GJDNet: Robust Graph Neural Networks via Joint Disentangled Learning Against Adversarial Attacks
Title: GJDNet: Enhancing Graph Neural Network Resilience Through Joint Disentangled Learning in the Face of Adversarial Threats
Abstract:
Graph Neural Networks (GNNs) face significant vulnerabilities to adversarial attacks, which operate by fundamentally reversing connectivity patterns. Specifically, these attacks introduce disassortative edges into assortative graphs and vice versa. This structural inversion generates mismatches between structure and features, thereby disrupting the neighborhood aggregation process across various graph types. Current defensive strategies, however, are constrained; they typically assume fixed assortativity and treat neighborhoods as uniform entities, or they depend on standard softmax classifiers that overlook representation shifts caused by perturbations.
To address these limitations, we adopt a robustness-focused approach that jointly disentangles node representations and decision spaces. This method isolates the impact of perturbations while ensuring that decision regions remain distinctly separated. Building on this foundation, we introduce the Graph Joint Disentanglement Network (GJDNet), a comprehensive framework designed for robust node classification across varying degrees of graph assortativity.
GJDNet strengthens robustness at both the representation and decision-making levels. At the representation level, it utilizes feature-driven soft structural disentanglement combined with skewness-aware neighbor filtering to mitigate structure-feature mismatches induced by perturbations. At the decision level, it incorporates a Spherical Decision Boundary (SDB) to enhance intra-class compactness and inter-class separation within the embedding space, thus stabilizing decision boundaries against perturbations. Our theoretical analysis elucidates the efficacy of the proposed disentangled representation and decision mechanisms, while comprehensive experimental results confirm that GJDNet maintains superior robustness across graphs with diverse connectivity regimes.
Source: arXiv Generated at: 2026-06-02 00:00:00 UTC
