Research map / Federated learning and privacy / Differential privacy
Breaking the Communication-Privacy-Accuracy Tradeoff with $f$-Differential Privacy
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 |
|---|---|---|---|
| Use ternary stochastic compression instead of SQKR | Distributed mean estimation · Distributed mean estimation (N=1000, d=250) | type II error at α = 0.5: 0.068 -> 0.484 | Helped |
| Use ternary stochastic compression rather than SQKR at the same MSE and communication cost | Distributed mean estimation · Distributed mean estimation | ε-DP: {1,2,5} -> {0.05,0.2,3.9} | Helped |
| Apply tight f-DP analysis rather than the prior analysis | Binomial noise · Binomial noise (M=500, p=0.5) | ε in (ε, δ)-DP: 3.18 -> 1.67 | Helped |
| Apply tight f-DP analysis rather than the prior analysis | Binomial noise · Binomial noise (M=500, p=0.5) | δ in (ε, δ)-DP: 0.044 -> 0.039 | Helped |
| Add sparsification to the two-level stochastic sign compressor (ablation) | Ternary stochastic compressor · Scalar compressor | f-DP privacy guarantee: f^ternary(α) > f^sto-sign for α ∈ [(A-c)/(2B), 1-(A+c)/(2B)] | Helped |
| Increase sparsity by increasing B at fixed A (ablation) | Ternary stochastic compressor · Scalar compressor | f-DP privacy guarantee: f^ternary(α) approaches 1-α | Helped |
| Remove sparsification (set A=B) (ablation) | Ternary stochastic compressor · Distributed mean estimation | communication overhead | Hurt |
| Use the ternarize compressor special case (ablation) | Ternary stochastic compressor · Scalar compressor | privacy guarantee: δ = c/B and ε = 0 | Mixed |
| Sparsify the Gaussian mechanism (ablation) | Gaussian mechanism · Distributed mean estimation | MSE | Hurt |
| Set AB = c² + σ² and use ternary compression at large d (ablation) | Ternary stochastic compressor · Distributed mean estimation | privacy guarantee and variance | No clear effect |
| Use ternary compression instead of the Gaussian mechanism for xᵢⱼ ∈ {-c,c} at large d (ablation) | Distributed mean estimation · Distributed mean estimation | privacy guarantee and variance | No clear effect |
| Increase M for binomial noise (ablation) | Binomial noise · Binomial noise | privacy protection | Helped |
Closest papers by meaning
- Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy (NeurIPS 2025 · similarity 0.92)
- Universal Exact Compression of Differentially Private Mechanisms (NeurIPS 2024 · similarity 0.91)
- Improved Communication-Privacy Trade-offs in $L_2$ Mean Estimation under Streaming Differential Privacy (ICML 2024 · similarity 0.89)
- Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation (ICML 2025 · similarity 0.88)
- The Privacy Power of Correlated Noise in Decentralized Learning (ICML 2024 · similarity 0.88)
- A Generalized Shuffle Framework for Privacy Amplification: Strengthening Privacy Guarantees and Enhancing Utility (AAAI 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.