HyWE: A Hybrid Word Embedding Method for Smart Contract Vulnerability Detection

Abstract

With the rise of blockchain technology, the use of smart contracts has increased, alongside challenges in detecting and addressing unknown vulnerabilities. Existing systems face limitations, including inadequate manual testing and flaws in automated tools. To improve smart contract security, this study introduces a Hybrid Word Embedding (HyWE) method that combines Word2Vec, FastText, and GloVe models, augmented by a channel self-attention mechanism for enhanced feature extraction in the field of deep learning smart contract vulnerability detection. HyWE captures semantic and contextual relationships in code more accurately, aiding in precise vulnerability detection. The method involves preprocessing smart contract data, extracting features with various embeddings, and applying self-attention to highlight critical features. Applied within the SCVD-SA framework, HyWE’s performance was evaluated experimentally, demonstrating superior accuracy and efficiency in vulnerability detection. This method is intuitive but effective, and can be easily adapted to other models, increasing performance.