A Scalable Hardware Trojan Detection Framework Utilizing Self-Attention Enhanced FCNN and GAN-SMOTE-based Data Balancing

Abstract

With the globalization of integrated circuit (IC) supply chains, hardware Trojans (HTs) have posed severe security threats to ICs and the entire information infrastructure. Design-time HT detection is recognized as a critical proactive countermeasure; however, existing feature-based analysis methods suffer from inherent limitations: a large volume of candidate Trojan features, minor discrepancies in several key features, and extreme class imbalance between Trojan-infected and Trojan-free samples. These challenges collectively lead to suboptimal detection performance. In this paper, a scalable Hardware Trojan detection framework is proposed for gate-level netlists. Seven universal features encompassing structural and testability characteristics are extracted from benchmark circuits to represent most types of HTs. A Generative Adversarial Network (GAN) and Synthetic Minority Oversampling Technique (SMOTE) are integrated to generate synthetic Trojan feature samples and refine the classification boundaries of genuine samples. Furthermore, a self-attention mechanism is embedded into Fully Connected Neural Networks (FCNN), which highlights discriminative HT features and mitigates the obscuring effect of irrelevant features on critical Trojan-related characteristics. Experimental results demonstrate that the proposed approach outperforms state-of-the-art methods with respect to detection accuracy, true positive rate, true negative rate, and precision.