ML-Based Hardware Trojan Detection in AI Accelerators via Power Side-Channel Analysis

Our work discusses the security risks associated with outsourcing AI accelerator design due to the threat of hardware Trojans (HTs), a problem traditional testing methods fail to address. We introduce a novel solution based on Power Side-Channel Analysis (PSCA), where we collect and preprocess power traces by segmenting them and extracting features from both time and frequency domains. This technique, coupled with an ML-based detection and identification method, was shown to effectively characterize various inserted HTs across several accelerators, achieving up to 99% accuracy.