DeFi Smart Contract Vulnerability Detection Based On Teacher-student Network

Decentralized finance(DeFi)is a peer-to-peer financial ecosystem based on blockchain technology,which allows users to trade financial products on a distributed blockchain system.Decentralized finance has made significant progress in recent years and has become a key core technology for independent innovation of the country’s new infrastructure.So far,decentralized finance applications have held more than $1 trillion worth of digital assets,and various attacks against the decentralized finance ecosystem have caused more than $20 billion in economic losses,it seriously threatening the security of the entire decentralized finance ecosystem.Although researchers and security engineers have designed a large number of smart contract vulnerability detection tools to check,most of these methods are tailored for traditional smart contracts.Based on these,this paper takes decentralized finance smart contract vulnerability detection as the research theme,investigates decentralized finance attack events and the detection effect of traditional detection tools,and proposes a new detection technology based on the Teacher-Student network model:(1)Decentralized finance attack classification and experiments.We comprehensively studied 51 decentralized finance attack incidents and summarized four common types of decentralized finance attacks.In order to describe the current decentralized finance security issues in more detail,we constructed a decentralized finance annotation dataset containing6347 decentralized finance smart contracts and involving four types of attacks,and finally we conducted effectiveness analysis and compatibility evaluation of 35 state-of-the-art smart contract vulnerability detection tools for decentralized finance protocols.(2)Vulnerability detection of DeFi smart contract based on Teacher-Student network.This paper designs a new source code assisted Teacher-Student network mutual learning framework,which utilizes a code semantics aware module to encode source code and bytecode into corresponding graph-structured data,and uses a network structure aware module to construct single-modal student network and a dual-modal teacher network.To facilitate efficient knowledge transfer between two networks,we propose a cross-modal mutual learning framework to assist in learning source-bytecode correlations by introducing bytecode-bytecode and source-bytecode transfer losses.The empirical results show that our method produces the best detection results for the detection of decentralized finance smart contract vulnerabilities,and the accuracy rates of four attack types are 84.35%,80.72%,88.33% and 77.53%,respectively.Compared with the state-of-the-art detection tools,the relative accuracy rates are increased by 8.86%,6.78%,7.91% and 7.75%,respectively.Finally,we explore the pressing issues and challenges of decentralized finance,which we hope can help future work.