Research map / Federated learning and privacy / Privacy attacks and secure learning
Bounding training data reconstruction in DP-SGD
Jamie Hayes, Borja Balle, Saeed Mahloujifar
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 new ReRo bound instead of prior RDP-based bound | DP-SGD reconstruction upper bound · advantage bound, prior size 100, sigma=1 | advantage upper bound: 0.346 -> 0.077 | Improved |
| use the new ReRo bound instead of prior Monte-Carlo bound | DP-SGD reconstruction upper bound · prior size 10, sigma=0.5 | advantage upper bound: 0.771 -> 0.737 | Improved |
| increase noise sigma | new ReRo upper bound · prior size 10 | advantage upper bound: 0.737 -> 0.080 | Improved |
| replace gradient-based prior-oblivious attack with prior-aware attack | reconstruction attack · CIFAR-10 | reconstruction success | Improved |
| increase size of the prior | upper bound and attack · MNIST | reconstruction success | Hurt |
| reduce sampling rate q from 0.99 to 0.01 at fixed eps=4 | upper bound · MNIST | reconstruction upper bound: 0.35 -> 0.20 | Improved |
| reduce sampling rate q from 0.99 to 0.01 at fixed eps=4 | improved prior-aware attack · MNIST | attack success: 0.32 -> 0.15 | Improved |
| use improved attack using only top qT inner products (ablation) | prior-aware attack · MNIST q=0.02 | attack success | Improved |
| increase T or q | upper bound · DP-SGD | reconstruction upper bound | Hurt |
| use eps in {10,100} with RDP-based bound | prior RDP bound · CIFAR-10 | bound | Hurt |
| larger model size (7,960 vs 36.5M parameters) | reconstruction attack tightness · MNIST vs CIFAR-10 | attack tightness | No clear effect |
| more samples for gamma estimation | Algorithm 1 · gamma std | std of gamma: <0.019 -> <0.0017 | Improved |
Closest papers by meaning
- Rethinking DP-SGD in Discrete Domain: Exploring Logistic Distribution in the Realm of signSGD (ICML 2024 · similarity 0.88)
- Tighter Privacy Auditing of DP-SGD in the Hidden State Threat Model (ICLR 2025 · similarity 0.87)
- Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement (NeurIPS 2024 · similarity 0.86)
- Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach (ICLR 2024 · similarity 0.86)
- Unifying Re-Identification, Attribute Inference, and Data Reconstruction Risks in Differential Privacy (NeurIPS 2025 · similarity 0.86)
- Optimal Differentially Private Model Training with Public Data (ICML 2024 · similarity 0.86)
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.
