Research map / Federated learning and privacy / Differential privacy
Differentially Private Graph Neural Networks for Whole-Graph Classification
Tamara T. Mueller, Johannes Christian Paetzold, Chinmay Prabhakar, Dmitrii Usynin, Daniel Rueckert, Georgios A. Kaissis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with DP-SGD at eps=0.5 instead of SGD | GCN · Synthetic | accuracy: 0.934 -> 0.756 | Hurt |
| train with DP-SGD at eps=5 instead of SGD | GCN · Synthetic | accuracy: 0.934 -> 0.918 | No clear effect |
| train with DP-SGD at eps=5 instead of SGD | GCN · Fingerprint | accuracy: 0.785 -> 0.794 | No clear effect |
| train with DP-SGD at eps=0.5 instead of SGD | GCN · Fingerprint | accuracy: 0.785 -> 0.693 | Hurt |
| train with DP-SGD at eps=1 instead of SGD | GCN · ECG | accuracy: 0.932 -> 0.824 | Hurt |
| train with DP-SGD at eps=10 instead of SGD | GCN · Molbace | accuracy: 0.655 -> 0.609 | Hurt |
| tighten privacy budget from eps=5 to eps=1 | GCN · ECG | accuracy: 0.923 -> 0.824 | Hurt |
| switch GNN architecture among GCN, GAT, GraphSAGE under DP-SGD | DP-SGD training · all datasets | accuracy | No clear effect |
| use DP-SGD with GAT instead of SGD on Molbace at eps=20 | GAT · Molbace | accuracy: 0.726 -> 0.580 | Hurt |
| increase graph size in synthetic data from 10 to 50 nodes | GCN DP-SGD · Synthetic | accuracy | Improved |
| tighten privacy: eps=5 vs eps=0.5 on GNNExplainer overlap with non-DP model | GCN DP-SGD · Fingerprint | IoU(DP, non-DP): 0.765 -> 0.609 | Hurt |
Closest papers by meaning
- Going Deeper into Locally Differentially Private Graph Neural Networks (ICML 2025 · similarity 0.90)
- Can Graph Neural Networks Expose Training Data Properties? An Efficient Risk Assessment Approach (NeurIPS 2024 · similarity 0.87)
- Poincaré Differential Privacy for Hierarchy-Aware Graph Embedding (AAAI 2024 · similarity 0.87)
- GRAIN: Exact Graph Reconstruction from Gradients (ICLR 2025 · similarity 0.85)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.85)
- Rethinking Independent Cross-Entropy Loss For Graph-Structured Data (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
