Enhancing Multi-Label Vulnerability Detection of Smart Contract Using Language Model

Abstract

Smart contracts are decentralized applications that play an important role in blockchain-based applications. Smart contracts are written by programming languages (e.g., Solidity, Python, etc.), so it is error-prone and suffers from vulnerabilities, leading to a huge amount of economic loss for the blockchain ecosystem. In the past, there were many existing vulnerability detection tools such as MythX, Oyente, Slither, and so on. However, these tools contain several limitations related to low accuracy and high execution time. Therefore, many studies focus on vulnerability detection mechanisms using Deep Learning which takes into account the bytecode of smart contracts to detect its vulnerabilities. Despite achieving good accuracy, these studies make an assumption that there is only one vulnerability in a smart contract. When there is more than one vulnerability in a smart contract, these studies can not obtain good performance. Therefore, in this paper, we propose a multi-label vulnerability detection of smart contracts using a language model. Concretely, the proposal takes into account the bytecode by using the SecBERT pre-trained model to extract the implicit features and analyzes it using the Multi-Layer Perceptron algorithm to identify multiple vulnerabilities in a smart contract. The experimental results show that the proposal outperforms benchmarks and obtains 92.55 percent accuracy.