Defending adversarial attacks in Graph Neural Networks via tensor enhancement

Graph Neural Networks (GNNs) have demonstrated remarkable success across diverse fields, yet remain susceptible to subtle adversarial perturbations that significantly degrade performance. Addressing this vulnerability remains a formidable challenge. Current defense strategies focus on edge-specific regularization within adversarial graphs, often overlooking the inter-edge structural dependencies and the interplay of various robustness attributes. This paper introduces a novel tensor-based framework for GNNs, aimed at reinforcing graph robustness against adversarial influences. By employing tensor approximation, our method systematically aggregates and compresses diverse predefined robustness features of adversarial graphs into a low-rank representation. This approach harmoniously combines the integrity of graph structure and robustness characteristics. Comprehensive experiments on real-world graph datasets demonstrate that our framework not only effectively counters diverse types of adversarial attacks but also surpasses existing leading defense mechanisms in performance.