TD-Zero: Automatic Golden-Free Hardware Trojan Detection Using Zero-Shot Learning

Supply chain vulnerability provides the opportunity for the attackers to implant hardware Trojans in System-on-Chip (SoC) designs. While machine learning (ML) based Trojan detection is promising, it suffers from three practical limitations: 1) golden model may not be available; 2) lack of human expertise for Trojan feature selection; and 3) limited learning transferability can lead to unacceptable performance in new benchmarks with unseen Trojans. While recent approach based on transfer learning addresses some of these concerns, it still requires retraining for fine-tuning the model using domain-specific (e.g., hardware Trojan features) knowledge. In this article, we propose a Trojan detection framework utilizing zero-shot learning to address the above challenges. The proposed framework adopts the idea of self-supervised learning, where a pretrained graph convolutional network (GCN) is utilized to extract underlined common sense about hardware Trojans, and a metric learning task is used to measure the similarity between test inputs and malicious samples to make classification. Extensive experimental evaluation demonstrates that our approach has four major advantages compared to state-of-the-art techniques: 1) does not require any golden model during Trojan detection; 2) can handle both unknown Trojans and unseen benchmarks without any changes to the network; 3) drastic reduction in training time; and 4) significant improvement in detection efficiency (10.5% on average).