2024 | Yizhou Chen, Zeyu Sun, Zhihao Gong, Dan Hao
The paper "Improving Smart Contract Security with Contrastive Learning-based Vulnerability Detection" addresses the critical issue of smart contract vulnerabilities (SCVs) in blockchain systems. Existing methods often treat each contract as an independent entity, neglecting the correlation between contracts, which limits their effectiveness. To tackle this, the authors propose Clear, a novel approach that integrates contrastive learning (CL) to capture fine-grained correlation information among smart contracts. Clear generates correlation labels based on the relationships between contracts and combines these labels with semantic information to enhance vulnerability detection.
The method is evaluated on a large dataset of over 40,000 real-world smart contracts, comparing it against 13 state-of-the-art baseline methods. The results show that Clear outperforms all baseline methods, achieving a significant improvement in precision, recall, and F1-score. Specifically, Clear achieves an average F1-score of 94.52%, a 9.73% increase over the best-performing method. Ablation studies further demonstrate the effectiveness of Clear's components, including the importance of both types of contract relationships and the role of the masked language model (MLM) module.
The paper also explores the impact of Clear's CL module on other deep learning models, showing that it can enhance the performance of traditional models like RNN, LSTM, and GRU. Overall, Clear provides a robust solution for improving the security of smart contracts by leveraging the correlation information among contracts.The paper "Improving Smart Contract Security with Contrastive Learning-based Vulnerability Detection" addresses the critical issue of smart contract vulnerabilities (SCVs) in blockchain systems. Existing methods often treat each contract as an independent entity, neglecting the correlation between contracts, which limits their effectiveness. To tackle this, the authors propose Clear, a novel approach that integrates contrastive learning (CL) to capture fine-grained correlation information among smart contracts. Clear generates correlation labels based on the relationships between contracts and combines these labels with semantic information to enhance vulnerability detection.
The method is evaluated on a large dataset of over 40,000 real-world smart contracts, comparing it against 13 state-of-the-art baseline methods. The results show that Clear outperforms all baseline methods, achieving a significant improvement in precision, recall, and F1-score. Specifically, Clear achieves an average F1-score of 94.52%, a 9.73% increase over the best-performing method. Ablation studies further demonstrate the effectiveness of Clear's components, including the importance of both types of contract relationships and the role of the masked language model (MLM) module.
The paper also explores the impact of Clear's CL module on other deep learning models, showing that it can enhance the performance of traditional models like RNN, LSTM, and GRU. Overall, Clear provides a robust solution for improving the security of smart contracts by leveraging the correlation information among contracts.