作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于可控制性度量的图神经网络门级木马检测方法

  • 发布日期:2023-11-14

Hardware Trojan Detection with GCN Based on controllability in Gate-Level Netlist

  • Published:2023-11-14

摘要: 随着全球化的不断深入,第三方知识产权(IP)核应用越来越广泛;而随着硬件木马攻击技术的逐步成熟,使得在芯片设计阶段植入硬件木马成为可能,因此,芯片设计过程中面临IP核被植入木马的严重威胁。现有研究中所提出的硬件木马检测方法具有以下缺陷:依赖黄金参考电路,需要完备的测试向量,需要大量的样本进行学习等。本文面向IP核的硬件木马检测需求,提出了一种基于可控制性度量的图神经网络检测方法。该方法以门级网表作为输入,首先以可控制性值为指导,得到可疑的门节点,用于缩小搜索范围;然后利用可疑门节点生成对应的子图,利用图卷积神经网络从子图中提取特征,实现对子图的分类和检测,最终识别硬件木马的存在。本文的方法无需测试激励和黄金模型,通过利用硬件木马的隐蔽特性与结构特征相结合的方法提升了硬件木马的检测准确率。该方法平均真阳率为100%,假阳率为0.75%,结果表明本文提出的方法在保证较高真阳率的同时,可有效降低假阳率,达到较好的检测效果。

Abstract: With the continuous deepening of globalization, third-party intellectual property (IP) core applications are becoming increasingly widespread. The gradual maturity of hardware Trojan attack technology makes it possible to implant hardware Trojan in the design process, posing a serious threat to the security of chip design. The hardware Trojan detection methods proposed in current work have the following drawbacks: relying on golden reference circuits, requiring complete test patterns, requiring a large number of samples for learning and so on. This paper proposes a graph neural network detection method based on controllability metrics for hardware Trojan detection requirements of IP cores. This method takes a gate level netlist as input, and first uses controllability values as guidance to obtain suspicious gate nodes for narrowing the search range; Then, the suspicious gate nodes are generated into corresponding subgraphs, and the graph convolutional neural network is used to extract features from the subgraphs, achieving detection of the subgraphs, and ultimately identifying the existence of hardware Trojans. This method eliminates the need for testing patterns and golden models, and improves the detection accuracy of hardware Trojans by combining their rare triggered features and structural characteristics. The experimental results show that the average True Positive Rate of this method is 100%, and the False Positive Rate is 0.75%. The results show that the proposed method can effectively reduce the false positive rate while ensuring a high true positive rate, achieving good detection results.