SVM算法在硬件木马旁路分析检测中的应用

Application of SVM Machine Learning to Hardware Trojan Detection Using Side-channel Analysis

集成电路(ICs)面临着硬件木马(HTs)造成的严峻威胁。传统的旁路检测手段中黄金模型不易获得,且隐秘的木马可以利用固硬件联合操作将恶意行为隐藏在常规的芯片运行中,更难以检测。针对这种情况,该文提出利用机器学习支持向量机(SVM)算法从系统操作层次对旁路分析检测方法进行改进。使用现场可编程门阵列(FPGA)验证的实验结果表明,存在黄金模型时,有监督SVM可得到86.8%的训练及测试综合的平均检测准确率,进一步采用分组和归一化去离群点方法可将检测率提升4%。若黄金模型无法获得,则可使用半监督SVM方法进行检测,平均检测率为52.9%~79.5%。与现有同类方法相比,验证了SVM算法在指令级木马检测中的有效性,明确了分类学习条件与检测性能的关系。

Integrated Circuits(ICs) are suffering severer threats caused by Hardware Trojans(HTs), some of which hide in routine operations by coercing firmware or hardware. Along with conventional side-channel detection not always getting golden-chip, HTs become more difficult to detect. An improved Support Vector Machine(SVM) machine learning frameworks for this is proposed using system-level side-channel analysis.Cross validation experimental results on Field Programmable Gate Array(FPGA) show that in the condition of golden-chip, supervised SVM achieves 85.8% test accuracy in average. After grouping, outlier-removing and normalization, it rises by 4%. Even if golden-chip is out of hand, semi-supervised SVM has accuracy to judge HTs existence, averaging in 52.9%-79.5% under different test modes. Comparing with existing researches, this work verifies the efficiency of SVM for HT detection in instruction level, and points out the relationship between diversified learning conditions with detection performance.