Journal of Cryptographic Engineering Computer Networks and Communications
An explainable semi-supervised self-training approach for hardware Trojan detection using LIME
Abstract
The rapid proliferation of Internet of Things (IoT) and Artificial Intelligence (AI) applications has escalated the prominence of Hardware Trojans (HTs) as a profoundly menacing threat to Integrated Circuits (ICs). Recent research has harnessed the power of Machine Learning (ML) and Deep Learning (DL) models for HT detection across diverse IC phases. However, the efficacy of these models in HT detection is hampered by three critical challenges: the absence of standardized large datasets prevalent in other domains, the pervasive class imbalance issue, and non-interpretable predictions. These challenges diminished the ability of ML/DL models to achieve the desired accuracy in HT detection. Further, supervised learning entails laborious and error-prone manual labeling of training samples, while unsupervised learning underperforms due to the unavailability of labeled samples. Moreover, the hidden nature of ML models also poses a limitation to understanding the predictions made by them. Addressing these issues, this paper proposes an Explainable Semi-Supervised Self-Training approach for HT detection in the pre-silicon IC phase. The proposed approach addresses all three challenges present in both supervised and unsupervised approaches. The proposed self-training SSL-based approach effectively handles the scarcity of labeled training data by leveraging both labeled and unlabeled samples. Further, a hybrid ADASYN-ENN sampling strategy is proposed, which integrates oversampling and undersampling to tackle severe class imbalance arising due to a small number of Trojan-inserted samples. Finally, Local Interpretable Model-agnostic Explanations (LIME) is utilized to interpret model predictions, i.e., which influential features are driving the ML model decisions. Experimental results show that the proposed technique accurately detects HT from TrustHub benchmarks of small and large size with 98% accuracy & 95% recall. Further, it also accurately detects the HTs from DeTrust & Detest benchmarks respectively.Furthermore, global SHAP analysis, component-wise ablation study, and adversarial robustness analysis further validate the proposed approach.