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

Yuyuan Feng, Bin Ma, Enyan DaiarXiv 2025 · DOI · arXiv

Introduction

Graph Neural Networks (GNNs) have become the standard for modeling relational data, yet they remain notoriously susceptible to adversarial perturbations. Because GNNs rely on message-passing mechanisms—aggregating features from local neighborhoods—attackers can easily degrade model performance by manipulating edges, injecting malicious nodes, or embedding structural backdoors.

Existing defense strategies are often fragmented, typically specializing in a single threat type while remaining vulnerable to others. Furthermore, many robust GNNs suffer from $O(N^2)$ complexity, limiting their utility in large-scale, real-world applications. This paper introduces RGMoE (Robust Graph Mixture of Experts), a unified and scalable defense framework designed to neutralize multiple attack vectors simultaneously.

The RGMoE Framework

RGMoE leverages a Sparse Mixture of Experts (MoE) architecture. In this setup, a "router" dynamically assigns input nodes to a subset of specialized sub-networks, or "experts." Unlike traditional GNNs that process all nodes through a monolithic structure, RGMoE forces experts to specialize in different structural patterns. This architectural choice ensures that an attack targeting a specific local feature cannot fool the entire system at once.

The RGMoE architecture integrates a robustness-aware router and diverse experts to defend against multiple graph attack types.

Mechanisms for Robustness

The core innovation of RGMoE lies in two primary mechanisms designed to amplify the inherent robustness of MoE architectures:

1. MI-Based Logic Diversity Loss

A significant risk in MoE systems is that experts might converge on similar decision logic, making them collectively vulnerable to the same adversarial triggers. To prevent this, the authors implement a Mutual Information (MI) based diversity loss ($L_{ED}$). This penalty forces different experts to rely on distinct local neighborhoods for their predictions. By diversifying the "logic" of each expert, the framework ensures that even if an attacker successfully perturbs a specific structural feature, a significant portion of the expert pool remains unaffected.

Comparison of expert Attack Success Rates with and without Logic Diversity Loss ($L_{ED}$).

2. Robustness-Aware Router

The second pillar of RGMoE is a router that intelligently identifies and bypasses poisoned experts. The authors observed that while experts generally agree on clean nodes, they exhibit high variance—or "disagreement"—when processing poisoned nodes. The Robustness-Aware Router uses this variance as an unsupervised signal to detect potential attacks. When high disagreement is detected, the router adaptively shifts its confidence away from the majority (potentially fooled) experts toward those offering alternative, robust predictions.

Experimental Performance and Scalability

RGMoE was evaluated against state-of-the-art attacks, including backdoor injections (GTA, UGBA) and edge manipulations (PRBCD), across benchmarks like Cora, Pubmed, Flickr, and OGB-Arxiv.

  • Backdoor Defense: RGMoE demonstrates remarkable efficacy, reducing Attack Success Rates (ASR) from nearly 100% in standard GCNs to approximately 0%, without sacrificing accuracy on clean data.
  • Unified Defense: Unlike specialized baselines, RGMoE maintains high performance across edge manipulation and node injection scenarios simultaneously.
  • Scalability: By utilizing sparse routing and avoiding dense adjacency matrix computations, RGMoE maintains linear time complexity $O(N)$. This allows it to scale to large datasets like OGB-Arxiv, where many traditional robust GNNs fail due to memory or compute constraints.

Conclusion

RGMoE represents a shift from specialized GNN defenses toward a unified, architectural approach to robustness. By combining the efficiency of Sparse MoE with targeted diversity and awareness mechanisms, the framework provides a scalable solution for deploying trustworthy GNNs in adversarial environments. Future research will continue to explore the untapped potential of MoE for building inherently safe and powerful graph models.

Summary updated

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.
BibTeX
@misc{feng2025backdoor,
  author = {Yuyuan Feng and Bin Ma and Enyan Dai},
  title = {{Backdoor or Manipulation? Graph Mixture of Experts Can Defend Against Various Graph Adversarial Attacks}},
  year = {2025},
  eprint = {2510.15333},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.15333},
}
Figure