Research map / Federated learning and privacy / Differential privacy
Multi-Epoch Matrix Factorization Mechanisms for Private Machine Learning
Christopher A. Choquette-Choo, Hugh Brendan McMahan, J Keith Rush, Abhradeep Guha Thakurta
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 |
|---|---|---|---|
| Optimize the factorization directly for (k,b)=(20,100) multi-epoch participation rather than use the single-epoch MF baseline | Matrix factorization mechanism · StackOverflow next-word prediction at ε=8.84 | accuracy: 24.11% -> 24.94% | Improved |
| Optimize the factorization directly for multi-epoch participation | Matrix factorization mechanism · StackOverflow next-word prediction | privacy-accuracy tradeoff: 24.94% at ε=8.84; 24.86% at ε=21.0 | Improved |
| Optimize the factorization directly for multi-epoch participation | Matrix factorization mechanism · StackOverflow next-word prediction | accuracy: 25.25% | Improved |
| Add a multi-epoch-optimized factorization instead of using the un-stamped single-epoch MF mechanism | MF mechanism · CIFAR10 prefix-sum workload, (k,b)=(20,100) | loss: 1.6e6 -> 6.5e5 | Improved |
| Stamp the single-epoch MF mechanism using n=1000 blocks ×2 (ablation) | MF (k=1,n=2000) · CIFAR10 prefix-sum workload, (k,b)=(20,100) | loss: 1.6e6 -> 1.37e6 | Improved |
| Stamp the multi-epoch MF mechanism as n=1000 blocks ×2 rather than use one n=2000 block (ablation) | MF (k=20,n=2000) · CIFAR10 prefix-sum workload, (k,b)=(20,100) | loss: 6.5e5 -> 8.8e5 | Hurt |
| Stamp the multi-epoch MF mechanism as n=500 blocks ×4 (ablation) | MF (k=20,n=2000) · CIFAR10 prefix-sum workload, (k,b)=(20,100) | loss: 6.5e5 -> 1.2e6 | Hurt |
| Stamp the single-epoch MF mechanism as n=100 blocks ×20 (ablation) | MF (k=1,n=2000) · CIFAR10 prefix-sum workload, (k,b)=(20,100) | loss: 1.6e6 -> 2.5e6 | Hurt |
| Stamp the Online Honaker mechanism as n=200 blocks ×10 rather than use n=2000 (ablation) | Online Honaker · CIFAR10 prefix-sum workload, (k,b)=(20,100) | loss: 5.8e6 -> 2.0e6 | Improved |
| Stamp the Optimal Decoder Honaker mechanism as n=400 blocks ×5 rather than use n=2000 (ablation) | Optimal Decoder Honaker · CIFAR10 prefix-sum workload, (k,b)=(20,100) | loss: 2.4e6 -> 1.2e6 | Improved |
| Remove noise contributions from virtual tree steps (ablation) | Online Honaker tree-completion mechanism · CIFAR10 image classification | accuracy | Improved |
| Encode momentum and cooldown in the query matrix rather than apply them as post-processing (ablation) | StackOverflow language-model training · StackOverflow next-word prediction | accuracy | Improved |
Closest papers by meaning
- Scaling up the Banded Matrix Factorization Mechanism for Large Scale Differentially Private ML (ICLR 2025 · similarity 0.89)
- Privacy Amplification for Matrix Mechanisms (ICLR 2024 · similarity 0.89)
- Banded Square Root Matrix Factorization for Differentially Private Model Training (NeurIPS 2024 · similarity 0.87)
- DMM: Distributed Matrix Mechanism for Differentially-Private Federated Learning Based on Constant-Overhead Linear Secret Resharing (ICML 2025 · similarity 0.87)
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · similarity 0.87)
- Differential Privacy in Scalable General Kernel Learning via $K$-means Nystr{\"o}m Random Features (NeurIPS 2024 · 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.
