Robust Graph Attention for Graph Adversarial Attacks: An Information Bottleneck Inspired Approach

27 Sept 2024 (modified: 15 Jan 2025)ICLR 2025 Conference Withdrawn SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Graph Adversarial Attacks, Robust Graph Attention, Information Bottleneck
TL;DR: We propose a novel and robust graph attention method termed Robust Graph Attention inspired by Information Bottleneck (IB), or RGA-IB, which renders superior robust accuracies for various graph adversarial attacks.
Abstract: Graph Neural Networks (GNNs) have shown exceptional performance in learning node representations for node-level tasks such as node classification. However, traditional message-passing mechanisms solely based on graph structure in GNNs make them vulnerable to adversarial attacks. Attention-based GNNs have been utilized to improve the robustness of GNNs due to their capabilities to selectively emphasize informative signals over noisy or less relevant ones. However, existing works on robust graph attention methods do not realize the correlation between improved robustness and better adherence to the IB principle of attention-based GNNs. In this work, we find that the IB loss of attention-based GNNs is a strong indicator of their robustness against variant graph adversarial attacks. Attention-based GNNs with lower IB loss learn node representations that correlate less to the input training data while aligning better with the target outputs. Due to better adhering to the IB principle, attention-based GNNs with lower IB loss usually show stronger robustness against graph adversarial attacks. Inspired by such observation, we propose a novel graph attention method termed Robust Graph Attention inspired by Information Bottleneck, or RGA-IB, which explicitly minimizes the IB loss of a multi-layer GNN through a carefully designed graph attention mechanism. Extensive experiment results on semi-supervised node classification under variant graph adversarial attacks show that GNNs equipped with RGA-IB exhibit lower IB loss, which indicates better adherence to the IB principle, and show significantly improved node classification accuracy under graph adversarial attacks compared to existing robust GNNs. The code of RGA-IB is available at \url{https://anonymous.4open.science/r/RGA-IB-A47F/}.
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11353
Loading