AURA: Structural and Semantic Calibration for Robust Federated Graph Learning

ICLR 2026 Conference Submission542 Authors

Keywords: Federated Learning, Graph Learning, Robustness, Geometric Learning

Abstract: Training highly generalizable server model necessitates requires data from multiple sources in Federated Graph Learning. However, noisy labels are increasingly undermining federated system due to the propagation of erroneous information between nodes. Compounding this issue, significant variations in data distribution among clients make noise node detection more challenging. In our work, we propose an effective structural and semantic calibration framework for Robust Federated Graph Learning, AURA. We observe that spectral discrepancies across different clients adversely affect noise detection. To address this, we employs SVD for self-supervision, compelling the model to learn an intrinsic and consistent structural representation of the data, thereby effectively attenuating local high-frequency perturbations induced by noisy nodes. We introduce two metrics, namely "Depth Influence" and "Breadth Influence". Based on these metrics, the framework judiciously selects and aggregates the most consensual knowledge from the class prototypes uploaded by each client. Concurrently, clients perform knowledge distillation by minimizing the KL divergence between their local model's output distribution and that of the global model, which markedly enhances the model's generalization performance and convergence stability in heterogeneous data environments. AURA demonstrates remarkable robustness across multiple datasets, for instance, achieving a $7.6\%$ $\uparrow$ F1-macro score under a 20\%-uniform noise on Cora. The code is available for anonymous access at \url{https://anonymous.4open.science/r/AURA-F351/}.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 542