Abstract: Security vulnerabilities in smart contracts on blockchain platforms such as Ethereum have long been a focus of industry attention.Bytecode analysis and vulnerability detection have become one of the mainstream approaches for identifying smart contract vulnerabilities.However,traditional methods,such as symbolic execution,rely on predefined vulnerability rules,leading to inefficiencies and low precision.Deep learning-based methods,on the other hand,lack a comprehensive understanding of bytecode semantics and struggle to simultaneously filter noise generated during the compilation process while capturing complete control flow and data flow information.To address these challenges,this paper proposes a novel method for constructing critical semantic graphs to detect smart contract vulnerabilities.Firstly,a set of specific denoising preprocessing rules is defined to remove irrelevant data while preserving key semantic information related to vulnerabilities.Next,a heterogeneous graph representation method is introduced to capture rich program semantics.Finally,a vulnerability detection model based on the HAN is designed.Experimental results demonstrate that the proposed method outperforms existing approaches for smart contract vulnerability detection.For denial of service,integer overflow,timestamp dependency,and unchecked function return value vulnerabilities,the F1 scores of the proposed method are improved by 17.75,5.94,28.94,and 27.85 percentage points,respectively.

Key words: Smart contract, Smart contract security, Graph neural network, Smart contract bytecode