Graph Neural Networks forHardware Vulnerability Detection : A Feasibility Study

Sammanfattning

The globalization of the Integrated Circuit (IC) supply chain has introduced severe security vulnerabilities, one of the most dangerous being the stealthy insertion of malicious hardware, known as hardware Trojans (HTs). However, current automated HT localization and detection techniques often suffer from high computational overhead, opaque, incomprehensible decision-making processes, and blurry results caused by data oversmoothing. This thesis presents a feasibility study of a new, decoupled machine learning pipeline that separates circuit-level HT detection from the gate-level localization. To optimize the detection process, the proposed architecture combines the benefits of a Graph Attention Network with added Jumping Knowledge Connections (GAT+JK). The proposed GNN model successfully detects hardware Trojans, achieving a final average macro F1 score of 92.27% across a 5-fold split. The pipeline then introduces the GNNExplainer tool to transparently map the functional neighborhoods that were most important in the GNN’s decision-making process to the infected circuit regions. Strictly quantitative metrics, such as Intersection over Union or the F1 score, were unable to accurately reflect the pipeline’s success, but structural analysis confirmed that an explainability tool can accurately isolate malicious logic. Ultimately, this research establishes the successful integration of explainability tools with Graph Neural Networks, effectively mitigating the ’black box’ nature of Machine Learning models, and offering a scalable, trustworthy, and resource-efficient framework for modern Hardware security verification.