Intelligent Smart Contract Vulnerability Detection Using Graph Neural Networks and Explainable Artificial Intelligence

Abstract

Smart contracts automate asset transfers and decentralized application logic, yet their immutability and compositional complexity amplify the consequences of implementation defects. This paper proposes XG-SCV, an explainable heterogeneous graph neural network framework for vulnerability detection and source-level localization. Solidity source code is transformed into a typed graph combining abstract syntax, control flow, data dependence, external calls, and cross-function storage interactions. Relation-aware message passing and attention-based pooling produce multi-label vulnerability probabilities, while integrated gradients and graph-substructure attribution provide auditable explanations. The evaluation protocol emphasizes project-disjoint splits, duplicate removal, calibrated thresholds, class imbalance, and explanation fidelity. An illustrative numerical study demonstrates how the proposed model could be compared with static analysis and neural baselines; these illustrative numbers are not empirical findings. The framework offers a reproducible research design for detecting reentrancy, access-control weaknesses, unchecked external calls, arithmetic risks, and timestamp dependence while retaining interpretable evidence for developer review.