Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Backdoor or Manipulation? Graph Mixture of Experts Can Defend Against Various Graph Adversarial Attacks

Yuyuan Feng, Bin Ma, Enyan Dai

TL;DR

This paper addresses the unified defense of graph neural networks against multiple adversarial threats, including backdoors, edge manipulation, and node injection. It introduces RGMoE, a Graph Mixture-of-Experts framework that (1) diversifies expert decision logics via an MI-based mechanism to focus on different neighborhood structures, and (2) trains a robustness-aware router to route perturbed samples to robust experts. The approach achieves superior robustness across diverse attacks and datasets, with scalability maintained by sparse routing and a modular MoE design; increasing the number of experts further improves resilience while preserving clean accuracy. The work advances practical graph security by offering a scalable, unified defense that leverages the intrinsic diversity and routing flexibility of MoE architectures, and points to future enhancements in trustworthy GNNs through adaptive, robust MoE systems.

Abstract

Extensive research has highlighted the vulnerability of graph neural networks (GNNs) to adversarial attacks, including manipulation, node injection, and the recently emerging threat of backdoor attacks. However, existing defenses typically focus on a single type of attack, lacking a unified approach to simultaneously defend against multiple threats. In this work, we leverage the flexibility of the Mixture of Experts (MoE) architecture to design a scalable and unified framework for defending against backdoor, edge manipulation, and node injection attacks. Specifically, we propose an MI-based logic diversity loss to encourage individual experts to focus on distinct neighborhood structures in their decision processes, thus ensuring a sufficient subset of experts remains unaffected under perturbations in local structures. Moreover, we introduce a robustness-aware router that identifies perturbation patterns and adaptively routes perturbed nodes to corresponding robust experts. Extensive experiments conducted under various adversarial settings demonstrate that our method consistently achieves superior robustness against multiple graph adversarial attacks.

Backdoor or Manipulation? Graph Mixture of Experts Can Defend Against Various Graph Adversarial Attacks

TL;DR

This paper addresses the unified defense of graph neural networks against multiple adversarial threats, including backdoors, edge manipulation, and node injection. It introduces RGMoE, a Graph Mixture-of-Experts framework that (1) diversifies expert decision logics via an MI-based mechanism to focus on different neighborhood structures, and (2) trains a robustness-aware router to route perturbed samples to robust experts. The approach achieves superior robustness across diverse attacks and datasets, with scalability maintained by sparse routing and a modular MoE design; increasing the number of experts further improves resilience while preserving clean accuracy. The work advances practical graph security by offering a scalable, unified defense that leverages the intrinsic diversity and routing flexibility of MoE architectures, and points to future enhancements in trustworthy GNNs through adaptive, robust MoE systems.

Abstract

Extensive research has highlighted the vulnerability of graph neural networks (GNNs) to adversarial attacks, including manipulation, node injection, and the recently emerging threat of backdoor attacks. However, existing defenses typically focus on a single type of attack, lacking a unified approach to simultaneously defend against multiple threats. In this work, we leverage the flexibility of the Mixture of Experts (MoE) architecture to design a scalable and unified framework for defending against backdoor, edge manipulation, and node injection attacks. Specifically, we propose an MI-based logic diversity loss to encourage individual experts to focus on distinct neighborhood structures in their decision processes, thus ensuring a sufficient subset of experts remains unaffected under perturbations in local structures. Moreover, we introduce a robustness-aware router that identifies perturbation patterns and adaptively routes perturbed nodes to corresponding robust experts. Extensive experiments conducted under various adversarial settings demonstrate that our method consistently achieves superior robustness against multiple graph adversarial attacks.
Paper Structure (35 sections, 20 equations, 11 figures, 8 tables, 1 algorithm)

This paper contains 35 sections, 20 equations, 11 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Preliminaries of Mixture of Experts and Graph Adversarial Attacks
  4. Robustness Analysis of Graph MoE
  5. Problem Formulation
  6. Methodology
  7. Mixture of Experts Architecture
  8. Diversifying Expert Decision Logics
  9. Capturing Logic by Mutual Information.
  10. Logic Diversity Loss.
  11. Diverse Experts Training
  12. Robustness-Aware Router Training
  13. Identification of Potential Perturbed Samples.
  14. Encouraging to Route to Robust Experts.
  15. Final Objective Function
  16. ...and 20 more sections

Figures (11)

  • Figure 1: Performance distribution of individual experts in Graph MoE under various attacks. We report Attack Success Rate (ASR) for backdoor attacks and accuracy drop compared with clean graphs for manipulation attacks.
  • Figure 2: (a) Impacts of existing diversity loss to the proportion of robust individual experts. (b) Rates of routing to roust and non-robust experts.
  • Figure 3: Overall framework of RGMoE.
  • Figure 4: Impact of model capacity under UGBA attack.
  • Figure 5: Ablation studies under UGBA across datasets.
  • ...and 6 more figures