Deny Prasetyo, Siska Narulita, Ahmad Jurnaidi Wahidin, Rosalina Yani Widiastuti · 7 authors
This study introduces a trust centric machine learning framework designed to improve decision making reliability and security in decentralized digital service ecosystems. Traditional machine learning models often focus on accuracy and efficiency but fail to address the challenges of trust and security in decentralized environments. In contrast, the proposed framework integrates dynamic trust indicators and employs Federated Learning (FL) to ensure privacy while enhancing decision making performance. The framework also incorporates Zero Knowledge Proofp based Verifiable Machine Learning (ZKP-VML), which ensures transparency and security without compromising sensitive data. Through continuous real time trust assessments, the framework adapts to changing conditions, improving the accuracy and reliability of decisions in environments where participants may not fully trust each other. The application of this framework in autonomous vehicles and IoT networks demonstrated its ability to make robust, secure decisions, even in complex and uncertain scenarios. The framework’s ability to incorporate both trust and security into its decision making processes sets it apart from traditional models, which typically do not address the trustworthiness of data or participants. This research highlights the importance of integrating trust and security into machine learning models, particularly in decentralized systems, and offers a robust solution to trust management challenges. However, challenges such as scalability and computational efficiency remain, and future work should focus on enhancing these aspects, along with exploring the framework's applicability in other decentralized domains like finance or supply chain management. The integration of privacy preserving technologies and improvements in adversarial robustness are also potential areas for future research.
Deep neural networks (DNNs) are one of the most effective methods available for detecting smart contract vulner-abilities (SCVs). The performance of current DNN approaches relies heavily on a large number of training samples and labels. The semi-supervised learning (SSL) trains the pseudo-labeling mechanism and performs label propagation on unlabeled data, thereby mitigating this problem. However, these approaches are not suitable for tasks related to Smart Contract Vulnerability Detection (SCVD) due to their limitations in capturing subtle faults, which are the primary causes of SCVD. We believe that subtle faults are contained in the correlation feature, which encompasses both commonalities among vulnerable contracts and differences between vulnerable and non-vulnerable contracts. Therefore, we propose a correlation-driven SSL method called Jupiter to solve this limitation. Specifically, Jupiter incorporates a contrastive learning module that conducts pairwise comparisons between smart contracts and captures correlation features. These correlation features facilitate the separation of feature distributions in vulnerable and non-vulnerable contracts. Then, a support vector machine with a built-in RBF kernel function is utilized to establish a decision boundary between the two types of contract distributions. We employ the distance from data to the decision boundary as a confidence score and propagate high-quality pseudo-labels to each unlabeled data surpassing the threshold. This process is iterated, where the labeled dataset is constructed using a combination of pseudo-labeled and reallabeled data to be re-input into the contrastive learning model until all data is labeled. Finally, the combination of semantic information and correlation features accurately detects SCVs.By conducting an empirical evaluation on a large-scale realworld dataset comprising over 40,000 smart contracts, we compare the performance of 6 state-of-the-art SSL methods and 5 state-of-the-art SCVD methods. Our findings demonstrate the effectiveness of our proposed method, Jupiter, in two key aspects: (1) Jupiter achieves optimal performance across all SSL methods, outperforming them by 18.89% to 28.42% in terms of F1- score; (2) Current state-of-the-art SCVD methods fail to deliver satisfactory results when only small amounts of labeled data are available. Specifically, these methods achieve F1-scores ranging from 49.72% to 50.98% when utilizing just 10% of the labeled data. In contrast, under the same conditions, Jupiter outperforms all baselines, achieving an F1-score of 89.28%, which represents an improvement of 75.12% to 79.56%.
Modern cloud-based AI training relies on extensive telemetry and logs to ensure accountability. While these audit trails enable retrospective inspection, they struggle to address the inherent non-determinism of deep learning. Stochastic operations, such as dropout, create an ambiguity surface where attackers can mask malicious manipulations as natural random variance, granting them plausible deniability. Consequently, existing logging mechanisms cannot verify whether stochastic values were generated and applied honestly without exposing sensitive training data. To close this integrity gap, we introduce Verifiable Dropout, a privacy-preserving mechanism based on zero-knowledge proofs. We treat stochasticity not as an excuse but as a verifiable claim. Our approach binds dropout masks to a deterministic, cryptographically verifiable seed and proves the correct execution of the dropout operation. This design enables users to audit the integrity of stochastic training steps post-hoc, ensuring that randomness was neither biased nor cherry-picked, while strictly preserving the confidentiality of the model and data.
Federated Learning (FL) enables distributed model training while preserving data privacy; however, it remains vulnerable to poisoning attacks and lacks computational integrity. Recent solutions integrating Zero-Knowledge Proofs (ZKPs) and blockchain have successfully established process-level verifiability but suffer from prohibitive computational overhead due to the requirement of generating cryptographic proofs for every local update. To address this efficiency bottleneck, this paper proposes Pro-ZkFL, a reputation-aware probabilistic verification framework. Unlike deterministic approaches that verify every transaction, Pro-ZkFL utilizes Verifiable Random Functions (VRF) on-chain to dynamically select a subset of clients for auditing based on their historical reputation scores. We design a dual-commitment scheme where clients submit lightweight cryptographic commitments for every round but generate heavy ZKPs only when challenged. Experimental results on FashionMNIST and CIFAR-10 demonstrate that Pro-ZkFL reduces computational overhead by approximately 82 % and gas costs by 73 % compared to full-verification baselines while maintaining a 99 % detection rate against persistent adversaries, offering a scalable trade-off between security and efficiency.
Elif Nur Kucur, Tolga Büyüktanır, Muharrem Ugurelli, Kazım Yıldız
Privacy-preserving machine learning (PPML) constitutes a core element of responsible AI by supporting model training and inference without exposing sensitive information. This survey presents a comprehensive examination of the major cryptographic PPML techniques and introduces a unified taxonomy covering technical models, verification criteria, and evaluation dimensions. The study consolidates findings from both survey and experimental works using structured comparison tables and emphasizes that recent research increasingly adopts hybrid and verifiable PPML designs. In addition, we map PPML applications across domains such as healthcare, finance, Internet of Things (IoT), and edge systems, indicating that cryptographic approaches are progressively transitioning from theoretical constructs to deployable solutions. Finally, the survey outlines emerging trends—including the growth of zero-knowledge proofs (ZKPs)-based verification and domain-specific hybrid architectures—and identifies practical considerations that shape PPML adoption in real systems.
Decision-OS V5 Revised (SiriusA2) addresses a practical AI safety problem: how human oversight can prevent irreversible decisions from being executed under pressure, confusion, coercion, or panic. It is designed for safety-critical, non-medical decision support settings where a user may be authenticated, yet the execution path may still be unsafe. The framework proposes a human-in-the-loop confirmation layer for irreversible risk. Instead of allowing a valid user action to move directly into execution, SiriusA2 routes protected actions through auditable confirmation states such as Request, Observe, Hold, Approve, Reject, Stop Candidate, Execute, and Revoke. The core mechanism is a trajectory-aware duress_score. This score is not an intent classifier, diagnosis, truthfulness score, or autonomous veto. It is an operational control-routing signal used to detect when a valid execution path deviates from an ordinary decision trajectory and approaches irreversible harm. SiriusA2 preserves human final consent through two-step confirmation, an explicit revoke path, optional family multisig, and a Zero-Knowledge approval layer (310/320) that verifies authorization without exposing personally identifiable information. The duress_score does not replace ZK approval or multisig; it routes actions into the confirmation path, while ZK qualification and multisig provide independent authorization conditions before irreversible execution. The revised manuscript integrates the SiriusA Adoption Gate into the main paper. The gate provides a Hold-first confirmation path for irreversible, externally pressured, unusually urgent, or high-stake actions: Request → Observe/Hold → Approve/Reject → Execute/Revoke. A score-based Stop Candidate does not automatically become Stop or Freeze. Stop or Freeze requires an independent non-score condition, such as verified revoke input, ZK-qualified m-of-k approval, policy-defined guardian confirmation, or an emergency protocol condition. If no such condition is available, SiriusA intentionally prefers continued Hold and evidence preservation over an AI-only execution veto. This release also clarifies cold-start behavior, causal bridge support, baseline maturity, corrected event terms, disclosure boundaries for calibration-sensitive parameters, and non-PII audit requirements. Safety is operationalized through auditable state transitions, non-PII KPIs, five-line gate outputs, evidence packaging (ZIP + SHA256), and explicit prohibitions on automatic transmission, payment, or reporting. A minimal proof-of-concept gate exists as a public runtime artifact, demonstrating PASS / DELAY / BLOCK routing, conservative severity merging, evidence union, pre-execution checking, and auditable JSON outputs. Deployment-level validation, calibration, and domain-specific robustness remain future work. Gateway / series index:https://github.com/shin4141/decision-os-paper Recommended read order:V5 Revised (SiriusA2) → V6 (PIC) → V8 (v2)Optional: V7 (AGI definition) Related repositories:- V5 Policy Pack / specification and adoption materials: https://github.com/shin4141/paper-public- Gate engine / MMAR-L0: https://github.com/shin4141/mmar-l0-core- SiriusA core runtime: https://github.com/shin4141/siriusA-core SSOT:GitHub repository “decision-os-paper”.This PDF corresponds to the revised SiriusA2 release candidate committed to the SSOT repository. Release note:This revised release integrates the trajectory-aware duress_score definition, SiriusA Adoption Gate, ZK qualification layer, family multisig, independent non-score Stop conditions, non-PII audit structure, V4-compatible escalation interface, and proof-of-concept gate positioning into the main paper. Transparency / Author’s Note:https://github.com/shin4141/decision-os-paper/blob/main/AUTHORS_NOTE.md
Healthcare AI needs large, diverse datasets, yet strict privacy and governance constraints prevent raw data sharing across institutions. Federated learning (FL) mitigates this by training where data reside and exchanging only model updates, but practical deployments still face two core risks: (1) privacy leakage via gradients or updates (membership inference, gradient inversion) and (2) trust in the aggregator, a single point of failure that can drop, alter, or inject contributions undetected. We present zkFL-Health, an architecture that combines FL with zero-knowledge proofs (ZKPs) and Trusted Execution Environments (TEEs) to deliver privacy-preserving, verifiably correct collaborative training for medical AI. Clients locally train and commit their updates; the aggregator operates within a TEE to compute the global update and produces a succinct ZK proof (via Halo2/Nova) that it used exactly the committed inputs and the correct aggregation rule, without revealing any client update to the host. Verifier nodes validate the proof and record cryptographic commitments on-chain, providing an immutable audit trail and removing the need to trust any single party. We outline system and threat models tailored to healthcare, the zkFL-Health protocol, security/privacy guarantees, and a performance evaluation plan spanning accuracy, privacy risk, latency, and cost. This framework enables multi-institutional medical AI with strong confidentiality, integrity, and auditability, key properties for clinical adoption and regulatory compliance.
Aaron Chan, Alex Ding, Frank Sicong Chen, Alan Wu · 6 authors
The rapid integration of Large Language Models (LLMs) into decentralized physical infrastructure networks (DePIN) is currently bottlenecked by the Verifiability Trilemma, which posits that a decentralized inference system cannot simultaneously achieve high computational integrity, low latency, and low cost. Existing cryptographic solutions, such as Zero-Knowledge Machine Learning (ZKML), suffer from superlinear proving overheads (O(k NlogN)) that render them infeasible for billionparameter models. Conversely, optimistic approaches (opML) impose prohibitive dispute windows, preventing real-time interactivity, while recent "Proof of Quality" (PoQ) paradigms sacrifice cryptographic integrity for subjective semantic evaluation, leaving networks vulnerable to model downgrade attacks and reward hacking. In this paper, we introduce Optimistic TEE-Rollups (OTR), a hybrid verification protocol that harmonizes these constraints. OTR leverages NVIDIA H100 Confidential Computing Trusted Execution Environments (TEEs) to provide sub-second Provisional Finality, underpinned by an optimistic fraud-proof mechanism and stochastic Zero-Knowledge spot-checks to mitigate hardware side-channel risks. We formally define Proof of Efficient Attribution (PoEA), a consensus mechanism that cryptographically binds execution traces to hardware attestations, thereby guaranteeing model authenticity. Extensive simulations demonstrate that OTR achieves 99% of the throughput of centralized baselines with a marginal cost overhead of $0.07 per query, maintaining Byzantine fault tolerance against rational adversaries even in the presence of transient hardware vulnerabilities.
Frequent digitalization of power systems has increased the areas of cyber attack on smart grids that require more developed security frameworks capable of responding to cyber and physical threats. The paper is a detailed discussion on the implementation of Artificial Intelligence (AI) and Block chain in improving the cyber security of smart grids. To prevent new cyber threats, we suggest a new hybrid security model, which uses AI-based anomaly detection and decentralized integrity checks, based on block chains. The architecture is a fusion of machine learning based real time threat detection and immutable block chain ledgers to keep transactions safe and under control. We prove with the help of the performance evaluation and case studies that the model is efficient in the detection of false data injection, distributed denial of service (DDOS) attacks, and other advanced threats. The paper also cites major issues such as the scalability, interoperability, and computational overhead, as well as outlining opportunities of future research in resilient smart grid infrastructure. We have found out that the AI-block chain implementation has the potential to enhance the accuracy of threat detection up to 30 percent relative to traditional approaches as well as integrity and transparency of data management in smart grids.
Large Language Models (LLMs) have demonstrated significant potential in smart contract auditing. However, they are still susceptible to hallucinations and limited context-aware reasoning. In this article, we propose SmartAuditFlow, a dynamic Plan-Execute framework that customizes audit strategies based on the unique characteristics of each smart contract. Unlike static, rule-based workflows, our approach iteratively generates and refines audit plans in response to intermediate outputs and newly detected vulnerabilities. To improve reliability, the framework incorporates structured reasoning, prompt optimization, and external tools such as static analyzers and Retrieval-Augmented Generation (RAG). This multi-layered design reduces false positives and enhances the accuracy of vulnerability detection. Experimental results show that SmartAuditFlow achieves 100% accuracy on common vulnerability benchmarks and successfully identifies 13 additional CVEs missed by existing methods. These findings underscore the framework’s adaptability, precision, and practical utility as a robust solution for automated smart contract security auditing. The source code is available at: https://github.com/JimmyLin-afk/SmartAuditFlow .
Mohammed Alhayani, Wisam Hazım Gwad, Shahab Wahhab Kareem, Moustafa Fayad
The increasing use of blockchain smart contracts has introduced new security challenges, as small coding errors can lead to major financial losses. While rule-based static analyzers remain the most common detection tools, their limited adaptability often results in false positives and outdated vulnerability patterns. This study presents a comprehensive comparative analysis of machine learning (ML) and deep learning (DL) methods for smart contract vulnerability detection using the BCCC-SCsVuls-2024 benchmark dataset. Six models (Random Forest, k-Nearest Neighbors, Simple and Deep Multilayer Perceptron, and Simple and Deep one-dimensional Convolutional Neural Networks) were evaluated under a unified experimental framework combining RobustScaler normalization and Principal Component Analysis (PCA) for dimensionality reduction. Our experimental results from a five-fold cross-validation show that the Random Forest classifier achieved the best overall performance with an accuracy of 89.44% and an F1-score of 93.20%, outperforming both traditional and neural models in stability and generalization. PCA-based feature analysis revealed that opcode-level features, particularly stack and memory manipulation instructions (PUSH, DUP, SWAP, and RETURNDATASIZE), were the most influential in defining contract behavior.
High-mobility Internet of Things (IoT) for Vehicle-to-Grid (V2G) Demand Response (DR), including roaming between Charge Point Operators (CPOs), requires privacy-preserving authentication with sub-1 ms responses and cross-domain scalability as devices exceed 200km/h. Mechanisms must run on constrained hardware while remaining compatible with EV-charging message flows such as ISO 15118–20 and OCPP 2.0.1. Many deployed schemes re-authenticate from scratch, which inflates computation and airtime; static credentials also ignore trajectory context and struggle with rapid mobility. We present a Zero-Knowledge Proof-based Authentication Scheme (ZKPAS) for V2G/DR that proves possession without disclosure and replaces heavy handshakes with compact, mobility-aware proofs, targeting latencyLO(n) toO(logn). (iii) Predictive token generation with Long Short-Term Memory (LSTM) models trained on GeoLife and T-Drive pre-computes material, yielding 84.7% token reuse along trajectories. (iv) Cross-domain authentication employs (t,n)-threshold cryptography for Byzantine-tolerant roaming across operators. We prove resistance to impersonation, replay, man-in-the-middle, and trajectory inference; under the Computational Diffie–Hellman Problem (CDHP), the adversary’s success probability satisfies Pr[break] ≤ 2−λ. On real transportation topologies, ZKPAS cuts computation by 71.8%, authentication latency by 93.9%, and energy by 69.5%, while interfacing with V2G/DR control flows. The protocol sustains a 98.5% authentication success rate at 250km/h.
AI development requires reliable datasets, yet today’s data supply chains face challenges in traceability, authenticity and ethical compliance. This study introduces a blockchain-assisted data integrity framework that ensures transparent and verifiable provenance for AI model training. The proposed system uses smart contracts to record data lineage, ownership, preprocessing transformations and annotation events. IPFS-based off-chain storage reduces blockchain load while ensuring immutability. A verification engine allows auditors to evaluate dataset compliance with ethical and regulatory standards, including bias mitigation and consent validation. Experiments utilized three real AI workflows: medical imaging, sentiment analysis and environmental sensor classification. Findings show a 92 percent reduction in provenance disputes and an improvement in audit efficiency by 41 percent. The system also provides tamper-resistant documentation supporting responsible AI governance. Latency tests show minimal performance impact due to parallelized validation nodes. This research demonstrates that blockchain can provide a robust backbone for ethical AI ecosystems, where transparency and trust are critical. Future work will explore confidentiality enhancements using zero-knowledge proofs.
As artificial intelligence becomes deeply embedded in critical sectors like finance and medicine, we face a pressing challenge: how to guarantee its integrity. At the heart of this issue is a conflict between the proprietary nature of AI models, which are valuable assets, and the growing need for transparency in their operations. This paper lays out an architectural blueprint that resolves this tension by bringing together blockchain technology and Zero-Knowledge Proofs (ZKPs). We show how it's possible to verifiably confirm that an AI model has run correctly without exposing any of its confidential internal parameters. We walk through a simulation where a Deep Neural Network (DNN) produces an inference, and a ZKP is generated to prove the calculation used the legitimate model weights. This proof, along with the public data, is then recorded on a decentralized ledger, creating a permanent, auditable trail. A key part of our work is a comparison of two major ZKP technologies, zk-SNARKs and zk-STARKs, where we break down their respective trade-offs. Our simulation's effectiveness is demonstrated through resilience testing; it successfully identified and rejected 100% of fraudulent attempts, including both tampered outputs and counterfeit models. This demonstrates the architecture's efficiency in creating a provably secure and auditable trail, lighting a path toward genuinely trustworthy AI.
The need for strong cybersecurity frameworks in the healthcare industry has increased due to the Internet of Medical Things (IoMT) devices and electronic health records (EHRs) exponential growth. This paper presents HealthSentinel-ZKP, a novel framework that leverages federated neuro-symbolic deep learning and zero-knowledge blockchain to secure electronic health data. In contrast to earlier models, HealthSentinel-ZKP combines CNNs, Transformers, and symbolic reasoning for multi-perspective intrusion detection. Federated learning and zero-knowledge proof mechanisms protect data privacy. Using an immutable ZKPenhanced blockchain, the system guarantees GDPR-compliant auditing and has a dual-stream anomaly detection architecture. HealthSentinel-ZKP is a next-generation healthcare cybersecurity paradigm, as demonstrated by experimental results on benchmark datasets that demonstrate superior performance in zero-day attack detection and privacy preservation.
Vivekrabinson K, C Yogesh, Indra Kumar M, Hari Suriya K · 6 authors
The rapid progress of generative AI has already seen the rise of highly realistic artificial and deepfake content that has created a problems related to trust on information, privacy, issues related to cyber security loss of general trust. This research introduces TrueSightQ, a unified full stack web application framework through with multimodal detection of AI generated content through use of Quatnum enhancement and Web3 Integration. This system is a hybrid between heuristic and deep learning methods with the added feature of GPU accelerated training combined with quantum advantage classifiers. The trustworthiness and transparency of blockchain technology, as well as IPFS storage and Ethereum Smart contract to improve verification process. The model is further enhanced with modality wise fusion and, decentralized trust based mechanisms for the defense mechanism to adversarial attack. Results show that in general TrueSightQ significantly outperforms standard unimodal detectors overall, with additional gains to verifiability, precision and interpretability demonstrating how the multimodal and decentralized methodology within the model mitigates the issues of AI generated content very efficiently.
Vinoth Punniyamoorthy, Ashok Gadi Parthi, Mayilsamy Palanigounder, Ravi Kiran Kodali · 6 authors
Distributed machine learning systems require strong privacy guarantees, verifiable compliance, and scalable deployment across heterogeneous and multi-cloud environments. This work introduces a cloud-native privacy-preserving architecture that integrates federated learning, differential privacy, zero-knowledge compliance proofs, and adaptive governance powered by reinforcement learning. The framework supports secure model training and inference without centralizing sensitive data, while enabling cryptographically verifiable policy enforcement across institutions and cloud platforms. A full prototype deployed across hybrid Kubernetes clusters demonstrates reduced membership-inference risk, consistent enforcement of formal privacy budgets, and stable model performance under differential privacy. Experimental evaluation across multi-institution workloads shows that the architecture maintains utility with minimal overhead while providing continuous, risk-aware governance. The proposed framework establishes a practical foundation for deploying trustworthy and compliant distributed machine learning systems at scale.
Nur Shahidah Mohammad, Md. Mafizur Rahman, Md. Abdur Razzaque
Federated Learning (FL) enables multiple entities to collaboratively train models without sharing sensitive data, but it faces critical privacy, security, and efficiency challenges in healthcare intrusion detection systems. These issues are intensified by adversarial attacks, non-IID data, and the need for real-time performance. Existing FL methods struggle with gradient inversion, model poisoning, Sybil attacks, and high computational overhead, limiting their effectiveness in secure and scalable healthcare applications. This work proposes the PrivacyPreserving Personalized Federated Learning Intrusion Detection in Healthcare applications (P3FL-HIDS), integrating Byzantinerobust aggregation, gradient masking, and Zero-Knowledge Proof based authentication. Key features include strong adversarial resilience, protection of privacy against gradient inversion, personalized model adaptation for heterogeneous data, and secure participant authentication. Additional contributions include a dual-network training approach, adaptive clustering for personalization, and optimized secure communication for real-time healthcare scenarios. Experimental results on a Brain Tumor magnetic resonance imaging (MRI) dataset show that P3FLHIDS outperforms state of the art works in terms of accuracy, resilience, and resistance.
Uisang Lee, Changhoon Chung, Junmo Lee, Sung Jun Moon
The rapid growth of Ethereum has made it more important to quickly and accurately detect smart contract vulnerabilities. While machine-learning-based methods have shown some promise, many still rely on rule-based preprocessing designed by domain experts. Rule-based preprocessing methods often discard crucial context from the source code, potentially causing certain vulnerabilities to be overlooked and limiting adaptability to newly emerging threats. We introduce BugSweeper, an end-to-end deep learning framework that detects vulnerabilities directly from the source code without manual engineering. BugSweeper represents each Solidity function as a Function-Level Abstract Syntax Graph (FLAG), a novel graph that combines its Abstract Syntax Tree (AST) with enriched control-flow and data-flow semantics. Then, our two-stage Graph Neural Network (GNN) analyzes these graphs. The first-stage GNN filters noise from the syntax graphs, while the second-stage GNN conducts high-level reasoning to detect diverse vulnerabilities. Extensive experiments on real-world contracts show that BugSweeper significantly outperforms all state-of-the-art detection methods. By removing the need for handcrafted rules, our approach offers a robust, automated, and scalable solution for securing smart contracts without any dependence on security experts.
Basil Hanafi, Mohammad Ubaidullah Bokhari, Mudasir Ahmad Wani, Kashish Ara Shakil · 5 authors
Various types of research are being carried out to advance in the field of cryptography and develop a more robust technique for security. Adversarial neural cryptography (ANC) is a recent development in this extension, which possesses huge potential to be implemented in various domains. There is a continuous need for the development of more adaptive techniques to secure data while in communication using deep learning and other applied artificial intelligence techniques, which serves as the motivation for this work stems from the increasing need for adaptive, robust encryption mechanisms to address the limitations of traditional cryptographic techniques in securing sensitive blockchain transactions. This article proposes a new approach for the protection of private smart contracts on blockchain systems via ANC. The proposed method in the research is dynamic adversarial training using three neural networks to secure smart contract transactions. It optimizes the encryption and decryption processes against evolving cyber threats. The algorithm strives to attain a high key agreement rate (KAR) and a lower Eve’s decryption failure rate (EDFR) eventually proving its efficacy in attaining privacy and security in blockchain applications and adaptability. This research will incite more studies on ANC and its practical implementations in ensuring private smart contracts and overcoming the present cryptographical approaches with significant development because of their potential.
Mohammad M Maheri, Sunil Cotterill, Alex Davidson, Hamed Haddadi
Machine unlearning aims to remove the influence of specific data points from a trained model to satisfy privacy, copyright, and safety requirements. In real deployments, providers distribute a global model to many edge devices, where each client personalizes the model using private data. When a deletion request is issued, clients may ignore it or falsely claim compliance, and providers cannot check their parameters or data. This makes verification difficult, especially because personalized models must forget the targeted samples while preserving local utility, and verification must remain lightweight on edge devices. We introduce ZK APEX, a zero-shot personalized unlearning method that operates directly on the personalized model without retraining. ZK APEX combines sparse masking on the provider side with a small Group OBS compensation step on the client side, using a blockwise empirical Fisher matrix to create a curvature-aware update designed for low overhead. Paired with Halo2 zero-knowledge proofs, it enables the provider to verify that the correct unlearning transformation was applied without revealing any private data or personalized parameters. On Vision Transformer classification tasks, ZK APEX recovers nearly all personalization accuracy while effectively removing the targeted information. Applied to the OPT125M generative model trained on code data, it recovers around seventy percent of the original accuracy. Proof generation for the ViT case completes in about two hours, more than ten million times faster than retraining-based checks, with less than one gigabyte of memory use and proof sizes around four hundred megabytes. These results show the first practical framework for verifiable personalized unlearning on edge devices.
Anurag Anand Duvey, Chandrashekhar Goswami, Amit Kumar Goel
Federated Learning (FL) gives opportunity to decentralized model training without the raw data's revealing. But in actual real-world implementation faces certain number of challenges. These include trust in client updates, verifiable end-to-end privacy promises, equitable contributor compensation, and accountable aggregation. This paper provides a design-first architecture that addresses these issues by integrating concise zero-knowledge proofs (zk-SNARKs) for trustless verification with smart-contract arrangements. This approach is for verification, secure aggregate pooling, and reward settlement. Our design is consisting a structure of five-layer stack, named as Client, Proof, Blockchain, Incentive, and Governance. It highlights end-to-end workflows for the generation of proof for updates, verifying them on chain, anchor-off chain aggregation sequence anchoring, and allocate contribution-matching token payouts. We specify clearly smart-contract interfaces called as aggregation, registry, incentivization, zk-circuit objectives, and several scaling controls like proofs/aggregated in batch or Layer-2/rollup rollouts and the anchor-Merkel. The design also implements Shapley-like contribution measures and ERC-compact reward settlement. It incorporates Sybil-resistance and vesting primitives to lean against game. We define a crisp threat model, discuss security and privacy trade-offs, and suggest evaluation using healthcare and IoT benchmarks. These assess the learning utility, resilience to poisoning, fairness of the payouts, and system costs in terms of gas fee. Moreover, by training and incentivization by smart contracts at design level, we create a foundation for future prototyping and rigorous empiric testing. This sets a realistic path from the design framework to effective workable, end-to-end Privacy-preserving Federated Learning deployment.
Federated Learning (FL) enables multiple participants to collaboratively train a shared model without sharing their private data. However, FL remains vulnerable to malicious clients submitting incorrect updates to disrupt training. To address this, we formalize each client’s local training step as a Nondeterministic Polynomial-time (NP) statement, verifiable via zero-knowledge proofs (ZKPs) at every round. We propose Jolt-FL, the first general-purpose verifiable FL framework that immediately detects and excludes malicious clients upon their first dishonest action – without relying on heuristics, statistical assumptions, or multi-round analysis. Built on Jolt’s zkVM, a state-of-the-art zero-knowledge virtual machine (zkVM) developed by a16zcrypto, Jolt-FL guarantees training integrity and data privacy without trusted hardware or third-party intermediaries. By witnessing every computation step, it defends against a wide range of attack vectors, securely filtering dishonest updates even if up to 50% of clients are malicious, while preserving convergence and final model performance. To demonstrate feasibility, we implement a prototype featuring a complete end-to-end Convolutional Neural Network (CNN) for image classification using the MNIST dataset. To our knowledge, this is the first fully verifiable end-to-end CNN training under ZKPs without any custom circuit design. Our solution achieves competitive proof generation times, compact proof sizes, and low verification costs–all while preserving model accuracy on par with standard FL.