D4A: An efficient and effective defense across agnostic adversarial attacks

Xianxian Li, Zeming Gan, Yan Bai, Linlin Su, De Li, Jinyan Wang

Abstract: Highlights•Investigate four potential abnormal behaviors of GNNs under adversarial attacks from the perspective of posterior distribution shifts.•Identify the unfitting abnormal behavior of smoothing-based GNNs under adversarial attacks.•Propose smooth-less message passing to address model unfitting, enhancing model tolerance to structure perturbations.•Introduce CMD metric to align posterior distributions, improving model robustness against adversarial attacks.