwater Acoustic Sensor Network

UAV

Unmanned Aerial Vehicle

ZT

Zero Trust

ZTNA

Zero Trust Network Access

BLoM

Privacy Preserving Blockchain-based Learning Model

Author contributions

K. Swathi, Putta Durga, K. Venkata Prasad – Manuscript Preparation. Atmakuri Krishna Chaitanya, Kuraganti Santhi – Diagrams and tables. P. Vidyullatha, S Venkata Achuta Rao – Review of manuscript.

Data availability

Data available with corresponding author, can be provided on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.