Research map / Federated learning and privacy / Differential privacy
(Amplified) Banded Matrix Factorization: A unified approach to private training
Christopher A. Choquette-Choo, Arun Ganesh, Ryan McKenna, Hugh Brendan McMahan, J Keith Rush, Abhradeep Guha Thakurta, Zheng Xu
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 banded matrix (b=342) instead of DP-SGD (1 band) (ablation) | matrix mechanism · prefix sums, n=2052 | RMSE under b>=342-min-sep: 9.63 -> 1.05 | Improved |
| use banded matrix (b=342) instead of Optimal TreeAgg | matrix mechanism · prefix sums, n=2052 | RMSE under b>=342-min-sep: 1.53 -> 1.05 | Improved |
| reduce bands from 342 to 128 (ablation) | BandMF · prefix sums, n=2052 | RMSE under b>=342-min-sep: 1.05 -> 1.27 | Hurt |
| add equal column norm constraint (ablation) | MF (b=342) · prefix sums, n=2052 | RMSE under b>=342-min-sep: 1.08 -> 1.05 | Improved |
| add equal column norm constraint (ablation) | MF (b=342) · prefix sums, n=2052 | RMSE under (6,342)-participation: 1.04 -> 1.05 | Hurt |
| switch from (k,b)-participation to b-min-sep-participation | MF without column normalization · prefix sums | RMSE: 1.00 -> 1.05 | Mixed |
| BandMF with amplification vs prior MF | BandMF · CIFAR-10, eps in [2,5] | accuracy: ~1 percentage point | Improved |
| BandMF with amplification vs DP-SGD | BandMF · CIFAR-10, eps below 2 | accuracy: 3 percentage points | Improved |
| BandMF vs DP-SGD as eps becomes very small | BandMF · CIFAR-10 | accuracy | No clear effect |
| combine amplification with MF | BandMF · StackOverflow NWP, eps in [1,8] | accuracy | Improved |
| use BandMF instead of Multi-epoch MF at eps=16 | BandMF · StackOverflow NWP | accuracy | No clear effect |
| use BandMF instead of Online TreeAgg | BandMF · production Spanish NWP FL | privacy guarantee: 0.52 -> 0.24 | Improved |
Closest papers by meaning
- DMM: Distributed Matrix Mechanism for Differentially-Private Federated Learning Based on Constant-Overhead Linear Secret Resharing (ICML 2025 · similarity 0.90)
- Scaling up the Banded Matrix Factorization Mechanism for Large Scale Differentially Private ML (ICLR 2025 · similarity 0.90)
- Banded Square Root Matrix Factorization for Differentially Private Model Training (NeurIPS 2024 · similarity 0.90)
- Sampling-Free Privacy Accounting for Matrix Mechanisms under Random Allocation (NeurIPS 2026 · similarity 0.90)
- Privacy Amplification for Matrix Mechanisms (ICLR 2024 · similarity 0.89)
- Near-Exact Privacy Amplification for Matrix Mechanisms (ICLR 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.
