GJDNet: Robust Graph Neural Networks via Joint Disentangled Learning Against Adversarial Attacks
Researchers have developed GJDNet, a novel framework designed to enhance the robustness of Graph Neural Networks (GNNs) against adversarial attacks. These attacks exploit structural inversions in graph connectivity, leading to mismatches that disrupt GNN performance. GJDNet addresses this by disentangling node representations and decision spaces, isolating the effects of perturbations and ensuring clearer decision boundaries. The framework utilizes feature-driven soft structural disentanglement and a Spherical Decision Boundary mechanism to improve accuracy and stability across various graph types. AI
IMPACT Enhances GNN security, potentially enabling more reliable deployment in sensitive applications.