Clean-Anchor Alignment: Toward Robust and Stable Graph Neural Networks

Graph neural networks (GNNs) are sensitive to small, plausible changes in structure and node features that frequently occur in real data. We formalize robustness as neighborhood stability: the model should yield the same output in a reasonable disturbance range around each clean graph and propose a clean anchor alignment (CAA) framework. A teacher provides a distribution on the clean input, while a student learns on perturbed views and is softly aligned to that clean anchor through a confidence gate and a short schedule that increases alignment after the task head has settled. CAA is architecture agnostic, requires only one additional clean forward per batch, and uses an exponential-moving-average (EMA) teacher without gradient updates, which keeps optimization stable and the cost low. In our implementation, this extra teacher pass keeps total training time in the range of seconds to at most a few tens of seconds on a single GPU. On five standard benchmarks (Cora, Citeseer, PubMed, MUTAG, PROTEINS) and multiple non-adversarial perturbation families, CAA improves Robust@ρ=0.10 by about +2.7–3.9 percentage points over a strong supervised baseline and reduces a prediction-dispersion metric by approximately 29.3%. Under feature-space FGSM attacks, CAA further mitigates the accuracy drop, with relative reductions between 2.5% and 13.1% as the perturbation strength increases. Clean accuracy remains competitive, indicating that the gains in robustness do not come at the expense of performance on unperturbed data.