Smart Contract Vulnerability Detection Based on Multimodal Feature Fusion

The wide application of smart contracts advances the growth of blockchain technology. Still, the corresponding security issues also posed serious challenges to the stability of the blockchain contract layer. Many smart contract vulnerability detection methods are limited to a single modal representation, failing to capture the contracts’ semantic and structural information completely. To address this issue, this paper suggests a novel approach that combines multimodal feature fusion and deep learning to detect smart contract vulnerability, using Word2Vec and Gated Graph Neural Network (GGNN) to extract word vectors and semantic graph features, Multilayer Perceptron (MLP) to extract expert rule features from source code of smart contracts. A multiple-encoder network combining the self-attention mechanism and multi-channel convolution is used to fuse the extracted features, learn feature representation, and train the model for vulnerability detection. In the experiments, the proposed method is tested against a dataset of 40,932 smart contract codes. Results show that for reentrancy and timestamp dependency vulnerabilities, the accuracy is 92.01% and 93.31%, respectively, and the corresponding F1-scores are 88.66% and 92.28%.