PallasGNN: Curriculum-Based Pattern Mining for Robust GNNs

Graph Neural Networks (GNNs) are vulnerable to adversarial perturbations. Existing defenses typically treat all edges equally during training, limiting their ability to distinguish reliable from adversarial connections. Inspired by the human ability to learn progressively and prioritize trustworthy information, we propose PallasGNN, a curriculum-based defense framework, dynamically prioritizes edges based on structural difficulty, enabling the model to learn from simpler, reliable edges first and gradually incorporate more complex, uncertain ones. PallasGNN is easy to implement, compatible with standard GNNs, and supports various difficulty metrics, such as global homophily and the even-hop emphasis from EvenNet, making it applicable to both homophilic and heterophilic graphs. Experiments show that PallasGNN improves robustness by 4–10% over the second-best method under strong adversarial attacks and is orthogonal to existing defense strategies, offering a flexible and practical solution for robust graph learning.