TrojanSDF: improving the performance of node-wise hardware Trojan detection via state distribution and feature fusion

The early detection of Hardware Trojans (HTs) is crucial, as the difficulty of detection and the cost of rollback would be significantly greater in the post-silicon stage. Currently, due to the rise of Graph Neural Network (GNN) and the natural compatibility between circuits and graph structures, training GNN for detection has become a feasible and effective method. However, most deep learning-based detection methods suffer from a lack of dynamic behavioral features reflecting circuit runtime states, resulting in limited detection capabilities and low recall rates. To address this, we propose TrojanSDF, a GNN-based node-wise hardware Trojan (HT) detection method for gate-level netlists without the need for a golden model. TrojanSDF incorporates signal state distributions to represent dynamic behaviors and introduces a feature fusion mechanism within the graph attention network framework. This enables the model to simultaneously learn both static structural features and dynamic behavioral features of circuits, significantly enhancing detection performance and generalization capabilities. Experimental results demonstrate that TrojanSDF improves recall from 86.54% to 96.44% compared to the current state-of-the-art node-level HT detection methods.