Kanta Tanahashi, Megha Khosla · ACM SIGSAC Conference on Computer and Communications Security (CCS) 2026 · 2026
DOI: 10.1145/3830454.3846450
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Graph Neural Networks (GNNs) achieve strong performance across many graph-based tasks but remain vulnerable to membership inference (MI). Prior work has shown that graph structure plays a central role in shaping MI risk. Since nodes differ in their structural and feature properties, we ask whether they also face unequal MI risk and find that structurally isolated nodes and nodes with high intra-class feature dissimilarity are particularly exposed. Motivated by this uneven vulnerability, we make the case for systematically exploring self-supervised learning (SSL) as a privacy-mitigation strategy for graph machine learning. As a first step, we investigate contrastive SSL pretraining on a public graph before fine-tuning on private data. In standard GNN training, a low backbone learning rate disproportionately protects the most vulnerable nodes, while SSL preserves utility and provides additional privacy gains across node groups. We then integrate SSL into PrivGNN, improving the privacy--utility trade-off by reducing the privacy budget by more than 50% on benchmark datasets.
No comments yet — start the discussion below.