Reproduce “Preserving Node-level Privacy in Graph Neural Networks” on OpenResearch

About this paper

Zihang Xiang, Tianhao Wang, Di Wang

Researchers at KAUST and the University of Virginia introduced HeterPoisson, a novel method for node-level differential privacy in Graph Neural Networks, achieving higher classification accuracy than existing baselines in high-privacy regimes. The study additionally provides a critical re-evaluation of the "private node embedding" paradigm, proving its inherent utility limitations.

Key insights

• A node's influence on a GNN's gradient update is approximately proportional to its out-degree, which can be leveraged to design tailored sampling strategies for privacy protection.
• Preventing central-peripheral node overlap during sub-graph generation improves utility by allowing for a tighter privacy bound, especially under strong privacy requirements.
• The "private node embedding" paradigm, used by some prior works, suffers from fundamental privacy flaws and an inherent impossibility to simultaneously achieve strong privacy and acceptable utility in graph classification tasks.

Summary by alphaXiv.