Research map / Federated learning and privacy / Differential privacy
Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy
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 f-DP accounting for decentralized DP-SGD with random-walk communication (PN-f-DP, user- and record-level) and for correlated-noise DecoR-style methods (Sec-f-LDP), yielding tighter (epsilon, delta) bounds than RDP-based accounting.
Problem
Privacy accounting for differentially private decentralized federated learning is loose under RDP or (epsilon, delta)-DP when algorithms combine random-walk communication, local iterations, and correlated noise.
Approach
Model a random-walk observer's view as a hitting-time mixture and apply joint concavity of trade-off functions plus Markov chain concentration to get f-DP guarantees; extend f-DP to secret-shared correlated noise (Sec-f-LDP).
Key results
- On MNIST with Expander graph and epsilon=5, test accuracy 0.792 for the f-DP method vs 0.647 for RDP.
- On the complete graph at epsilon=10, the f-DP method reaches 0.891 test accuracy vs 0.867 for RDP.
- Pairwise epsilon values at delta=1e-5 on a hypercube with n=32 are noticeably smaller for PN-f-DP than PN-RDP across nearly all node pairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PN-f-DP instead of RDP (HT) for noise calibration | Decentralized DP-GD · MNIST, Complete graph (ε=10, 8, 5) | test accuracy: RDP (HT): 0.884, 0.876, 0.852; Ours: 0.891, 0.885, 0.869 | Helped |
| Use PN-f-DP instead of RDP (HT) for noise calibration | Decentralized DP-GD · MNIST, Expander graph (ε=10, 8, 5) | test accuracy: RDP (HT): 0.823, 0.797, 0.706; Ours: 0.854, 0.840, 0.792 | Helped |
| Replace RDP-based accounting with PN-f-DP accounting | Random-walk decentralized DP-GD · UCI Housing logistic regression | objective value and test accuracy | Helped |
| Use Sec-f-LDP accounting for correlated-noise training instead of SecLDP or local DP | DecoR-style correlated-noise decentralized FL · a9a logistic regression, ring and 2D torus | privacy-utility trade-off | Helped |
| Use PN-f-DP instead of PN-RDP accounting | Random-walk decentralized DP-SGD · Synthetic graphs, including the hypercube | privacy ε at fixed δ | Helped |
Limitations
- Gaussian lower bounds on trade-off functions are used, so tightness is not guaranteed for general mixtures
- Theorem has no closed form and must be computed numerically
- Experiments limited to synthetic graphs and small classification tasks (logistic regression and a small CNN)
Experimental setup
Datasets: UCI Housing (binarized) MNIST a9a synthetic hypercube and expander graphs
Benchmarks: MNIST UCI Housing (binarized) a9a
Metrics: epsilon at fixed delta test accuracy optimality gap
Baselines: PN-RDP PN-RDP with hitting time SecLDP Local DP DecoR
Builds on: Pairwise Network DP Secret-based Local DP f-DP (Dong et al.) Privacy amplification by iteration DP-SGD
Closest papers by meaning
- Differentially Private Decentralized Learning with Random Walks (ICML 2024 · similarity 0.91)
- The Privacy Power of Correlated Noise in Decentralized Learning (ICML 2024 · similarity 0.90)
- Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation (ICML 2025 · similarity 0.89)
- Private and Federated Stochastic Convex Optimization: Efficient Strategies for Centralized Systems (ICML 2024 · similarity 0.89)
- One-shot Empirical Privacy Estimation for Federated Learning (ICLR 2024 · similarity 0.89)
- Towards the Robustness of Differentially Private Federated Learning (AAAI 2024 · 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.