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

Recently, due to the increase of outsourcing in integrated circuit (IC) design and manufacturing, the threat of injecting a malicious circuit, called a hardware Trojan, by third party has been increasing. Machine learning has been known to produce a powerful model to detect hardware Trojans. But it is recently reported that such a machine learning based detection is weak against adversarial examples (AEs), which cause misclassification by adding perturbation in input data. Referring to the existing studies on adversarial examples, most of which are discussed in the field of image processing, this paper first proposes a framework generating adversarial examples for hardware-Trojan detection for gate-level netlists utilizing neural networks. The proposed framework replaces hardware Trojan circuits with logically equivalent circuits, and makes it difficult to detect them. Second, we define Trojan-net concealment degree (TCD) as a possibility of misclassification, and modification evaluating value (MEV) as a measure of the amount of modifications. Third, judging from 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 true positive rate (TPR) by at most 30.15 points.