On Measuring Unnoticeability of Graph Adversarial Attacks: Observations, New Measure, and Applications

The paper introduces HideNSeek, a learnable measure for assessing graph adversarial attacks' noticeability. It addresses limitations in existing measures by learning to distinguish original from potential attack edges and conducting imbalance-aware aggregation of edge scores. HideNSeek effectively mitigates bypass and overlooking issues, outperforming eleven competitors in distinguishing attack edges under various methods. Additionally, the learnable edge scorer (LEO) boosts the performance of robust graph neural networks by removing attack-like edges.