Research map / Federated learning and privacy / Differential privacy
Clients Collaborate: Flexible Differentially Private Federated Learning with Guaranteed Improvement of Utility-Privacy Trade-off
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FedCEO, a differentially private federated learning method that applies truncated tensor SVD (tensor low-rank proximal optimization) on stacked noisy client models at the server, claiming an improved utility-privacy trade-off bound of O(sqrt(d)/N).
Problem
Differential privacy noise added to client updates in federated learning degrades model utility, and the noise accumulates over communication rounds; the utility-privacy trade-off in DP federated learning needs improvement.
Approach
Stack noisy client parameters into a third-order tensor at the server and apply a tensor nuclear norm proximal objective (equivalent to T-tSVD soft-thresholding in the Fourier domain) with an adaptive geometric threshold, so that inter-client semantic correlation smooths out DP noise.
Key results
- On CIFAR-10 with LeNet-5 at sigma_g=2.0, FedCEO (vartheta>1) reaches 45.35% vs 36.70% for CENTAUR and 26.88% for UDP-FedAvg.
- On EMNIST with MLP at sigma_g=1.0, FedCEO (vartheta>1) achieves 78.05% vs 77.26% for CENTAUR.
- Theoretical utility-privacy trade-off bound improved to O(sqrt(d)/N) from O(d/N) (CENTAUR) and O(d^1.5/N) in prior work.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedCEO with ϑ>1 instead of CENTAUR | FedCEO · EMNIST, σ_g=1.0 | testing accuracy: 77.26% -> 78.05% | Helped |
| Use FedCEO with ϑ>1 instead of CENTAUR | FedCEO · EMNIST, σ_g=1.5 | testing accuracy: 71.86% -> 72.44% | Helped |
| Use FedCEO with ϑ>1 instead of CENTAUR | FedCEO · EMNIST, σ_g=2.0 | testing accuracy: 62.12% -> 64.20% | Helped |
| Use FedCEO with ϑ>1 instead of CENTAUR | FedCEO · CIFAR-10, σ_g=1.0 | testing accuracy: 50.14% -> 54.16% | Helped |
| Use FedCEO with ϑ>1 instead of PPSGD | FedCEO · CIFAR-10, σ_g=1.5 | testing accuracy: 47.56% -> 50.00% | Helped |
| Use FedCEO with ϑ>1 instead of CENTAUR | FedCEO · CIFAR-10, σ_g=2.0 | testing accuracy: 36.70% -> 45.35% | Helped |
| Set ϑ=1 instead of using ϑ>1 (ablation) | FedCEO · EMNIST, σ_g=1.0 | testing accuracy: 78.05% -> 77.14% | Hurt |
| Set ϑ=1 instead of using ϑ>1 (ablation) | FedCEO · EMNIST, σ_g=1.5 | testing accuracy: 72.44% -> 71.56% | Hurt |
| Set ϑ=1 instead of using ϑ>1 (ablation) | FedCEO · EMNIST, σ_g=2.0 | testing accuracy: 64.20% -> 63.38% | Hurt |
| Set ϑ=1 instead of using ϑ>1 (ablation) | FedCEO · CIFAR-10, σ_g=1.0 | testing accuracy: 54.16% -> 50.09% | Hurt |
| Set ϑ=1 instead of using ϑ>1 (ablation) | FedCEO · CIFAR-10, σ_g=1.5 | testing accuracy: 50.00% -> 48.89% | Hurt |
| Set ϑ=1 instead of using ϑ>1 (ablation) | FedCEO · CIFAR-10, σ_g=2.0 | testing accuracy: 45.35% -> 37.39% | Hurt |
Limitations
- Experiments limited to small MLP and LeNet-5 models on EMNIST and CIFAR-10.
- Scalability to large models requires restricting T-tSVD to last layers.
- Theoretical analysis relies on strong convexity, smoothness and bounded dissimilarity assumptions; heterogeneous FL left to future work.
Experimental setup
Datasets: EMNIST CIFAR-10
Benchmarks: EMNIST CIFAR-10
Metrics: testing accuracy PSNR
Baselines: UDP-FedAvg PPSGD CENTAUR FedAvg
Builds on: DP-FedAvg tensor nuclear norm tensor SVD DLG Opacus
Closest papers by meaning
- Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation (ICML 2025 · similarity 0.90)
- Sketched Gaussian Mechanism for Private Federated Learning (NeurIPS 2025 · similarity 0.89)
- Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy (NeurIPS 2025 · similarity 0.88)
- Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.88)
- Private and Federated Stochastic Convex Optimization: Efficient Strategies for Centralized Systems (ICML 2024 · similarity 0.88)
- Differentially Private Federated Low Rank Adaptation Beyond Fixed-Matrix (NeurIPS 2025 · 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.