New attack reconstructs private graph data from GNN explanations

Researchers have developed a new attack called PRIVX that can reconstruct hidden graph structures from differentially private Graph Neural Network (GNN) explanations. The attack exploits the Gaussian differential privacy mechanism, treating reconstruction as a reverse diffusion process. Experiments show that PRIVX can achieve high accuracy even with typically deployed privacy budgets, suggesting that differential privacy alone may not be sufficient to protect sensitive graph data. AI

IMPACT Demonstrates that differential privacy may be insufficient for protecting sensitive graph data when GNN explanations are released.

RANK_REASON This is a research paper detailing a novel attack on differentially private GNN explanations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• GDPR
• GNNExplainer
• GraphLIME
• GNN
• PRIVX

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →