Topology-aware dynamic co-evolution for robust graph learning.

Abstract

Graph Neural Networks (GNNs) have achieved remarkable success in learning from graph-structured data. However, recent studies have revealed their vulnerability to adversarial attacks, where imperceptible perturbations to the graph structure can easily mislead predictions. Existing defense methods predominantly rely on global heuristic strategies, which often lead to over-smoothing in deep architectures and fail to capture localized, fine-grained structural anomalies. In this paper, we propose Topology-Aware Dynamic Co-evolution (TADC), a novel framework that reformulates graph defense as a closed-loop optimization process. Unlike static purification paradigms, TADC introduces a closed-loop dynamic co-evolution framework, in which graph topology and model parameters are jointly optimized through validation-guided meta-gradients, enabling continuous and adaptive structural refinement during training. To further ensure the reliability of topology updates, we identify the topology-agnostic limitation of conventional meta-gradients and propose a topology-aware gradient calibration mechanism, which incorporates structural priors to guide gradient directions toward more plausible and semantically consistent graph modifications. Concretely, TADC first localizes vulnerable nodes via a structural-semantic consistency probe and constructs an initialized graph through topology-aware refinement. It then performs structure decoupling to disentangle reliable and potentially adversarial edges into distinct channels. Building upon this decomposition, a fully differentiable attention-based architecture DeepGCN enables end-to-end optimization of the co-evolution process. Extensive experiments on multiple benchmark datasets demonstrate that TADC consistently outperforms state-of-the-art methods across various attack settings, with particularly pronounced advantages under strong adversarial perturbations and high perturbation rates. Our code is available at https://github.com/robustgcn/TADC.