Equipping Graph Autoencoders: Revisiting Masking Strategies from a Robustness Perspective
Shuhan Song, Ping Li, Ming Dun, Yuan Zhang, Huawei Cao, Xiaochun Ye · Society for Industrial and Applied Mathematics eBooks · 2025
Masked Graph Autoencoders (MGAEs), represented by GraphMAE and GraphMAE2, which utilize masked feature (or structure) reconstruction strategies, have demonstrated the potential to surpass contrastive learning. However, current masked reconstruction strategies primarily rely on random strategies, only prove effective on reliable graph data. Therefore, these popular methods face immediate robustness deficiencies issues. Firstly, when the graph is unreliable or under adversarial attacks, the selection of nodes for masked reconstruction has a significant impact on downstream tasks. Secondly, the reconstructed features contains redundant components. In this paper, to overcome the non-robustness caused by randomness, we provide a theoretical analysis and evaluation of the robustness of state-of-the-art MGAEs. Additionally, we design two lightweight plug-and-play tools: Box-Based Weighted Reliability Ranking Masking Strategy and Decoupled Feature Reconstruction. Without incurring additional time overhead, these tools provide a defense armor against adversarial attacks for MGAEs, significantly boosting the robustness performance of downstream tasks. Extensive experiments on real-world graphs attacked by various attacks demonstrate our designs have a considerable robust expressive ability. Especially on datasets with large perturbations, the defense performance could even be improved by up to 20%.