Differentially Private Graph Neural Networks for Whole-Graph Classification.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Privacy-Preserving Data Analysis

Background:

• Graph Neural Networks (GNNs) excel in analyzing complex networks but often handle sensitive data.
• Existing privacy techniques for GNNs are underexplored due to graph structure challenges.
• Differential privacy offers formal guarantees for sensitive data analysis.

Purpose of the Study:

• To develop a framework for differential private graph-level classification.
• To enable privacy-preserving deep learning on multi-graph datasets.
• To evaluate the performance and scalability of private GNNs.

Main Methods:

• Implemented a framework for differential private graph classification.
• Utilized differentially private stochastic gradient descent (DP-SGD) for training.
• Applied explainability techniques to compare private and non-private learned representations.

Main Results:

• DP-SGD can be applied to graph classification with reasonable utility.
• Evaluated the impact of GNN architectures and hyperparameters on private model performance.
• Demonstrated scalability on a large medical dataset.

Conclusions:

• The proposed framework facilitates privacy-preserving graph classification.
• The study provides a baseline for future research in private GNNs.
• Privacy-preserving GNNs can learn comparable representations to non-private models.