DPAR: Decoupled Graph Neural Networks with Node-Level Differential Privacy.

Area of Science:

• Machine Learning
• Graph Analytics
• Data Privacy

Background:

• Graph Neural Networks (GNNs) excel at learning from graph data.
• Training GNNs raises privacy concerns regarding sensitive node features and graph structure.
• Node-level differential privacy (DP) for GNNs is challenging due to unbounded node sensitivity in message passing.

Purpose of the Study:

• To develop a node-level differentially private GNN training method.
• To improve the privacy-utility tradeoff in GNNs.
• To protect sensitive node features and graph structure information.

Main Methods:

• Proposed a Decoupled GNN with Differentially Private Approximate Personalized PageRank (DPAR).
• Decoupled feature projection and message passing using a DP PageRank algorithm.
• Utilized top-K neighbors identified by PageRank for feature aggregation, bounding node sensitivity.

Main Results:

• Achieved improved privacy-utility tradeoff compared to existing methods.
• Bounded node sensitivity by avoiding layer-wise message passing.
• Demonstrated superior utility for DPAR at equivalent node DP levels.

Conclusions:

• DPAR offers a robust solution for privacy-preserving GNN training.
• The method effectively balances privacy guarantees with model utility.
• DPAR advances the field of private graph representation learning.