Generating Adversarial Examples for Hardware-Trojan Detection at Gate-Level Netlists

抄録

Recently, the great demand for integrated circuits (ICs) drives third parties to be involved in IC design and manufacturing steps. At the same time, the threat of injecting a malicious circuit, called a hardware Trojan, by third parties has been increasing. Machine learning is one of the powerful solutions for detecting hardware Trojans. However, a weakness of such a machine-learning-based classification method against adversarial examples (AEs) has been reported, which causes misclassification by adding perturbation in input samples. This paper firstly proposes a framework generating adversarial examples for hardware-Trojan detection at gate-level netlists utilizing neural networks. The proposed framework replaces hardware Trojan circuits with logically equivalent ones, and makes it difficult to detect them. Secondly, we propose a Trojan-net concealment degree (TCD) and a modification evaluating value (MEV) as measures of the amount of modifications. Finally, based on the MEV, we pick up adversarial modification patterns to apply to the circuits against hardware-Trojan detection. The experimental results using benchmarks demonstrate that the proposed framework successfully decreases the true positive rate (TPR) by a maximum of 30.15 points.