Research map / Federated learning and privacy / Communication-efficient FL
Personalization Improves Privacy-Accuracy Tradeoffs in Federated Learning
Alberto Bietti, Chen-Yu Wei, Miroslav Dudik, John Langford, Steven Wu
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 smaller α (more local personalization) instead of global learning (ablation) | PP-SGD · Synthetic regression | excess risk | Improved |
| Use local-only learning (α=0) instead of joint learning (ablation) | PP-SGD · Synthetic regression with larger privacy noise | excess risk | Improved |
| Use intermediate personalization (α=10) instead of local or global learning (ablation) | PP-SGD · Stackoverflow tag prediction | top-1 accuracy | Improved |
| Use smaller α instead of large α (ablation) | PP-SGD · Stackoverflow tag prediction with privacy noise | top-1 accuracy | Improved |
| Use local-only learning (α=0) instead of joint learning (ablation) | PP-SGD · Federated EMNIST | test accuracy | Hurt |
| Use smaller α to increase personalization (ablation) | PP-SGD · Federated EMNIST | test accuracy at good privacy guarantees | Improved |
| Use smaller α instead of large α (ablation) | PP-SGD with a shared CNN representation · Federated EMNIST CNN | test accuracy at good privacy guarantees | Improved |
| Use large α instead of small α (ablation) | PP-SGD with a shared CNN representation · Federated EMNIST CNN | test accuracy at poor privacy levels | Improved |
| Use local-only learning (α=0), leaving the shared representation untrained (ablation) | PP-SGD with a shared CNN representation · Federated EMNIST CNN | test accuracy | Hurt |
| Optimize α along with privacy noise and learning rate instead of choosing only local, global, or intermediate settings (ablation) | PP-SGD · Stackoverflow and EMNIST | accuracy-privacy tradeoff | Improved |
| Sample one user per round on average instead of sampling all users each round (ablation) | PPSGD with client sampling · Theoretical generalization bound | bias-term convergence rate: 1/Nn instead of 1/n | Improved |
| Optimize average user risk instead of sample-weighted risk with heterogeneous user sample sizes (ablation) | PPSGD with client sampling · Heterogeneous sample sizes | privacy cost | Improved |
Closest papers by meaning
- Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation (ICML 2025 · similarity 0.90)
- Clients Collaborate: Flexible Differentially Private Federated Learning with Guaranteed Improvement of Utility-Privacy Trade-off (ICML 2025 · similarity 0.90)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.90)
- Private Model Personalization Revisited (ICML 2025 · similarity 0.89)
- Private and Federated Stochastic Convex Optimization: Efficient Strategies for Centralized Systems (ICML 2024 · similarity 0.89)
- Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy (NeurIPS 2025 · similarity 0.89)
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.
