SheAttack: A Silhouette Score Motivated Restricted Black-Box Attack on Graphs

Runlin Lei, Haipeng Ding, Zhewei Wei

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.

Keywords: Graph Neural Networks, Graph Adversarial Attack, Graph Representation Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a restricted black-box attack on graphs that handles both homophilic and heterophilic settings.

Abstract:

Graph Neural Networks (GNNs) have gained large popularity in various applications, with their vulnerability against adversarial attacks also being brought up.
Despite the numerous graph attacks proposed, few have focused on the Restrict Black-box attack, where attackers only have access to node features and the graph structure.
Existing works in this setting aim to perform destructive attacks by degrading the quality of victim graphs yet imposing the homophily assumption or requiring high computational complexity.
To address these challenges, we propose the Modified Silhouette Score (MSS) as a measure of a graph's quality, and demonstrate its generalizability across graphs of different homophily levels through theoretical analysis.
Using MSS as the objective, we present SheAttack, an efficient attack that effectively reduces the distinguishability of nodes.
We conduct experiments on both synthetic and real-world graphs to validate the effectiveness of SheAttack in both homophilic and heterophilic settings.
We find that even without prior knowledge of labels or the victim model, our method shows comparable performance to split-unknown white-box attacks.

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: 128