UCGVulDetector: A Unified Contract Graph-Based Framework for Enhanced Smart Contract Vulnerability Detection

Abstract

In recent years, deep learning (DL) has shown remarkable performance in smart contract vulnerability detection, with graph neural networks (GNNs) serving as a key technique for learning structured code representations. However, existing graph-based approaches suffer from semantic fragmentation, noisy node interference, and weak semantic alignment, which limit detection robustness and deployment efficiency. To address these challenges, we propose UCGVulDetector, a unified and efficient framework designed for smart contract security in blockchain-based communication systems. It consists of three modules: (1) Structural Simplification (UDP): a hierarchical pruning strategy that refines abstract syntax trees by removing redundant nodes while preserving key semantics; (2) Graph Information Enhancement (UGSF): constructing heterogeneous graphs from Solidity ASTs and integrating control-flow, data-flow, and state-slot-chain (SSC) relations to capture multi-dimensional semantics; and (3) Graph Encoding and Alignment (PGE+TSCC): employing a pairwise graph encoder combined with temperature-scaled contrastive learning to align vulnerability semantics in a shared latent space. These modules collaboratively unify structural and semantic information to enhance feature representation. Experiments on the SolidiFI and MESSI datasets demonstrate that UCGVulDetector achieves F1-score improvements of 10.53%$\mathbf{1 1. 5 0 \%}$%ver state-of-the-art methods, delivering more accurate and robust vulnerability detection performance.