Robust Graph Learning Against Adversarial Evasion Attacks via Prior-Free Diffusion-Based Structure Purification

• DOI: 10.48550/arxiv.2502.05000
• Publication Date: 2025-02-07
• Authors: Jiayi Luo; Qingyun Sun; Haonan Yuan; Xingcheng Fu; Jianxin Li
• Keywords: Evasion (ethics); Pursuit-evasion

Adversarial evasion attacks pose significant threats to graph learning, with lines of studies that have improved the robustness Graph Neural Networks (GNNs). However, existing works rely on priors about clean graphs or attacking strategies, which are often heuristic and inconsistent. To achieve robust learning over different types diverse datasets, we investigate this problem from a prior-free structure purification perspective. Specifically, propose novel Diffusion-based Structure Purification framework named DiffSP, creatively incorporates diffusion model learn intrinsic distributions purify perturbed structures by removing adversaries under direction captured predictive patterns without relying priors. DiffSP is divided into forward process reverse denoising process, during achieved. avoid valuable information loss an LID-driven nonisotropic mechanism selectively inject noise anisotropically. promote semantic alignment between purified generated reduce generation uncertainty proposed transfer entropy guided mechanism. Extensive experiments demonstrate superior against attacks.