Topology-Driven Explainable GNNs for Trojan Detection in Deep Learning

Deep neural networks (DNNs) are widely used in critical applications such as autonomous systems, healthcare, and finance. However, their increasing deployment raises security concerns, particularly against adversarial threats like Trojan attacks. These attacks embed hidden triggers during training, causing misclassifications when activated while maintaining normal performance otherwise, making them difficult to detect. We propose a novel graph-based detection framework that models neuron activations as graph structures, capturing topological anomalies indicative of Trojan behavior. We leverage Graph Neural Networks (GNNs) to identify these structural deviations effectively. Additionally, we integrate Explainable AI (XAI) techniques to enhance interpretability. This not only improves trust and transparency in our framework but also aids in understanding the fundamental characteristics of backdoored networks, potentially guiding the development of more robust defense mechanisms.