Hardware Trojan Detection with Feature Fusion
Senjie Zhang, Shan Zhou, Li Lü, Chang Jia, Jinbo Wang · 2024
The small scale, high stealthiness, and significant destructive potential of Hardware Trojans (HTs) have attracted researchers' attention to detect them accurately. Existing research methods have certain limitations, such as the inability to detect unknown HTs, poor scalability, and incomplete feature representation. To overcome these limitations, we propose DFF, a feature fusion model based on graph neural networks. We represent the register transfer level (RTL) code as an abstract syntax tree (AST) and extract features from different perspectives using graph convolutional networks (GCN) and graph attention networks (GAT). Then, we perform feature fusion based on long short-term memory (LSTM) and utilize the fused features for HT detection. We validated our model on an extended dataset based on Trust-hub [1], and the experimental results demonstrate that our model achieves an average improvement of approximately 6.6% in terms of F1 score compared to other works.