一种结合残差卷积注意力机制的二进制代码相似性分析方法  

Binary Code Similarity Analysis Method Combining Residual Convolutional Attention Mechanism

作　　者：周鑫 庞建民[1] 岳峰[1] 刘光明[1] ZHOU Xin;PANG Jianmin;YUE Feng;LIU Guangming(Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《信息工程大学学报》2024年第6期725-731,共7页Journal of Information Engineering University

基　　金：国家自然科学基金(61802433,61802435)。

摘　　要：针对跨架构的同源性漏洞检测问题,提出一种跨平台的二进制代码相似性分析方法 CodeMeld。该方法基于双向门控循环单元和自注意力机制实现指令序列信息提取,基于结合残差卷积注意力机制的卷积神经网络(CNN)模型实现邻接矩阵的图结构信息提取,通过对二者信息的融合实现函数的向量化表示。在不同架构下多种优化选项编译得到的数据集上进行评估实验。实验结果表明,CodeMeld可以有效捕获指令序列与控制流图结构的特征信息,准确实现跨平台的二进制代码相似性度量,准确率为94.26%,AUC高达0.980 6。To address the issue of cross-architecture same-origin vulnerability detection,a cross-plat-form binary code similarity analysis method called CodeMeld is proposed.Bidirectional gated recur-rent units(BGRU)and self-attention mechanism are relied upon by this method to extract instruction sequence information,and a convolutional neural network(CNN)model combining residual convolu-tion attention mechanism is used to extract graph structure information from the adjacency matrix.The two types of information are then fused to achieve function vectorization representation.Evaluation ex-periments are conducted on datasets compiled with various optimization options on different architec-tures.The experimental results demonstrate that the feature information of instruction sequence and control flow graph structure can be effectively captured by CodeMeld,which can accurately measure cross-platform binary code similarity with an accuracy of 94.26%and an AUC as high as 0.9806.

关 键 词：代码相似性分析 跨架构 神经网络 残差卷积 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]