Releasing Graph Neural Networks with Differential Privacy Guarantees

This repository contains additional details and reproducibility of the stated results.

Motivation:

Real-world graphs, such often are associated with sensitive information about individuals and their activities. Hence, they cannot always be made public. Moreover, graph neural networks (GNNs) models trained on such sensitive data can leak significant amount of information e.g via membership inference attacks.

Goal of the Paper:

Release a GNN model that is trained on sensitive data yet robust to attacks. Importantly, it should have differential privacy (DP) guarantees.

Proposed Method

Results

Privacy-accuracy Tradeoff

Privacy Budget vs Noise

Membership Inference Attacks

Performance on Attack 1

Performance on Attack 2

Reproducing Results

PrivGNN can be executed via

python3 knn_graph.py

Changing Parameters & Setting

Please adjust the all_config.py file to change parameters, such as the datasets, K, lambda, attack, baselines, or gamma.