Enhancing Graph Injection Attacks Through Over-Smoothing Amplification

Haoqiang Zhang, Yonggang Zhang, Rongfei Zeng, Bo Han, Hao Wang, Defu Lian, Enhong Chen

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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, Adversarial Machine Learning, Graph Adversarial Attack

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

TL;DR: This paper incorporates over-smoothing into the GIA attack and proposes a universal framework Over-Smoothing adversarial Injection (OSI) that can be combined with any GIA method to improve the attack power.

Abstract:

Graph Injection Attack (GIA) on Graph Neural Networks (GNNs) has attracted significant attention due to its serious threats to the deployment of GNNs by carefully injecting a few malicious nodes into graphs.
Existing GIA defense methods mostly follow a framework similar to the defense depicted in the images. Instead, we aim to enhance the attack capabilities of GIA by studying the properties of the graph itself. Considering the negative impact of the over-smoothing issue in GNNs, we propose $\textit{O}$ver-$\textit{S}$moothing adversarial $\textit{I}$njection (OSI), a universal method that can be combined with any GIA to enhance the attack power by amplifying the over-smoothing on graphs. Specifically, OSI proposes two metrics to evaluate the over-smoothing of the graph. We prove that these two metrics are highly correlated with singular values of the adjacency matrix. Thus, OSI further introduces a Smooth Injection Loss (SIL) which aims to smooth the singular values. By fine-tuning the adjacency matrix using SIL, OSI can amplify over-smoothing and enhance the attack power of GIA. We conduct experiments on 4 benchmark datasets and the state-of-the-art GNNs and GIA attacks. Empirical experiments show that OSI can significantly improve the attack capabilities of existing GIA attacks on different defense GNN models in most scenarios.

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: 1928