Research map / Graph learning / Graph self-supervised learning
Poincaré Differential Privacy for Hierarchy-Aware Graph Embedding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PoinDP, a hyperbolic (Poincaré ball) differential privacy framework that protects hierarchy-aware GNN node embeddings via personalized hierarchy-aware sensitivity and a Hyperbolic Gaussian Mechanism.
Problem
Hyperbolic GNNs that exploit graph hierarchy leak latent topological information to membership/inference attackers, and existing DP methods measure sensitivity in Euclidean space and so protect hierarchy poorly.
Approach
Learns Poincaré-ball node embeddings to derive inter-hierarchy (radius) and intra-hierarchy (angle) sensitivities, then adds noise from a Gaussian distribution wrapped onto hyperbolic space (Hyperbolic Gaussian Mechanism) with the privacy budget split adaptively between the two via attention.
Key results
- Among privacy-preserving models at epsilon=1, PoinDP reaches weighted-F1 of 75.5 on Citeseer and 83.8 on PubMed, outperforming VANPD, LaP, RdDP and AtDP on those datasets (AtDP is best on Cora at 81.0 vs PoinDP 78.2).
- In the membership inference attack on Photo (epsilon=0.2), PoinDP lowers AUC to 37.4 versus 68.1 for VANPD and 70.3 for LaP.
- Hierarchy-enhanced GCN (GCN+H) raises attack AUC on Photo from 79.7 to 82.4 relative to GCN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove inter-hierarchy sensitivity (ablation) | PoinDP · Cora, Citeseer, PubMed, Computers, Photo | W-F1: PoinDP: 59.9±1.4, 74.0±1.3, 79.4±0.5, 83.8±0.4, 91.9±0.5; without inter: 48.4±2.6, 60.6±4.1, 70.8±1.4, 79.5±1.8, 91.3±0. | Hurt |
| remove intra-hierarchy sensitivity (ablation) | PoinDP · Cora, Citeseer, PubMed, Computers, Photo | W-F1: PoinDP: 59.9±1.4, 74.0±1.3, 79.4±0.5, 83.8±0.4, 91.9±0.5; without intra: 48.8±0.4, 60.1±9.9, 76.6±1.4, 82.6±1.2, 91.3±0. | Hurt |
| remove adaptive privacy-budget allocation (ablation) | PoinDP · Cora, Citeseer, PubMed, Computers, Photo | W-F1: PoinDP: 59.9±1.4, 74.0±1.3, 79.4±0.5, 83.8±0.4, 91.9±0.5; without allocate: 51.2±2.4, 69.5±2.8, 77.2±0.7, 82.7±0.4, 91.5 | Hurt |
| add hierarchical information to the attack model | GCN membership inference attack · PubMed | AUC: 62.8±1.5 -> 63.2±0.2 | Helped |
| add hierarchical information to the attack model | GCN membership inference attack · Photo | AUC: 79.7±2.5 -> 82.4±0.4 | Helped |
| use PoinDP instead of LaP for membership-inference defense | LaP · PubMed | AUC: 51.9±0.5 -> 46.7±2.1 | Helped |
| use PoinDP instead of LaP for membership-inference defense | LaP · Photo | AUC: 70.3±0.1 -> 37.4±0.6 | Helped |
Surprising findings
- Hyperbolic-geometry-based privacy noise is reported to yield higher utility than Euclidean DP baselines, which the authors attribute to noise being better matched to hierarchical structure.
Experimental setup
Datasets: Cora Citeseer PubMed Amazon Computers Amazon Photo
Benchmarks: Cora Citeseer PubMed Computers Photo
Metrics: weighted-F1 micro-F1 AUC precision
Baselines: GCN GAT HyperIMBA VANPD LaP RdDP AtDP
Builds on: Poincaré embeddings Gaussian mechanism wrapped Gaussian distribution GCN
Closest papers by meaning
- On the trade-off between expressivity and privacy in graph representation learning (ICLR 2026 · similarity 0.88)
- Differentially Private Decoupled Graph Convolutions for Multigranular Topology Protection (NeurIPS 2023 · similarity 0.88)
- On the Price of Differential Privacy for Hierarchical Clustering (ICLR 2025 · similarity 0.87)
- Going Deeper into Locally Differentially Private Graph Neural Networks (ICML 2025 · similarity 0.87)
- On provable privacy vulnerabilities of graph representations (NeurIPS 2024 · similarity 0.85)
- Differentially Private Graph Diffusion with Applications in Personalized PageRanks (NeurIPS 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.