Research map / Federated learning and privacy / Privacy attacks and secure learning
Gaussian Membership Inference Privacy
Tobias Leemann, Martin Pawelczyk, Gjergji Kasneci
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 |
|---|---|---|---|
| use the analytical LRT-based GLiR attack with known gradient distribution parameters | GLiR attack on simulated gradients · Simulated gradient distribution, single step | match to analytical trade-off | Improved |
| estimate gradient mean and covariance from background data instead of knowing them (ablation) | GLiR attack · Real model gradients on Purchase, CIFAR-10, Adult | attack strength vs analytical curve | Hurt |
| compose five SGD steps instead of one | GLiR attack · CIFAR-10 at FPR=0.25 | attack performance (TPR): mu = 1.14 -> mu = 2.54 | Improved |
| compose five SGD steps instead of one | GLiR attack · Purchase and Adult | attack performance (TPR) | No clear effect |
| calibrate noise for mu-GMIP instead of mu-GDP at the same mu | noisy SGD · CIFAR-10, Purchase, Adult | test accuracy | Improved |
| increase mu above 0.86 under mu-GMIP | noisy SGD · CIFAR-10 | required noise | Improved |
| push mu toward 0 (high privacy) for both GDP and GMIP | noisy SGD · CIFAR-10, Purchase, Adult | test accuracy | Hurt |
| use loss-based attacks such as LiRA instead of the gradient attack | membership inference attack · Appendix C.4 comparison | attack success rate | Hurt |
| increase d relative to batch size n | one-step SGD bound · Theory (Corollary 5.1) | attacker success | Hurt |
| increase gradient susceptibility K | one-step SGD bound · Theory (Theorem 5.1) | attacker success | Hurt |
| increase d and n so the trade-off converges to the Gaussian curve | Theorem 5.1 trade-off function · Figure 1 (n=4,d=3; n=10,d=7; n=22,d=15) | closeness to mu-GMIP curve | Improved |
Closest papers by meaning
- Theoretically Unmasking Inference Attacks Against LDP-Protected Clients in Federated Vision Models (ICML 2025 · similarity 0.89)
- Low-Cost High-Power Membership Inference Attacks (ICML 2024 · similarity 0.89)
- Mitigating Privacy Risk in Membership Inference by Convex-Concave Loss (ICML 2024 · similarity 0.88)
- Chameleon: Increasing Label-Only Membership Leakage with Adaptive Poisoning (ICLR 2024 · similarity 0.88)
- Assessing Per-Sample Membership Inference Vulnerability without Retraining (NeurIPS 2026 · similarity 0.88)
- Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning (NeurIPS 2025 · similarity 0.87)
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.
