GREAT: Global Representation and Edge-Attention for Hardware Trojan Detection

Zhengyi Li, Wei Cheng, Huanrong Tang, Yaohua Wang

Abstract: With the increasing complexity of Integrated Circuit design and the globalization of the supply chain, the threat posed by Hardware Trojans is becoming increasingly significant. Currently, HT detection methods based on graph neural networks demonstrate the most promising performance. However, existing GNN-based methods fail to capture global representations and edge distinctiveness, and their corresponding sampling methods are not sufficiently efficient, resulting in a low F1 score. We propose GREAT, a detection framework that accurately locates Trojans in gate-level netlists. GREAT leverages feature fusion technology to mitigate the deficiency in global representation inherent in GNN. We introduce Wire Encoding, which assigns a unique identifier to each edge, enhancing sensitivity to shared wires. Additionally, GREAT incorporates a novel graph sampling technique to reduce resource consumption and improve training efficiency. The experimental results demonstrate that the GREAT has achieved a recall of 94.58% and an F1 score of 96.18%, surpassing all existing HT detection methods on gate-level netlists.

External IDs:dblp:conf/dsn/LiCTW25