Pre-silicon hardware Trojan detection in mixed-signal circuits using heterogeneous graph attention networks

Abstract

Detecting hardware Trojans (HTs) in mixed-signal circuits is challenging due to structural complexity and cross-domain vulnerabilities between analog and digital components. Existing methods often rely on post-silicon analysis, circuit modifications, or focus solely on leakage, limiting practicality. We propose HGAT4TJ, a pre-silicon detection approach based on heterogeneous graph attention networks, which models gate- and transistor-level structures in a unified graph. This enables effective cross-domain HT detection directly from netlists without requiring golden models. Experimental results on benchmark circuits indicate that HGAT4TJ achieves 100% detection rate at the circuit level and over 97% accuracy at the node level, making it a non-invasive solution for HT detection in mixed-signal circuits.