Improving Hardware Trojan Detection with Transformer-Based Power Analysis

Abstract The prevalence of hardware trojans (HTs) poses a significant threat to the integrity and security of Integrated Circuits (ICs). Rule-based Hardware Trojan Detection (HTD) techniques are limited in their effectiveness and scalability. Many domains such as Natural Language Processing (NLP), and cybersecurity are experiencing a surge in automation with the aid of modern Generative Artificial Intelligence (GenAI) techniques such as Generative pre-trained transformer (GPT), Bidirectional Encoder Representations from Transformers (BERT), which are applied in Large Language Models (LLMs). In HTD, transformers have only recently started to receive traction. This paper proposes a novel non-destructive golden-chip free transformer-based HTD framework. The proposed framework is applied to Power Side-Channel (PSC) data. Modern generative AI techniques such as GPT, BERT, and transformers are exploited to solve the HTD problem. The proposed framework combines the power of transformer-based networks with time-series side-channel analysis to achieve efficient and accurate HTD. The side-channel data are processed by different transformer networks, including GPT, BERT, and full transformer models to classify the trojan into three main categories: Enabled, Disabled, and Triggered HTs. The proposed framework effectively analyses side-channel measurements, accurately detecting abnormal IC behaviours. The experimental results demonstrated promising and superior performance, achieving an accuracy of 87.74% in HT detection compared with existing frameworks.