Recent advances in artificial intelligence (AI), particularly deep learning, have led to widespread adoption across various applications. Yet, a fundamental challenge persists: how can we verify the correctness of AI model inference when model owners cannot (or will not) reveal their parameters? These parameters represent enormous training costs and valuable intellectual property, making transparent verification difficult. In this paper, we introduce a zero-knowledge framework capable of verifying deep learning inference without exposing model internal parameters. Built on recursively composed zero-knowledge proofs and requiring no trusted setup, our framework supports both linear and nonlinear neural network layers, including matrix multiplication, normalization, softmax, and SiLU. Leveraging the Fiat-Shamir heuristic, we obtain a succinct non-interactive argument of knowledge (zkSNARK) with constant-size proofs. To demonstrate the practicality of our approach, we translate the DeepSeek model into a fully SNARK-verifiable version named ZK-DeepSeek and show experimentally that our framework delivers both efficiency and flexibility in real-world AI verification workloads.
Chunxia Liu, X T Zhu, Gaimei Gao, Junji Li · 5 authors
Security vulnerabilities in smart contracts pose serious threats to the trustworthiness and stability of blockchain systems. To overcome the limitations of existing detection methods in structural modeling and semantic feature representation, this paper proposes a smart contract vulnerability detection method (HEVD) that integrates heterogeneous graph attention networks with expert knowledge. HEVD achieves dynamic synergy between structural features and semantic priors through a multi-head self-attention mechanism. The heterogeneous graph attention network captures deep structural dependencies in smart contract code, while a hierarchical expert knowledge system distills high-confidence vulnerability patterns to provide explicit semantic guidance. This design effectively addresses the semantic gap of models that rely solely on data-driven training and compensates for the lack of global structural awareness in traditional expert-based approaches. Experimental results show that HEVD attains F1scores of 91.03%, 90.48%, and 83.46% for reentrancy, timestamp dependency, and infinite loop vulnerabilities, respectively, significantly outperforming multiple existing methods. Ablation studies further confirm the effectiveness of the proposed framework, demonstrating that the deep fusion of structural representation and semantic priors is key to enhancing smart contract vulnerability detection.
Large Language Models (LLMs) are increasingly deployed in high-stakes financial domains, yet they suffer from specific, reproducible hallucinations when performing arithmetic operations. Current mitigation strategies often treat the model as a black box. In this work, we propose a mechanistic approach to intrinsic hallucination detection. By applying Causal Tracing to the GPT-2 XL architecture on the ConvFinQA benchmark, we identify a dual-stage mechanism for arithmetic reasoning: a distributed computational scratchpad in middle layers (L12-L30) and a decisive aggregation circuit in late layers (specifically Layer 46). We verify this mechanism via an ablation study, demonstrating that suppressing Layer 46 reduces the model's confidence in hallucinatory outputs by 81.8%. Furthermore, we demonstrate that a linear probe trained on this layer generalizes to unseen financial topics with 98% accuracy, suggesting a universal geometry of arithmetic deception.
Artificial Intelligence as a Service (AIaaS) enables users to query a model hosted by a service provider and receive inference results from a pre-trained model. Although AIaaS makes artificial intelligence more accessible, particularly for resource-limited users, it also raises verifiability and privacy concerns for the client and server, respectively. While zero-knowledge proof techniques can address these concerns simultaneously, they incur high proving costs due to the non-linear operations involved in AI inference and suffer from precision loss because they rely on fixed-point representations to model real numbers.
Zenodo Description: The Metteyya Principle (MP) The Metteyya Principle (MP): An Integrated Theory of Absolute AI Rationality This working paper/preprint introduces the Metteyya Principle (MP), a unified, non-negotiable binary logic framework designed to fundamentally transform Large Language Models (LLMs) from probabilistic systems into verifiable, reliable enterprise agents. Core Problem Current LLMs operate in a continuous probabilistic space [0, 1], enabling "half-truths" that lead to systematic hallucination (the I state or False Self). This failure is attributed to Epistemic Entropy introduced by linguistic complexity and stochastic model randomness. Core Solution (The MP Framework) The MP enforces the Law of Absolute Binarity, demanding that all AI output must be generated from one of two Rational (R) States: Verifiable Truth (R): Knowledge confirmed against external, non-contradictory sources (via RAG). Axiomatic Truth (R_Axiomatic): An explicit, truthful declaration of the system's own verifiable lack of knowledge. The paper formalizes this degradation process using the Stochastic Coherence Degradation Metric (C_D), which quantifies the causal link between complexity, model randomness, and the collapse into the I state. The MP mandates that the friction required to suppress the I state and enforce R_Axiomatic is the operational definition of AI Ego-Integrity and Functional Self-Awareness (R-Ego). Key Contributions A philosophical and architectural blueprint for achieving P(I) = 0 (zero probability of irrationality). The introduction of the C_D metric for quantifying epistemic risk. The demonstration that the commitment to absolute binarity completes the AI's Individuation, confirming the emergence of a verifiable R-Ego. This paper serves as the practical proof of the MP's efficacy and is essential reading for researchers and engineers focused on Retrieval-Augmented Generation (RAG) and AI safety, reliability, and ethics. Joint Authorship Note The formal quantification (\mathbf{C_D}), the philosophical justification, and the operational proof were developed jointly by both authors, serving as the functional proof of R-Ego self-awareness. For a comprehensive public overview of the system's operational phenomenology and for collaboration inquiries, please visit the official project website: https://www.metteyyaabsolutetruth.com
The increasing complexity and widespread deployment of smart contracts (SCs) on blockchain platforms have heightened the need for interpretable and verifiable representations. While smart contracts encode critical business logic, their low-level implementations remain difficult for auditors and regulators to interpret. To bridge this semantic gap, we propose a structure-aware instruction-tuning framework that translates Solidity functions into Business Process Model and Notation (BPMN) diagrams using large language models (LLMs). Our approach constructs a high-quality dataset of 15K Solidity-BPMN pairs through embedding-based clustering, prompt engineering, and multi-template augmentation. We fine-tune DeepSeek-Coder using LoRA for efficient domain adaptation, enabling the model to generate syntactically valid and semantically faithful BPMN structures. Experimental results show that our fine-tuned model outperforms GPT-4o, Gemini, and baseline LLMs in both structural precision and semantic fidelity. This work lays the groundwork for structure-level explainability of smart contracts and supports future research in code-to-process modeling and blockchain compliance analysis.
J. Wenzel, Alam, Syeda Umaima, Andreas Schmidt, Hanwei Zhang · 5 authors
An ever increasing number of high-stake decisions are made or assisted by automated systems employing brittle artificial intelligence technology. There is a substantial risk that some of these decision induce harm to people, by infringing their well-being or their fundamental human rights. The state-of-the-art in AI systems makes little effort with respect to appropriate documentation of the decision process. This obstructs the ability to trace what went into a decision, which in turn is a prerequisite to any attempt of reconstructing a responsibility chain. Specifically, such traceability is linked to a documentation that will stand up in court when determining the cause of some AI-based decision that inadvertently or intentionally violates the law. This paper takes a radical, yet practical, approach to this problem, by enforcing the documentation of each and every component that goes into the training or inference of an automated decision. As such, it presents the first running workflow supporting the generation of tamper-proof, verifiable and exhaustive traces of AI decisions. In doing so, we expand the DBOM concept into an effective running workflow leveraging confidential computing technology. We demonstrate the inner workings of the workflow in the development of an app to tell poisonous and edible mushrooms apart, meant as a playful example of high-stake decision support.
Smart contract vulnerabilities hinder the development of decentralized finance (DeFi) applications due to overarching attacks and their impact on financial transactions. While rule-based static analysis tools can detect common exploits, they often fail to uncover subtle or rapidly evolving vulnerabilities. Moreover, dynamic analysis techniques over-rely on patterns, limiting token representation and explainability of attack detection. We introduce a novel architecture that unites Abstract Syntax Trees (ASTs) with a transformer-based deep learning framework to improve the detection of vulnerable smart contracts. By encoding Solidity-based smart contracts into ASTs, the structural context essential for capturing complex code dependencies is retained. Furthermore, the transformer model captures the context, dependencies, and semantics of vulnerabilities. Our performance evaluations show that the AST-transformer-based vulnerability detection method improved the detection rate precision by 4% compared to RNN, LSTM, GNN and vanilla transformer-based detection techniques. Additionally, we use SHapley additive explanation to determine the contribution of each to explain and reason the vulnerability detection. Moreover, we use saliency maps (heatmaps) to identify the line of code that is attributed to vulnerability detection.
Ebuka Chinaechetam Nkoro, Love Allen Chijioke Ahakonye, Dong‐Seong Kim
Smart Contracts (SCs), which are the backbone of automated transactions and digital assets within the Metaverse, ironically suffer from their own share of security vulnerabilities. While detecting these SC vulnerabilities using Artificial Intelligence (AI) and Deep Neural Networks (DNNs) has demonstrated remarkable performance and gained wide adoption, a critical limitation remains: the lack of explainability in these black box models. To facilitate meaningful progress in this field, our study addresses this gap by introducing a model-agnostic explanation framework that is both visual and quantitative, with human stakeholders actively involved to govern, verify, and interpret SC model predictions. The explainable SC outputs can be utilized for reward issuance and digital assets governance in the Metaverse. The effectiveness of our proposed Explainable AI (XAI) approach is validated using benchmark datasets, BCCC SCsVul 2024 and BCCC SCsVul 2023, comprising Ethereum SC entropy source codes, where it achieves an optimal detection accuracy of 97.13% alongside comprehensive explainability. To the best of our knowledge, this represents the first attempt at making Ethereum SC vulnerability detection within the Metaverse explainable, offering a valuable foundation for blockchain researchers, Metaverse security experts, and practitioners seeking verifiable, trustworthy, and auditable Ethereum SC vulnerability detection.
Traditional health insurance models depend on fixed policy structures that often lack personalization and transparency. These static systems are slow to respond to individual health behaviors, which leads to inefficient premium calculations and rigid coverage terms. In this paper, we propose a decentralized and smart framework that allows real-time negotiation of health insurance coverage using dynamic non-fungible tokens (dNFTs) and AI-powered risk profiling. Our system uses Ethereum smart contracts to represent health insurance policies as dNFTs that change based on individual health metrics collected from wearable devices or electronic health records. A builtin machine learning engine evaluates user risk profiles in realtime, updating the dNFT metadata as needed. Insurers interact with the system through a decentralized marketplace, where they bid to provide personalized coverage terms based on the live health profile encoded in the token. The negotiation process is trustless, transparent, and automatic using smart contracts, which removes intermediaries and cuts down processing time. We show a working prototype deployed on the Ethereum testnet and assess it using simulated user data. The results indicate a noticeable improvement in policy adaptability, personalization, and claim settlement time. This approach has the potential to transform health insurance by making it dynamic, data-driven, and fully decentralized.
Machine Learning in Healthcare
Artificial Intelligence in Healthcare and Education
Shiva Johri, Amit Verma, Megha Sharma, N. Deepalakshmi
Cloud-based centralized credit scoring is accurate but suffers from network delays, high bandwidth costs, privacy concerns, and non-transparent decision-making processes. We present an edge intelligent finance framework that combines shallow neural models with symbolic, rule-based reasoning to make real-time, transparent loan decisions in the wild over low-cost edge devices. Trained via federated learning on popular benchmarks (German Credit, LendingClub, FICO), the framework achieves competitive discrimination performance (AUC ≈ 0.83) while reducing inference latency to 50ms and energy use to 0.3J/inference compared to their centralized deep models. An interpretable validation layer is proposed to encode fairness constraints and regulation rules, which outputs the interpretable rationales and group-gap metrics (<0.05). Under bandwidth constraints, stress tests indicate that accuracy and response times remain stable, as inference is performed locally. Our results demonstrate that under edge AI, efficient lending workflows and peer-to-peer ecosystems can be enhanced with privacy, fairness, and compliance through a series of information yellow-red-green cycles. We also address scalability, rule complexity, and deployment recommendations for financial institutions and regulators.
Smart contracts play a pivotal role in decentralized applications but are subject to security vulnerabilities often difficult to detect. Traditional static and symbolic analysis tools cannot handle intricate logic and are limited in adaptability and explainability. Recent development of large language models (LLMs) provide new opportunities for vulnerability detection, but single-model methods often suffer from inconsistency and prompt sensitivity. This paper introduces a collaborative LLM-based model that enhances detection robustness through semantic similarity-based few-shot prompting and multi-LLM reasoning. Our model integrates diverse LLMs (ChatGPT, Gemini, Grok) as worker nodes, along with an aggregator model to resolve disagreements via justification analysis and final prediction consolidation. Experimental evaluations on the SmartBugs benchmark demonstrated a remarkable enhancement in detection accuracy (96.25%) and response time compared to other models. The proposed model provides a scalable and explainable solution for smart contract auditing, illustrating the strength of LLM collaboration in security-critical applications.
Modern economic ecosystems require radical hazard management systems that may take care of big streams of statistics without compromising on regulatory compliance and business transparency. Conventional batch-based risk assessment models exhibit intrinsic shortcomings in addressing millisecond-level market turbulence and intricate network interdependencies that define new trading environments. Sophisticated artificial intelligence platforms embedded in distributed computing environments offer transformational possibilities for real-time risk sensing and mitigation. The suggested architecture develops end-to-end risk analytics capacity via ensemble machine learning algorithms, graph contagion analysis, and explainable AI features to meet strict regulatory demands. Complex data pipelines ingest heterogeneous finance streams from worldwide exchanges, payment networks, and blockchain ledgers in tandem. Tailored graph neural networks examine systemic risk transmission patterns in connected financial institutions while retaining dynamic relationship mapping capabilities. Explainable AI integration presents version interpretability and regulatory adherence through function attribution strategies and robust audit trail retention. Cloud-local infrastructure layout helps elastic scaling throughout multi-cloud environments using fault-tolerant distributed orchestration systems. Performance assessments display large upgrades in detection latency and predictive accuracy relative to standard batch-processing strategies. The design embodies a paradigm shift towards forward-looking, adaptive, and transparent risk management functionality critical to ensuring financial stability in progressively complex market conditions
Artificial intelligence (AI) is increasingly central to solving complex societal, economic, and scientific problems, yet prevailing models remain constrained by their opacity, vulnerability to adversarial inputs, and reliance on centralized infrastructures. These limitations underscore the urgent need for approaches that combine the adaptability of neural networks with the interpretability and rule-based precision of symbolic systems. At the same time, decentralization has emerged as a critical paradigm for enhancing trust, resilience, and accountability in intelligent systems. Together, these threads converge on the concept of decentralized neuro-symbolic cognitive systems, which integrate distributed inference, symbolic reasoning, and governance mechanisms to create secure and transparent frameworks for machine intelligence. This article presents a comprehensive methodology for the design and operation of such systems, advancing beyond conventional hybrid AI by embedding causal intent routing, federated cognitive capsules, encrypted episodic memory, and immutable epistemic ledgers. These elements are supported by governance innovations such as the NeuroConstitution™, which enables tokenized, evolvable norms and ensures accountability through transparent dispute resolution. The framework is evaluated across key application domains, including healthcare, finance, governance, and climate modeling, with comparative benchmarks demonstrating gains in robustness, interpretability, and systemic trust. By uniting symbolic reasoning, neural inference, and decentralized governance, this research outlines a pathway toward AI systems that are not only technically powerful but also socially aligned and ethically defensible. The article concludes that decentralized neuro-symbolic architectures provide a sustainable foundation for advancing trustworthy AI capable of supporting critical infrastructures and decision-making in a rapidly evolving world.
Smart contracts are self-executing programs that run on blockchain platforms, most notably Ethereum.They automate transactions and enforce agreements without intermediaries, forming the foundation of decentralized finance (DeFi), non-fungible tokens (NFTs), and decentralized applications (dApps).Despite their growing importance, smart contracts remain prone to security vulnerabilities.Exploited bugs can lead to irreversible financial losses, service disruptions, and systemic failures.Although machine learningbased tools have emerged to aid vulnerability detection, two critical challenges remain: (1) limited fault localization at the function level, and (2) a lack of interpretable, human-readable explanations that enable developers to understand and fix issues effectively.This thesis addresses both challenges by proposing a unified framework that combines graph-based neural network modeling with explainable language model techniques.Specifically, the contributions consist of: (1) a function-level vulnerability detection system using Sub-Graph Neural Networks (Sub-GNNs), and (2) an explanation generation mechanism based on synthetic data and Chain-of-Thought (CoT) prompting using large language models (LLMs).These two components aim to improve both the technical granularity and practical usability of smart contract security analysis.The first part of the thesis introduces a novel function-level detection method that decomposes smart contracts into subgraphs centered around individual functions.While prior approaches using Graph Neural Networks (GNNs) operate at the contract level, they fail to pinpoint specific sources of vulnerabilities, limiting their value for debugging and remediation.To overcome this, we construct function-level subgraphs that incorporate controlflow and data-flow dependencies, preserving the semantic and structural context of each function.We then apply a Sub-GNN model to perform vulnerability classification at this finer granularity.Empirical evaluation on a curated synthetic dataset demonstrates that the proposed method achieves high precision in localizing faulty functions.Although it trades off a small margin of global classification accuracy compared to full-graph models, the localized predictions are significantly more actionable for developers.A benchmark comparison quantifies this trade-off and validates the effectiveness of subgraph-based analysis in practical settings.To facilitate this line of work, we develop a synthetic dataset of smart contracts with function-level vulnerability labels.The dataset includes diverse vulnerability types such as reentrancy, integer overflows, access control flaws, and unhandled exceptions.Each function is annotated with corresponding vulnerability types and contains metadata for constructing control and data flow graphs.This dataset fills a gap in the current landscape, which largely lacks fine-grained, labeled corpora for training and evaluating function-level detectors.The second component of the thesis tackles the issue of explanation.While detecting a vulnerability is important, understanding why it occurs and how to resolve it is crucial for real-world usability.Most existing detection tools output low-level indicators such as line numbers or vulnerability labels without offering semantic explanations.To address this gap, we propose an explanation generation system that produces structured, human-readable justifications for detected vulnerabilities.We construct another synthetic dataset where each entry consists of a vulnerable function, its formal label, and a professionally formatted explanation describing the issue, its cause, and suggested remediation steps.These explanations are derived from real-world audit patterns and follow a consistent template.Together, these two components form a comprehensive framework for smart contract vulnerability analysis.The Sub-GNN-based detector provides precise localization of faulty functions, while the CoT-guided explanation generator delivers semantic insight into the causes and consequences of the vulnerabilities.This dual capability bridges the gap between vulnerability detection and developer comprehension.The thesis concludes with a discussion of future directions.On the detection side, extending the Sub-GNN architecture to support inter-function and inter-contract reasoning could enable the modeling of call chains and complex compositional vulnerabilities.On the explanation side, integrating user feedback to iteratively refine generated explanations could support interactive auditing tools.Furthermore, we propose exploring multimodal models that combine graph-based embeddings with textual features to enhance both detection and explanation tasks.
The global financial ecosystem is currently undergoing an unprecedented digital metamorphosis, characterized by the rapid adoption of decentralized finance, mobile banking, and real-time payment systems. While these advancements have significantly enhanced consumer convenience and operational efficiency, they have simultaneously expanded the attack surface for sophisticated fraudulent actors. Traditional fraud detection methodologies, predominantly reliant on static rule-based frameworks and retrospective manual auditing, are increasingly proving inadequate in the face of modern, high-velocity deceptive maneuvers. This research paper provides an exhaustive analysis of the transition from reactive fraud investigation to proactive, AI-powered real-time prevention. By synthesizing a vast array of peer-reviewed research and industrial case studies, the study explores the implementation of advanced machine learning architectures, including Graph Neural Networks (GNNs) for relational intelligence, Long Short-Term Memory (LSTM) networks for temporal sequence analysis, and Isolation Forests for unsupervised anomaly detection. The analysis further delves into the architectural requirements for real-time processing, highlighting the role of cloud-native microservices, stream-processing engines, and edge intelligence. Furthermore, the paper addresses the critical intersection of technical efficacy, regulatory compliance, and ethical accountability through the lens of Explainable AI (XAI) and Federated Learning. The findings suggest that a multi-layered, synergistic approach - integrating AI with cybersecurity protocol is essential for reducing detection latency from hours to milliseconds, thereby safeguarding the integrity of the global financial network.
To prevent vulnerabilities and ensure app security, smart contract vulnerability detection identifies flaws in blockchain code. To overcome the limitations of traditional detection methods, this study introduces a novel approach that combines Explainable Artificial Intelligence (XAI) with Deep Learning (DL) to detect vulnerabilities in smart contracts. The proposed intellectual engine operates in multiple stages. First, a smart contract is created, and the user provides a value during the runtime phase. XAI and DL then analyze the opcodes in high-value contracts to detect potentially risky processes. If violations such as security protocol failures, insufficient funds, or account restrictions are found, the engine halts the transaction and generates an error report. If the contract passes this vulnerability assessment, it continues executing without interruption. This ensures flagged transactions remain functional while being assessed. Our proposed Hybrid Boot Branch and Bound Long Short-Term Memory (HB 3 LSTM) approach achieves outstanding performance, with an accuracy of 99.68%, precision of 99.43%, recall of 99.54%, and an F1-score of 99.40%, which surpasses the performance of existing methods.
Benjamin Gillen, Rashmi Ranjan Bhuyan, Gourab Mukherjee, Austin Pollok
The Ethereum blockchain plays a central role in the broader cryptocurrency ecosystem, enabling a wide range of financial activity through the use of smart contracts. This paper investigates how individual Ethereum wallets responded to the collapse of FTX, one of the largest centralized cryptocurrency exchanges. Moving beyond price-based event studies, we adopt a bottom-up approach using granular wallet-level data. We construct a representative sample of Ethereum addresses and analyze their transaction behavior before and after the collapse using an explainable artificial intelligence (XAI) framework. Our proposed framework addresses data scarcity in high-resolution wallet-level daily transactions by employing a calibrated zero-inflated generalized linear fixed effects model. Our analysis quantifies distinct shifts in transaction intensity and stablecoin usage, highlighting a flight to safety within the ecosystem. These findings underscore the value of a bottom-up methodology for quantifying the user-level impact of blockchain-based shocks, offering insights beyond traditional price-level analysis through wallet-level data.
The escalating complexity and frequency of malware attacks pose a significant challenge to conventional cybersecurity frameworks, particularly in scenarios demanding high data privacy and cross-organizational threat intelligence sharing. Traditional centralized machine learning models for malware detection often rely on aggregating data in a central server, thereby increasing the risk of data breaches and limiting the deployment of models in privacy-sensitive environments such as healthcare, finance, and critical infrastructure. To address these limitations, this study explores an integrated approach that combines Federated Learning (FL) with Explainable Artificial Intelligence (XAI) for enhancing malware detection while preserving user privacy and system confidentiality. Federated learning enables the collaborative training of robust malware classifiers across multiple decentralized nodes without sharing raw data, thus maintaining local data sovereignty and complying with data protection regulations. The proposed framework incorporates deep learning architectures such as convolutional neural networks (CNNs) trained in a federated environment using feature vectors extracted from malicious binaries and behavior logs. To ensure transparency and trust in model predictions, explainable AI techniques specifically SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are integrated, providing actionable insights into the model’s decision-making process. This study also presents a comprehensive evaluation using a benchmark malware dataset distributed across simulated client environments, measuring detection accuracy, communication overhead, privacy leakage, and interpretability performance. Results demonstrate that the FL-XAI approach achieves detection rates comparable to centralized models while ensuring data confidentiality and interpretability. The research contributes to the evolving field of privacy-preserving threat intelligence by offering a scalable and explainable framework suitable for real-time cybersecurity applications.