Hardware Trojan Detection with Machine Learning and Power Side-Channels: A Post-Deployment Analysis

The detection of hardware Trojans has become increasingly complex due to evolving attack vectors and the growing separation between IP owners and foundries, a trend driven by the industry's shift toward fabless manufacturing. Hardware Trojans involve malicious modifications to hardware designs that create exploitable vulnerabilities. Traditional detection techniques, such as reverse engineering and side-channel analysis, have proven effective but often come with high costs and substantial preprocessing overhead. This paper explores machine learning algorithms combined with side-channel analysis to detect dormant hardware Trojans. We collected power consumption data from an Artix FPGA running an AES-128 algorithm that included various Trojan insertions. Experimental results show that applying machine learning directly to power consumption traces allows for effective Trojan detection without extensive feature extraction. Among the classifiers evaluated, the Random Forest model achieved a detection accuracy of 98%, outperforming existing methods. The K-Nearest Neighbors (KNN) classifier also showed strong performance, achieving an accuracy of 97%. Our proposed method reduces preprocessing requirements, lowers computational costs, and improves noise robustness, thereby enabling the post-deployment detection of Trojans in integrated circuits.