Reinforcement learning-based secure training for adversarial defense in graph neural networks

Dongdong An, Yi Yang, Xin Gao, Hongda Qi, Yang Yang, Xin Ye, Maozhen Li, Qin Zhao

Abstract: Highlights•Introduces RL-based secure training algorithm using MDP and deep Q-learning to defend against GNN adversarial attacks.•Enables formal verification of GNN training via model transformation, ensuring security with tools like Prism.•Achieves higher accuracy against node-level attacks than Robust/Median GCN across multiple datasets.