Robust Hardware Trojan Detection: Conventional Machine Learning vs. Graph Learning Approaches

Hardware Trojans (HTs) have emerged as a security threat to the integrated circuits (ICs) industry. To counteract this threat, various detection methods have been proposed, among which conventional machine learning (ML)-based techniques using heuristic features from gate-level netlists have gained wide acceptance. However, these methods are notably sensitive to minor perturbations in modifications of the test circuits, often resulting in decreased detection capabilities. Furthermore, the black-box nature of the ML models obscures the basis for their decisions. This lack of transparency makes it difficult to scrutinize and address potential flaws in the models, thereby further reducing the credibility of Trojan detection results. In response to these challenges, we propose a targeted solution that leverages the SHapley Additive exPlanations method (SHAP), which dismantles the black-box paradigm and clarifies the fundamental reasons behind the failure of existing detection methods under circuit sample perturbations. Building on these insights, we abandon classical ML-based detection in favor of a scheme based on graph learning (GL), which significantly reduces the average drop in Recall from 52.35% to 7.29% compared with the traditional method. Comparative experiments demonstrate that our proposed GL-based method effectively resolves the sensitivity issue related to the sample perturbations in existing HT detection approaches.