Poster: AuditVotes: A Framework towards Deployable Certified Robustness for GNNs
Yuni Lai, Kai Wen Zhou · 2024
Graph Neural Networks (GNNs) are powerful but vulnerable to adversarial attacks, necessitating the research on certified robustness that can provide GNNs with robustness guarantees. Existing randomized smoothing methods struggle with a trade-off between utility and robustness due to high noise levels. We introduce AuditVotes, which integrates randomized smoothing with two components, augmentation and conditional smoothing, aiming to improve data and vote quality. We instantiated AuditVotes with simple strategies, and preliminary results demonstrate its significant promise in enhancing certified robustness, representing a substantial step toward deploying certifiably robust GNNs in real-world applications.