Research map / Federated learning and privacy / Differential privacy
The Fundamental Price of Secure Aggregation in Differentially Private Federated Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Apply sparse random projection with r=5 and b=12 instead of quantization at r=1 and b=10 | Stack Overflow NWP LSTM · Stack Overflow NWP, 1000 clients, z=0.3 | total communication per parameter (bits): 10 -> 2.4 | Helped |
| Apply sparse random projection with r=5 and b=12 instead of quantization at r=1 and b=10 | Stack Overflow NWP LSTM · Stack Overflow NWP, 1000 clients, z=0.3 | final test performance (%): 21.90\pm 0.09 -> 21.68\pm0.02 | No clear effect |
| Increase compression from r=5 to r=10 while using b=12 | Stack Overflow NWP LSTM · Stack Overflow NWP, 1000 clients, z=0.5 | total communication per parameter (bits): 2.4 -> 1.2 | Helped |
| Apply sparse random projection with r=10 and b=12 instead of quantization at r=1 and b=10 | Stack Overflow NWP LSTM · Stack Overflow NWP, 1000 clients, z=0.5 | final test performance (%): 21.21 \pm 0.18 -> 20.74 \pm 0.05 | No clear effect |
| Combine dimensionality reduction with optimized quantization | F-EMNIST CNN · F-EMNIST, z=0.5 | communication per parameter (bits): 0.24 bits per parameter; quantization alone: 10 bits per parameter | Helped |
| Increase noise multiplier and use higher compression | Sparse random projection FL · Stack Overflow NWP | attainable compression rate: r=4x at z=0; r=5x at z=0.3; r=10x at z\geq0.5 | Helped |
| Train a smaller model instead of the original larger model at comparable latent update size (ablation) | F-EMNIST CNN · F-EMNIST, across privacy budgets | test performance: >5pp | Hurt |
| Increase quantization bitwidth from b=18 to b=22 (ablation) | Sparse random projection FL · Stack Overflow NWP | test performance: b=18\to22 | Helped |
| Sketch model layers separately instead of sketching the concatenated gradient (ablation) | Stack Overflow NWP LSTM · SONWP, z=0.05 and z=0 | final model test performance | No clear effect |
| Threshold low values in the random projection sketch (ablation) | Sparse random projection FL · Stack Overflow NWP LSTM | final model test performance | No clear effect |
Closest papers by meaning
- Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy (NeurIPS 2025 · similarity 0.90)
- Noise-Aware Algorithm for Heterogeneous Differentially Private Federated Learning (ICML 2024 · similarity 0.90)
- PREAMBLE: Private and Efficient Aggregation via Block Sparse Vectors (NeurIPS 2025 · similarity 0.90)
- DMM: Distributed Matrix Mechanism for Differentially-Private Federated Learning Based on Constant-Overhead Linear Secret Resharing (ICML 2025 · similarity 0.89)
- Towards Trustworthy Federated Learning with Untrusted Participants (ICML 2025 · similarity 0.89)
- Federated Online Prediction from Experts with Differential Privacy: Separations and Regret Speed-ups (NeurIPS 2024 · 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.