Multimodal Feature Fusion Framework for Smart Contract Vulnerability Detection

Smart contracts play a central role in automating processes on blockchain platforms, enabling operations such as identity management, access control, and logic execution. However, due to their inherent complexity and extensive interactions, they are often vulnerable to a range of security issues, such as reentrancy and timestamp dependency. In this work, we propose a novel vulnerability detection framework that integrates expert knowledge and multimodal representations to improve both accuracy and interpretability in smart contract analysis. Specifically, we leverage the Code Llama large language model to extract expert pattern features and construct three semantic graphs — textual, opcode-level, and transaction-based — to capture heterogeneous behavioral patterns of contracts. These multimodal features are aligned and fused through graph neural networks and attention mechanisms to form a unified representation, which is then used for binary classification of vulnerability presence. Experimental results on both public and custom datasets demonstrate that our approach achieves strong detection performance;for instance, the accuracy rate for detecting timestamp dependency vulnerabilities reaches 89 % on private datasets. These findings validate the effectiveness of combining multimodal fusion with large language model-guided knowledge extraction for enhanced smart contract security analysis.