Supervised and Unsupervised Machine Learning for Power Side-channel-based Trojan Detection

Abstract

The globalization of the semiconductor supply chain has created new challenges for security researchers. Hardware Trojans (HTs) are considered to be one of the most difficult challenges. This paper presents an effective HT detection method based on power side-channel features that can classify circuits under test (CUTs) into Trojan-inserted (TI) and Trojan-free (TF). It classifies the power traces based on the machine learning algorithms. The selected machine learning algorithms include supervised and unsupervised algorithms. The experimental results demonstrated on AES benchmarks show that the accuracy of TI power traces is 91.38% and 65.81% with supervised and unsupervised machine learning, respectively. Finally, it uses majority voting to perform the secondary classification on the CUTs based on the classification results of the power traces, which can mitigate the effects of process variations and noise. The experimental results show that the secondary classification can achieve 100% and 94.44% accuracy of TI circuits with supervised and unsupervised machine learning, respectively. The effect of dataset balance on machine learning performance was investigated, and a balanced dataset can improve accuracy by 13% to 30%. The experimental results on AES 8/128-bit HT demonstrate the effectiveness of the proposed method in detecting unknown Trojans.