The intersection of Artificial Intelligence (AI) and distributed systems has given rise to Federated Learning (FL), a paradigm that enables decentralized model training without compromising local data privacy. As organizational data silos grow, deploying complex machine learning models across highly distributed edge networks becomes a critical infrastructural challenge. Standard FL implementations suffer from severe vulnerabilities related to adversarial gradient updates and computational bottlenecks at the aggregation layer. This paper presents a novel, end-to-end distributed architecture that hardens FL pipelines using advanced cryptographic verification and optimized big data processing frameworks. We introduce a Zero-Knowledge Proof (ZKP) wrapper that cryptographically validates node computations before global aggregation, neutralizing model poisoning attacks without inspecting raw gradients. Additionally, we evaluate the system's performance using extreme gradient boosting models optimized for distributed edge execution. We formalize the mathematical transformation of the machine learning loss functions into Rank-1 Constraint Systems (R1CS) suitable for succinct verification. Extensive experimental results demonstrate that our hybrid architecture achieves a 94.2\% accuracy retention under adversarial conditions while maintaining scalable throughput across 1,000 parallel distributed nodes, effectively bridging the gap between rigorous cryptographic security and high-performance distributed AI.
AI hallucination is a cost problem, not a knowledge problem. This paper documents that three sentences of prompt-level instruction — IDK+COMP: a compression mandate paired with a refusal permission — reproduce hallucination suppression matching or exceeding a full multi-constraint methodology across three frontier AI models. Gemini: 6.3% hallucination rate (Baseline 57.5%). ChatGPT: 0.0% (Baseline 22.2%). Claude: 0.0%. The paper establishes hallucination as a utility-maximizing response to a cost structure that makes confident invention cheaper than refusal. Change the cost structure at the prompt level — without touching the model, without retraining, at near-zero cost — and the behavior changes. IDK is load-bearing. The compression mandate is the environment in which it operates. Secondary findings: hedging is not a mitigation — it is a co-symptom of unresolved uncertainty, and this dataset moves the hedge-hallucination relationship in both directions depending on directive design. Plausibility-trap strings (SPLAM, Vandermeer Effect) expose the limit of cost-structure interventions: the model cannot recognize the unrecognizable. In a plausibility-trap domain, IDK+COMP is worse than nothing. 410 trials. Three frontier AI models. Five governance conditions. Proof-of-concept dataset; results are directional. Companion resources: Kowalski et al. (2026a), A Puma in a Teacup: Signal Quality and Hallucination Suppression Through Prompt-Level Incentive Restructuring. https://doi.org/10.5281/zenodo.19502460 Kowalski, M. M. and Claude (Anthropic). (2026). Taxonomy of AI Bullshit: hallucination and hedging subcategories. Zenodo. https://doi.org/10.5281/zenodo.20631337. Kowalski, M. M. & Claude (Anthropic). (2026). Hallucination Test Suite and Execution Records: test strings, activation blocks, trial data and AI transcripts. Zenodo. https://doi.org/10.5281/zenodo.21325014.
This paper argues that trust scores — from credit ratings and ESG scores to AI-generated trust metrics — fail not because of poor implementation, but because trust itself is the wrong abstraction. Trust is not a scalar quantity but a contextual, relational, and topological phenomenon. Any attempt to reduce it to a universal numerical score leads to fragility, manipulation, exclusion, and systemic failure. We identify five structural failure modes (context collapse, Goodhart's Law, epistemic centralization, irreversibility, and metric substitution for truth), supported by historical case studies (Enron, Wirecard, Volkswagen Dieselgate, the 2008 subprime crisis, ESG rating failures). A formal impossibility argument demonstrates that no universal trust score can simultaneously satisfy context independence, temporal stability, observer neutrality, and manipulation resistance. We propose proof-based systems as the alternative paradigm, where trust is not measured but rendered unnecessary through local, irreversible verification. Examples include Bitcoin Proof-of-Work, zero-knowledge proofs, and blockchain-based supply chain traceability.
We describe a protocol that produces byte-identical outputs from frontier-scale language model inferenceand binds each output to a portable, offline-verifiable signed receipt. The construction has three parts.First, an inference substrate that runs models up to seventy-two billion dense parameters and forty-sevenbillion mixture-of-experts active parameters on NVIDIA H100, with cross-vendor extension to AMDInstinct MI300X. Output hashes match byte-for-byte across fresh process launches in every configurationmeasured; at single-GPU bf16 with eager attention the AMD and NVIDIA hashes are themselves byte-identical, including over fifty-one tokens of compounding frontier-scale generation, and at two-GPUtensor-parallel they differ as predicted by the underlying NCCL-ring versus RCCL-fabric all-reducetopology. Both are individually deterministic. Second, a canonical CBOR receipt schema with an Ed25519signature over a domain-separated message, implemented in Go, Python, and Rust, with cross-languagebyte-identity verified end-to-end and AMD-produced receipts verifying byte-for-byte through a Rustverifier built on x86 NVIDIA hardware. Third, a probabilistic spot-check verifier that re-executes asmall sample of receipts and rejects on mismatch; we prove a soundness lemma of the form 1−(1−f )kand validate it empirically across seventy adversary-verifier configurations with seven hundred thousandMonte Carlo trials. Verification costs about eighty microseconds per receipt on a single core. Eleventhousand sequential warm-model inferences ran without a single byte-identity failure. The contribution isthe construction itself: a primitive that gives issuer-independent fabrication soundness for AI inference atproduction cost, without a hardware-vendor dependency and without zero-knowledge proofs.
Large Reasoning Models (LRMs) and Multi-Agent Systems (MAS) in high-stakes domains demand reliable verification, yet centralized approaches suffer four limitations: (1) Robustness, with single points of failure vulnerable to attacks and bias; (2) Scalability, as reasoning complexity creates bottlenecks; (3) Opacity, as hidden auditing erodes trust; and (4) Privacy, as exposed reasoning traces risk model theft. We introduce TRUST (Transparent, Robust, and Unified Services for Trustworthy AI), a decentralized framework with three innovations: (i) Hierarchical Directed Acyclic Graphs (HDAGs) that decompose Chain-of-Thought reasoning into five abstraction levels for parallel distributed auditing; (ii) the DAAN protocol, which projects multi-agent interactions into Causal Interaction Graphs (CIGs) for deterministic root-cause attribution; and (iii) a multi-tier consensus mechanism among computational checkers, LLM evaluators, and human experts with stake-weighted voting that guarantees correctness under 30% adversarial participation. We prove a Safety-Profitability Theorem ensuring honest auditors profit while malicious actors incur losses. All decisions are recorded on-chain, while privacy-by-design segmentation prevents reconstruction of proprietary logic. Across multiple LLMs and benchmarks, TRUST attains 72.4% accuracy (4-18% above baselines) and remains resilient against 20% corruption. DAAN reaches 70% root-cause attribution (vs. 54-63% for standard methods) with 60% token savings. Human studies validate the design (F1 = 0.89, Brier = 0.074). The framework supports (A1) decentralized auditing, (A2) tamper-proof leaderboards, (A3) trustless data annotation, and (A4) governed autonomous agents, pioneering decentralized AI auditing for safe, accountable deployment of reasoning-capable systems.
The Model State Attestation Framework (MSAF) is the theoretical foundation of the Auburn Governance Stack. It establishes the intellectual authority for why cryptographic AI attestation is necessary, defines the three-tier attestation architecture, surveys over 200 papers from 2022 through 2026 confirming that no existing system composes hardware trust, continuous model health monitoring, and supply chain provenance into a unified attestation artifact, and documents the theoretical ceilings that bound what attestation can and cannot guarantee. The framework addresses a fundamental governance paradox: organizations deploying foundation models into critical infrastructure cannot definitively demonstrate that a specific output was generated by a model version in a compliant internal state. Current governance artifacts including model cards, training logs, and static benchmarks are descriptive rather than prescriptive and fundamentally non-verifiable. MSAF composes three verification tiers into a layered architecture. The first tier provides TEE-backed platform attestation for execution environment integrity, proving the silicon is real and uncompromised. The second tier provides signed continuous metrics for training and inference health monitoring, proving the model is healthy at the moment of inference. The third tier provides selective zero-knowledge proofs for high-value inference verification, enabling verification without exposing proprietary model internals. The composed attestation artifact is mapped to specific evidence requirements across five regulatory frameworks: EU AI Act Articles 11 and 15, FDA Software as a Medical Device Pre-Determined Change Control Plans, Federal Reserve SR 11-7 ongoing monitoring requirements, US Federal Procurement mandates, and AI insurance underwriting. The framework identifies the theoretical ceilings that bound attestation: Rice's theorem establishing the undecidability of semantic properties, the Impossibility Sandwich bounding what verification can achieve between Rice's theorem and Gödel's incompleteness, and TEE physical side-channel vulnerabilities limiting hardware trust guarantees. The honest framing: MSAF provides probabilistic risk reduction and accountability infrastructure, not behavioral safety guarantees. This is analogous to financial auditing, which certifies process compliance without guaranteeing future solvency. The building blocks exist individually. MSAF is the composition that binds them into an enforceable governance architecture. This work was previously hosted on Figshare, where the author maintained a portfolio of 29 publications with minted DOIs and an established ORCID record. The author's Figshare account was disabled without prior notice, without citation of a specific terms violation, and without opportunity for review, rendering all published items and their associated DOIs inaccessible. No communication was provided before or at the time of the disable action. This deposit and associated deposits on Zenodo ensure continued public accessibility of the author's research on institutional infrastructure with appropriate permanence guarantees.
Open access
2 source records
Adversarial Robustness in Machine Learning
Security and Verification in Computing
Physical Unclonable Functions (PUFs) and Hardware Security
Abstract Federated learning (FL) has emerged as a paradigm-shifting approach to distributed machine learning, enabling multiple participants to collaboratively train models without exposing raw data. However, conventional federated learning architectures remain susceptible to a broad spectrum of security and privacy threats, including model poisoning, gradient inversion, inference attacks, and Byzantine faults. This paper presents a unified and robust framework— Secure and Privacy-Preserving Federated AI (SPFA) — that integrates differential privacy, homomorphic encryption, secure multi-party computation, Byzantine fault tolerance, and zero-knowledge proofs into a unified, production-grade architecture. We formally analyze the threat model, prove privacy guarantees under the ( ε , δ )-differential privacy framework, and demonstrate Byzantine resilience under partial adversarial participation. Extensive experiments on heterogeneous data distributions across image classification, natural language processing, and medical diagnosis benchmarks demonstrate that SPFA achieves model accuracy within 2.3% of centralized baselines while providing provable ε = 1.0 privacy with a communication overhead of only 18% above standard FedAvg. To the best of our knowledge, our framework is among the first to consolidate all five protection layers into a unified, deployable system with formal analysis and an open-source reference implementation. The relevance of SPFA extends to privacy-sensitive applications in healthcare, cybersecurity, distributed edge computing, and smart city analytics.
The widespread deployment of neural networks has raised significant concerns regarding the integrity and privacy of model predictions, especially in malicious environments. Current approaches have explored zero-knowledge proofs for integrity verification. However, they suffer from inefficiency in proving runtime and a lack of rigorous integrity verification for non linear operations. To address these issues, we present a trustwor thy framework for Enhancing Integrity Verification of Convolutional Neural Network predictions (EIV-CNN) in a malicious model, whose key contributions are an efficient optimized sum check protocol and a robust enhanced verification mechanism. Specifically, we first propose an algorithm that enables efficient proving of both batch and collaborative CNN predictions by com bining sumcheck claims of multiple matrix multiplications into one. Moreover, we introduce a non-interactive sumcheck protocol with malicious security (NM-Sumcheck) to serve as a building block for publicly verifying matrix multiplication operations. Furthermore, we introduce a verifiable method for transforming nonlinear operations into matrix operations, enabling their sub sequent evaluation with the NM-Sumcheck protocol. Our EIV CNN provides malicious security, guarantees public verifiability, and preserves model privacy. Empirical results demonstrate that our sumcheck framework achieves constant prover time, verifier time, and proof size. Compared to the state-of-the-art, it achieves up to a 128.56× reduction in prover time, along with significant reductions in communication overhead and enhanced scalability.
This paper introduces ZKAP (Zero-Knowledge Audit Protocol), a cryptographic protocol in which a machine-learning inference is coupled to a zero-knowledge proof of constraint satisfaction in such a way that the output of the inference cannot be released on any observable channel unless the proof verifies. The protocol rests on two principal technical contributions: The Certified Stack — a composite object that cryptographically binds the model weights, a mandatory bit-integrity policy (integer quantisation), the inference runtime stack and the hardware configuration under a single commitment (RootHash), which is included as a public input to every proof produced by the system. Prove-before-output enforcement — a family of four embodiments (hardware output gate, trusted-execution-environment release path, silicon-level embodiment, syscall-intercepting software runtime) in which the output of the inference is physically blocked from leaving the system until the accompanying proof has verified against the Certified Stack commitment and against a constraint set signed by an external authority. Three supporting mechanisms are described: authority-signed formal constraints with a five-type taxonomy, a per-inference hash chain with external anchoring, and a three-party cryptographic separation of duties. The paper further discusses practical relaxations of the strict release semantics for real-time workloads. A soundness theorem is stated against a polynomial-time adversary controlling the operator of the inference, and regulatory implications for high-risk AI systems under Regulation (EU) 2024/1689 (Artificial Intelligence Act) are discussed. Priority declaration. The inventive mechanisms described in this paper are the subject of Bulgarian patent applications BG/P/2026/114317 (filed 30 March 2026) and PTBG202600000316742 (filed 12 April 2026). This preprint establishes the academic priority of the author, as of the date of the deposit timestamp, over the specific technical constructions described in Sections 3 and 4 of the paper. Access note. This deposit is under embargo until 31 March 2027. During the embargo, metadata (title, abstract, author, keywords, references) are publicly visible; the full text is not publicly accessible. The embargo date coincides with the Paris Convention priority deadline of the underlying Bulgarian patent application BG/P/2026/114317. The DOI assigned at deposit time serves as a timestamp for academic priority purposes, independent of the file's accessibility.
Introduction The proliferation of dApps is increasing the attack surface for exploitable vulnerabilities in smart contracts, and thus there is a need for verifiable detection methodologies. Methods In this work, we propose a machine learning framework with blockchain integration for explainable and note that “explainable” implies “verifiable” smart contract vulnerability detection. The SmartBugs-curated data was systematically pre-processed with metadata filtering, feature correlation analysis and encoding for model evaluation. Four ensemble learning methods, Random Forest, XGBoost, LightGBM and CatBoost were tested under identical experimental settings for comparison. Results The Random Forest classifier initially achieved the best balance in terms of stability and performance with an accuracy of 87.67%, successfully detecting important vulnerability classes such as re-entrancy, unchecked low-level calls, etc. To enhance the applicability of our blockchain-based machine learning framework for vulnerable smart contract analysis we extend it from the initial 143-contract dataset SmartBugs-Curated to evaluate it on on large-scale set, namely, SmartBugs-Wild which contains 47,398 real-world Ethereum contracts. Based on 29 static contract-level features, unsupervised clustering (k = 4, silhouette score = 0.3735) identifies discrete structural archetypes present in the dataset. Ensemble classifiers (such as XGBoost, CatBoost, Random Forest and LightGBM) can get excellent discriminative performance on these cluster labels: LightGBM achieves 99% accuracy and 0.98918 macro-F1. Discussion The additional results show that the approach scales, is robust and leads to stable models, even if interpretable. After injecting SHAP-based explainability, the interpretability and predictive power of CatBoost became similar to those of Random Forest. In order to guarantee end-to-end trust and traceability of our optimised classifier, this was linked to a blockchain oracle that independently store the outcomes as well as confidence scores for predictions directly onto an Ethereum-compatible ledger through a Vulnerability Registry smart contract. This integration provides the data is immutable, auditable and transparent in reporting.
Smart contracts extended blockchain functionality beyond simple transactions, powering complex applications like decentralized finance (DeFi). However, this complexity introduces serious security challenges, including price manipulation and inflation attacks. Despite the development of various security tools, the rapid rise in financially motivated exploits continues to pose a significant threat to the blockchain ecosystem. These financially motivated exploits often stem from Monetarily Exploitable Vulnerabilities (MEVuls), which refer to vulnerabilities arising from exploitable implementations in monetary transactions or value-transfer logic. Due to their complexity, intricate chains of function calls, multifaceted logic, and diverse manifestations across different smart contracts, MEVuls are particularly challenging for current security tools to identify. Instead of providing actionable insights, existing tools frequently generate excessive warnings that overwhelm developers without effectively mitigating risks. To address the challenge of recognizing MEVuls, we first formalize MEVuls based on common real-world financial exploits. Then, we introduce FAUDITOR, a specialized fuzzer designed to detect MEVuls in smart contracts. The key insight is that leveraging smart contracts' finance-related interfaces directly exposes critical vulnerabilities, making detection more targeted. We further integrate auditors' reports using NLP to extract valuable insights on exploitation patterns, enabling a more informed search strategy. Additionally, FAUDITOR employs a self-learning mechanism that refines its detection strategies over time, allowing it to improve based on prior fuzzing results. In our evaluation, FAUDITOR impressively reveals 220 zero-day MEVuls. Meanwhile, compared to existing fuzzers, FAUDITOR detects vulnerabilities faster and achieves better instruction coverage.
Decentralized Autonomous Organizations (DAOs) are inclined explore Small Language Models (SLMs) as edge-native constitutional firewalls to vet proposals and mitigate semantic social engineering. While scaling inference-time compute (System 2) enhances formal logic, its efficacy in highly adversarial, cryptoeconomic governance environments remains underexplored. To address this, we introduce Sentinel-Bench, an 840-inference empirical framework executing a strict intra-model ablation on Qwen-3.5-9B. By toggling latent reasoning across frozen weights, we isolate the impact of inference-time compute against an adversarial Optimism DAO dataset. Our findings reveal a severe compute-accuracy inversion. The autoregressive baseline (System 1) achieved 100% adversarial robustness, 100% juridical consistency, and state finality in under 13 seconds. Conversely, System 2 reasoning introduced catastrophic instability, fundamentally driven by a 26.7% Reasoning Non-Convergence (cognitive collapse) rate. This collapse degraded trial-to-trial consensus stability to 72.6% and imposed a 17x latency overhead, introducing critical vulnerabilities to Governance Extractable Value (GEV) and hardware centralization. While rare (1.5% of adversarial trials), we empirically captured "Reasoning-Induced Sycophancy," where the model generated significantly longer internal monologues (averaging 25,750 characters) to rationalize failing the adversarial trap. We conclude that for edge-native SLMs operating under Byzantine Fault Tolerance (BFT) constraints, System 1 parameterized intuition is structurally and economically superior to System 2 iterative deliberation for decentralized consensus. Code and Dataset: https://github.com/smarizvi110/sentinel-bench
Smart contracts exhibit increasingly complex semantics and interactions, yet existing vulnerability detection methods rely on single-granularity representations, limiting their ability to capture semantic mechanisms across bytecode execution and cross-contract interactions. In addition, vulnerability data are scarce and imbalanced, and most deep learning-based approaches lack semantic interpretability. To address these limitations, a semantic-aware cross multi-granularity representation enhanced detection framework (CMR-ED) is proposed. CMR-ED models opcode execution semantics, function-level behaviors, and cross-contract interactions, aligning multi-level semantic information within a unified representation space. A structure-aware vulnerability pattern generator mitigates data scarcity through semantically consistent sample synthesis, while an explainable detection mechanism provides triggering paths and reasoning-chain explanations. Extensive experiments show that CMR-ED outperforms state-of-the-art methods while improving interpretability under semantically complex scenarios.
Rohith Singh, Mr. Charan Singh, Abdul Rashad, Md. Abdur Rasheed · 6 authors
Prompt injection is a foundational security vulnerability in large language models (LLMs) deployed as autonomous agents with tool access and multi-step reasoning capabilities. Existing defenses rely on heuristic filters that fail under obfuscation, indirect injection, and multi-agent propagation. We present a Unified Cryptographic-Control Architecture (UCCA), a principled framework that integrates five complementary guarantees: (1) information-theoretic leakage bounds derived via Fano's inequality, (2) certified robustness via randomized smoothing, (3) token-level rejection via erase-and-check, (4) runtime trajectory enforcement via control barrier functions (CBFs), and (5) verifiable inference via zero-knowledge proofs (ZK-SNARKs). We formally prove that any successful prompt injection attack must simultaneously bypass all five mechanisms, a condition we show has probability at most δ under stated assumptions. We evaluate UCCA on three real LLMs (GPT-4o, Claude 3.5 Sonnet, Mistral-7B) across four established attack benchmarks (INJECAGENT, TensorTrust, PromptBench, HarmBench), achieving attack success rates below 8% while maintaining median latency overhead under 340 ms. Our framework bridges formal security guarantees and deployable system architecture, establishing a foundation for provably secure autonomous AI. • Information-theoretic bounds on system prompt leakage using mutual information and Fano's inequality. • Certified robustness for safety-critical classification through randomized smoothing, where the robustness radius R is determined from output probability gaps. • Token-level rejection guarantees using an erase-and-check procedure capable of detecting adversarial subsets of size ≤ k. • Runtime safety enforcement through control barrier functions (CBFs), ensuring LLM outputs remain within a verified safe set. • Verifiable inference using ZK-SNARKs, allowing cryptographic attestation of model outputs without revealing model weights. • UCCA, a deployable system integrating all five mechanisms, evaluated on real LLMs and standard benchmarks.
Gregorio Dalia, Tat Luat Nguyen, Andrea Di Sorbo, Corrado Aaron Visaggio · 5 authors
Ethereum smart contracts manage billions in digital assets, and vulnerability detection is critical given the immutability of deployed code and the irreversible nature of transactions. However, existing tools such as Slither rely on rigid, rule-based analysis, and general-purpose language models like ChatGPT often miss rare or context-dependent bugs. To address these limitations, this paper presents BreachT5, an ensemble of two fine-tuned CodeT5+ models designed for multi-label vulnerability detection in Solidity contracts. We first fine-tune a 220M parameter model on over 67,000 real contracts labeled with the Smart Contract Weakness Classification (SWC), revealing intrinsic detection differences across vulnerability types. We then explore the performance of a 770M variant, which improves accuracy on frequent classes but underperforms on rare ones. To balance this trade-off, BreachT5 combines both models via soft voting with per-class thresholds. Our results on the BCCC-SCsVuls2024 dataset show that BreachT5 achieves 0.556 Macro-F1 and 0.612 Micro-F1, outperforming the two standalone models, Slither, and GPT-5 in multi-label vulnerability detection.
This paper proposes a conceptual methodological framework based on a Dual-Domain Architecture mediated by a Zero-Knowledge Audit Proxy (ZKAP) to reconcile AI Act accountability with GDPR data minimization. Legal norms are polynomialized into R1CS constraints, transforming compliance into a formally verifiable computational property. For cognitively opaque exascale models, these invariants may be hardware-anchored through a Provable Arithmetic Logic Unit (pALU), ensuring determinism and resistance to algorithmic drift. For lower-risk or on-premise systems, ZKAP operates in a software-only configuration, enabling periodic asymmetric regulatory proofs without silicon-level integration. A calibrated threshold distinguishes admissible technical variance from structural divergence, triggering mandatory safeguards. The framework provides a proportional, scalable, and cryptographically verifiable oversight model applicable both to future non-explainable AI systems and to lighter local infrastructures. This Zenodo deposit contains both the original Bulgarian peer-reviewed version (version of record) and an unofficial English translation. The Bulgarian version was published in Artificial Intelligence Proceedings (ISSN 3033-2923 / 3134-1667), pp. 75–78, as presented at the XI International Scientific Conference "High Technologies. Business. Society", Borovets, Bulgaria, 23–26 March 2026.
We establish an information-theoretic lower bound on the prover overhead of any zero-knowledge proof system that verifies arbitrary neural network inference. We prove a minimum multiplicative overhead of 2x for general circuits, rising to 4x for neural networks with ReLU activations due to activation encoding, weight commitment, and layer dependency costs. We further prove that composing ZK with fully homomorphic encryption produces multiplicative overhead blowup, making ZK+FHE verification impractical beyond approximately 10^4 gates. We survey six contemporary proof systems and show their observed overheads are consistent with our bounds. Our results formalize the intuition that free verification of AI computation is impossible and provide concrete bounds for system designers.
Open access
2 source records
Adversarial Robustness in Machine Learning
Cryptography and Data Security
Physical Unclonable Functions (PUFs) and Hardware Security
With the rapid iteration of blockchain technology, smart contracts, as core components of decentralized applications, directly impact the stability of on-chain assets and ecosystems through their security. Traditional vulnerability detection methods primarily rely on expert rules and static analysis, facing bottlenecks such as high false positive rates and poor adaptability to complex logical vulnerabilities. In recent years, Large Language Models (LLMs), with their exceptional code understanding and reasoning capabilities, have provided new technical pathways for smart contract security auditing. This paper focuses on LLM-driven smart contract vulnerability detection technologies, systematically reviewing mainstream application paradigms from prompt engineering to model fine-tuning. The paper first reviews the current state of smart contract security and the limitations of traditional methods; subsequently, it provides in-depth analysis of the architectural design and core mechanisms of representative frameworks such as GPTLens and SmartVD, evaluating their performance in detection accuracy and recall rate; finally, addressing current challenges including data scarcity, model hallucinations, and computational overhead, it proposes future evolution directions such as multimodal fusion and human-in-the-loop auditing, providing reference for research and practice in related fields.
Faithful, Stable, Complete: Pick Two The Problem in Plain Language When a machine learning model makes a prediction — approving a loan, diagnosing a disease, flagging a transaction — practitioners use a tool called SHAP to answer "which input features mattered most?" SHAP is the most widely used explanation method in machine learning. Here is the problem: retrain the same model on the same data with a different random seed, and the explanation changes. The model's predictions barely move, but the "most important feature" can flip entirely. In 68% of 77 public datasets, the top feature is not stable across retrains. This is not a software bug. This is not fixable by tuning hyperparameters. We prove it is a mathematical impossibility. What We Prove No feature ranking can simultaneously be: Faithful — it reflects what the model actually learned Stable — it doesn't change when you retrain Complete — it ranks every pair of features …when features are correlated with similar importance. You must give up one. The proof is four lines long. It requires no assumptions about the model, the data, or the explanation method — only that correlated features admit models ranking them in opposite orders (the Rashomon property), which is true for every standard ML algorithm. How Bad Is It? We trained 50 XGBoost models on Breast Cancer Wisconsin — the dataset used in every SHAP tutorial — and counted how many different "top 3 most important features" appeared. Twenty-four. At 100 models: thirty-five. The "most common" answer appeared in only 12% of runs. Two randomly chosen models agree on the top-3 only 4.2% of the time. Every tutorial, textbook, and blog post showing SHAP on this dataset is showing one of two dozen equally valid answers. Three other datasets (California Housing, Heart Disease, Wine Quality) produce exactly one ranking every time — because their top features have clearly different importance. The theory correctly predicts which datasets are affected and which are safe. Dataset Distinct top-3 rankings (50 models) Two models agree? Breast Cancer 24 4.2% Diabetes 2 88.5% Wine Quality 1 100% (stable) Heart Disease 1 100% (stable) California Housing 1 100% (stable) It Gets Worse for Yes/No Questions For ranking questions (which feature is MORE important?), there is a fix: average across multiple models. But for binary questions — "does this feature contribute positively or negatively?", "is this feature selected?" — no fix exists. Even averaging doesn't help, because there's no middle ground between "positive" and "negative." We call this the bilemma. Real-World Consequences For loan applicants. We trained 30 models on German Credit data. Under standard settings, 45% of applicants receive a different "most important reason" for their decision depending on which model happens to be deployed. One applicant received six different top reasons across 30 models. For biomarker discovery. On a dataset of 10,935 genes distinguishing colon from kidney tissue, the "#1 most important gene" alternates between TSPAN8 (involved in tumor invasion) and CEACAM5/CEA (involved in immune evasion) depending on the random seed. A drug discovery pipeline targeting one gene makes a different bet than one targeting the other — and which bet gets made depends on a random number. For fairness audits. A SHAP-based audit checking whether a model relies on a protected attribute (like race or gender) reaches its conclusion with the reliability of a coin flip when the protected attribute is correlated with other features. The Fix DASH (Diversified Aggregation for Stable Hypotheses): train 25 models with different seeds, average their SHAP values. This is provably the best possible approach — no method can do better. Features that genuinely differ in importance get stable rankings. Features that are interchangeable get reported as tied, which is the honest answer. We also provide a 7-line diagnostic that identifies which features are at risk, requiring no statistical expertise and no assumptions about the data distribution. It outperforms the standard formula by 2× on real data. The practical workflow: Screen your model (1 model, seconds) Run the minority fraction diagnostic (7 lines of code) For flagged features, train 5 models and run a Z-test If unstable, use DASH with 25+ models Machine Verification Every mathematical claim is checked by a computer. The proofs are written in Lean 4 (a programming language for mathematics) and verified by its type-checker: 357 theorems, all machine-verified 6 axioms (the minimal assumptions the theory needs) Zero unproved claims across 58 files During the formalization, the computer caught two logical errors and one type mismatch that human reviewers missed. To our knowledge, this is the first formally verified impossibility result in explainable AI. Technical Details Architecture-dependent bounds Gradient boosting (XGBoost, LightGBM): instability diverges as correlation increases. At ρ = 0.9, the dominant feature gets 5× its fair share. Lasso: the ratio is infinite — one correlated feature gets everything, the other gets zero. Neural networks: 87% of feature pairs are unstable. Model instability dominates SHAP estimation noise by 8:1. Random forests: instability converges with more trees — the contrast case showing that parallel (not sequential) training helps. Cross-implementation. XGBoost, LightGBM, and Random Forest all show the same instability pattern. It is not specific to any one software package. Subsample sensitivity. Even at subsample = 0.95 (minimal randomness), 17 distinct rankings remain. Only fully deterministic training (subsample = 1.0) produces one ranking — but this sacrifices the regularization that makes the model accurate. Mechanistic interpretability. Preliminary evidence suggests the impossibility extends beyond feature importance to neural network circuit analysis. 10 transformers trained on modular addition (all achieving 100% accuracy) agree on only 36% of the top-3 circuit components. Design Space The achievable set of explanation methods has exactly two families: Family A (single model): faithful and complete, but unstable. Rankings flip up to 50% of the time. This is what standard SHAP does. Family B (DASH ensemble): faithful and stable, but reports ties for indistinguishable features. This is what DASH does. No third option exists. DASH is provably the best method in Family B. Associated Papers Companion paper (TMLR, under review). First-Mover Bias in Gradient Boosting Explanations: Mechanism, Detection, and Resolution.arXiv: https://arxiv.org/abs/2603.22346DOI: https://doi.org/10.5281/zenodo.19446088 Companion implementation: https://github.com/DrakeCaraker/dash-shap
Smart contracts, essential to Blockchain functionality, can be compromised by vulnerabilities like reentrancy attacks, allowing unscrupulous entities to misappropriate funds. A universal and efficient multi-modal vulnerability detection framework is created to tackle detection issues that exceed the capability of standard methods such as fuzzy testing and symbolic execution. The methodology incorporates BiLSTM, EfficientNet, and Transformer architectures, augmented by CNN2D and BiGRU for better feature extraction and sequence modeling. The SMARTBUG dataset is employed in two formats: compiled OPCODES and features extracted via Word2Vec from smart contract source code. Preprocessing entails utilizing Word2Vec to produce N-gram numerical representations, succeeded by an 80-20 division for training and testing. The system analyzes multi-modal inputs, such as grayscale image attributes, opcode frequency statistics, and source code sequences, facilitating comprehensive vulnerability characterisation. The experimental assessment assesses the proposed model in comparison to existing algorithms, including MLP, GRU, and BiLSTM, utilizing criteria such as accuracy, precision, recall, and F-score. The CNN2D + BiGRU + EfficientNet + Transformer setup attains the greatest detection accuracy of 91.9%, surpassing all benchmarks. The system reduces dependence on domain knowledge by automating feature extraction, enabling adaptation across diverse smart contract forms and improving security in blockchain contexts
Fine-tuning is the dominant paradigm for adapting large machine learning models, yet current deployment pipelines provide no way to verify how a released model was updated. In particular, a model provider or auditor cannot check whether a fine-tuned model adheres to a claimed update procedure without access to its parameters. We introduce \emph{fine-tuning integrity} (FTI), a cryptographic objective for verifying that a deployed model differs from a trusted base model only within a declared class of admissible updates. We construct \emph{succinct model difference proofs} (SMDPs), zero-knowledge protocols that certify structured parameter drift without revealing model weights. Our framework supports three fundamental update classes: norm-bounded, low-rank, and sparse drift, covering common fine-tuning methods such as regularized training, LoRA, and prefix tuning. In all cases, proof size and verification cost depend on the structure of the update rather than the number of parameters. We prove soundness, zero-knowledge, and succinctness for each construction, and establish a matching $Ω(n)$ lower bound showing that structural assumptions are necessary for succinct verification. A prototype evaluation on synthetic benchmarks and GPT-2 fine-tuning demonstrates that proofs remain compact and verification is efficient at realistic scales.
A deployed model can appear unchanged while ceasing to be the model it claims to be. Publicly available weight-level mutation toolchains now automate safety-alignment removal from open-weight models on ordinary hardware, producing checkpoints intended to preserve operational familiarity while discarding refusal behavior. This paper argues that safety-alignment removal is a model-identity failure: in tested published checkpoints from multiple toolchains across two model families, the mutation leaves measurable structural scars ranging from 7.6 to over 2,300 times the instrument's acceptance threshold. Artifact identity, workload identity, and agent authorization can all remain valid while structural model identity fails — a finding that the program's formally verified admissibility doctrine predicted before this threat class existed. A sentinel validation panel across four model families confirms that the hardened instrument configuration preserves or improves all tested positives. In an agentic deployment context, model-identity failure propagates upward into agent-integrity failure: the agent is authenticated, but the model inside it is no longer the model the surrounding controls were designed to govern. The practical implication is that runtime evaluation frameworks — including those emerging under the EU AI Act — implicitly depend on a model continuity that weight-level mutation can break, and that structural identity verification offers a candidate evidentiary layer for closing that gap. The Neural Network Identity Series — Mathematical foundations, empirical validation, and governance frameworks for verifying which model is running Newest addition: Technical Note: The Disappearing Window — AI Logprob Access Withdrawal and the Structural Verifiability of Frontier Model Contracts (DOI: 10.5281/zenodo.20362098) Paper 1: The δ-Gene: Inference-Time Physical Unclonable Functions from Architecture-Invariant Output Geometry (DOI: 10.5281/zenodo.18704275) Paper 2: Template-Based Endpoint Verification via Logprob Order-Statistic Geometry (DOI: 10.5281/zenodo.18776711) Paper 3: The Geometry of Model Theft: Distillation Forensics, Adversarial Erasure, and the Illusion of Spoofing (DOI: 10.5281/zenodo.18818608) Paper 4: Provenance Generalization and Verification Scaling for Neural Network Forensics (DOI: 10.5281/zenodo.18872071) Paper 5: Beneath the Character: The Structural Identity of Neural Networks — Mathematical Evidence for a Non-Narrative Layer of AI Identity (DOI: 10.5281/zenodo.18907292) Paper 6: Which Model Is Running?: Structural Identity as a Prerequisite for Trustworthy Zero-Knowledge Machine Learning (DOI: 10.5281/zenodo.19008116) Paper 7: The Deformation Laws of Neural Identity (DOI: 10.5281/zenodo.19055966) Paper 8: What Counts as Proof? — Admissible Evidence for Neural Network Identity Claims (DOI: 10.5281/zenodo.19058540) Paper 9: Composable Model Identity — Formal Hardening of Structural Attestations in the Enterprise Identity Stack (DOI: 10.5281/zenodo.19099911) Paper 10:Where Identity Comes From: Path Sensitivity and Endpoint Underdetermination in Neural Network Training (DOI: 10.5281/zenodo.19118807) Paper 11: Post-Hoc Disclosure Is Not Runtime Proof: Model Identity at Frontier Scale (DOI: 10.5281/zenodo.19216634) Paper 12: Family-Dependent Response to Reasoning Distillation Across Structural and Functional Identity Layers (DOI: 10.5281/zenodo.19298857) Paper 13: Safety-Alignment Removal as a Model-Identity Failure — Structural Evidence from Published Weight-Level Mutation Checkpoints (DOI: 10.5281/zenodo.19383019) Technical Note: Agent Identity Is Not Model Identity (DOI: 10.5281/zenodo.19240883) Technical Note: Gap Invariance: Why PPP Measurements Are Domain-Independent by Construction (DOI: 10.5281/zenodo.19275524) Technical Note: Measured Model Substitution Under Valid Agent Credentials (DOI: 10.5281/zenodo.19342848) Technical Note: Artifact Identity Is Not Runtime Identity — Trustfall Lite and the Boundary of File-Level Model Verification (DOI: 10.5281/zenodo.20019127) Formal Verification Stack for Neural Network Structural Identity (IT-PUF Coq Proofs) (DOI: 10.5281/zenodo.18930621) Copyright (c) 2026 Anthony Ray Coslett / Fall Risk AI, LLC. All Rights Reserved. Confidential and Proprietary. Patent Pending (Applications 63/982,893, 63/990,487, 63/996,680, 64/003,244).
Open access
2 source records
Adversarial Robustness in Machine Learning
Physical Unclonable Functions (PUFs) and Hardware Security
Wang Yishun, Wenkai Li, Xiaoqi Li, Zongwei Li · 6 authors
Smart contracts are self-executing programs that manage financial transactions on blockchain networks. Developers commonly rely on third-party code libraries to improve both efficiency and security. However, improper use of these libraries can introduce hidden vulnerabilities that are difficult to detect, leading to significant financial losses. Existing automated tools struggle to identify such misuse because it often requires understanding the developer's intent rather than simply scanning for known code patterns. This paper presents LibScan, an automated detection framework that combines large language model (LLM)-based semantic reasoning with rule-based code analysis, identifying eight distinct categories of library misuse in smart contracts. To improve detection reliability, the framework incorporates an iterative self-correction mechanism that refines its analysis across multiple rounds, alongside a structured knowledge base derived from large-scale empirical studies of real-world misuse cases. Experiments conducted on 662 real-world smart contracts demonstrate that LibScan achieves an overall detection accuracy of 85.15\%, outperforming existing tools by a margin of over 16 percentage points. Ablation experiments further confirm that combining both analysis approaches yields substantially better results than either method used independently.