Research map / Federated learning and privacy / Differential privacy
Optimal privacy guarantees for a relaxed threat model: Addressing sub-optimal adversaries in differentially private machine learning
Georgios Kaissis, Alexander Ziller, Stefan Kolek, Anneliese Riess, Daniel Rueckert
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 |
|---|---|---|---|
| calibrate DP-SGD noise to the relaxed threat model (RTM) guarantee instead of the standard DP threat model | DP-SGD ResNet-9 · CIFAR-10 at (1, 1e-5)-DP | Top-1 accuracy: 0.67 -> 0.81 | Improved |
| calibrate DP-SGD noise to the RTM guarantee instead of the DPTM | DP-SGD BERT · Stanford SNLI at (1, 1.8e-6)-DP | Top-1 accuracy: 0.78 -> 0.80 | Improved |
| calibrate DP-SGD noise to the RTM guarantee instead of the DPTM | DP-SGD ResNet-18 · ImageNet at (10, 1e-6)-DP | Top-1 accuracy: 0.46 -> 0.48 | Improved |
| increase dimension from d = 1 to d = 30 for the Gaussian mechanism under RTM | Gaussian mechanism trade-off function · Delta2/sigma = 1 | lambda-DR privacy (blessing of dimensionality) | Improved |
| adversary cannot reduce effective gradient dimension (50-dim latent linear network) versus d = 1 worst case | gradient-space MI attack · Diabetes dataset, Delta2/sigma = 2 | hypothesis test power | Hurt |
| switch from DPTM to RTM adversary (no access to candidate record) | Gaussian mechanism MI attack · power at alpha = 10^-3 vs Delta2/sigma | adversary power 1 - beta | Hurt |
| add Poisson subsampling (rate 0.3) to the mechanism under RTM | Gaussian mechanism · worst-case 1-d task | adversary power loss | Improved |
| audit with Dirac-canary gradient-space attack instead of LiRA logit-space attack | empirical privacy auditing · 1-d synthetic binary task, Delta2/sigma = 1 | tightness to theoretical bound | Improved |
| evaluate reconstruction attack bound under RTM instead of DPTM (prior kappa = 0.1) | DP-SGD on CIFAR-10, SNLI, ImageNet · Reconstruction robustness | probability of reconstruction | Improved |
Closest papers by meaning
- A Unified Framework for Adversary-Aware Differential Privacy Bounds (NeurIPS 2026 · similarity 0.90)
- Provable Robustness against Backdoor Attacks via the Primal-Dual Perspective on Differential Privacy (NeurIPS 2026 · similarity 0.89)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.89)
- Attack-Aware Noise Calibration for Differential Privacy (NeurIPS 2024 · similarity 0.89)
- The Limits of Differential Privacy in Online Learning (NeurIPS 2024 · similarity 0.88)
- Tighter Privacy Auditing of DP-SGD in the Hidden State Threat Model (ICLR 2025 · similarity 0.88)
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.
