Research map / Federated learning and privacy / Communication-efficient FL
Pursuing Overall Welfare in Federated Learning through Sequential Decision Making
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.
Unifies existing fairness-aware adaptive aggregation methods for federated learning as online convex optimization (OCO) and proposes AAggFF, with cross-silo (ONS-based) and cross-device (entropic FTRL with doubly robust estimation) variants that achieve sublinear regret and better client-level fairness.
Problem
Single global models in federated learning perform unequally across clients; existing adaptive aggregation schemes that reweight client updates for client-level fairness are suboptimal as sequential decision making and are poorly suited when only a few local signals per round are observed.
Approach
Cast the server's mixing-coefficient update as an online portfolio selection / OCO problem with negative log-growth decision loss, replace static/stateless weights with FTRL-style stateful updates, bound responses via a CDF transform for Lipschitzness, and use a doubly robust estimator under client sampling for the cross-device setting.
Key results
- AAggFF-S on Berka (AUROC) improves worst-client AUROC to 52.08 vs 48.06 for FedAvg, with Gini 10.16 vs 10.87.
- AAggFF-D on Reddit (Acc. 1) yields worst-10% accuracy 4.75 vs 2.50 for FedAvg and average 12.95 vs 10.76.
- Theory: regret O(L sqrt(T log K)) for cross-device and O(L K log T) for cross-silo with full participation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AAggFF-S adaptive aggregation instead of the strongest shown baseline | Cross-silo FL · Berka | Worst AUROC: 49.44 -> 52.08 | Helped |
| Use AAggFF-S adaptive aggregation instead of the strongest shown baseline | Cross-silo FL · Berka | Gini: 10.47 -> 10.16 | Helped |
| Use AAggFF-S adaptive aggregation instead of the strongest shown baseline | Cross-silo FL · MQP | Worst AUROC: 41.71 -> 41.79 | Helped |
| Use AAggFF-S adaptive aggregation instead of the strongest shown baseline | Cross-silo FL · MQP | Gini: 8.56 -> 8.38 | Helped |
| Use AAggFF-S adaptive aggregation instead of the strongest shown baseline | Cross-silo FL · ISIC | Worst Acc. 5: 77.32 -> 85.17 | Helped |
| Use AAggFF-S adaptive aggregation instead of the strongest shown baseline | Cross-silo FL · ISIC | Gini: 3.77 -> 2.52 | Helped |
| Use AAggFF-D adaptive aggregation instead of the strongest shown baseline | Cross-device FL · CelebA | Worst 10% Acc. 1: 55.76 -> 56.71 | Helped |
| Use AAggFF-D adaptive aggregation instead of the strongest shown baseline | Cross-device FL · CelebA | Gini: 7.82 -> 7.54 | Helped |
| Use AAggFF-D adaptive aggregation instead of the strongest shown baseline | Cross-device FL · Reddit | Worst 10% Acc. 1: 2.35 -> 4.75 | Helped |
| Use AAggFF-D adaptive aggregation instead of the strongest shown baseline | Cross-device FL · Reddit | Gini: 23.34 -> 22.59 | Helped |
| Use AAggFF-D adaptive aggregation instead of the strongest shown baseline | Cross-device FL · SpeechCommands | Worst 10% Acc. 5: 5.98 -> 14.54 | Helped |
| Use AAggFF-D adaptive aggregation instead of the strongest shown baseline | Cross-device FL · SpeechCommands | Gini: 23.16 -> 21.42 | Helped |
Surprising findings
- FedMGDA attains the smallest accuracy parity gap on Reddit but at much lower average accuracy (10.58 vs 12.95 for AAggFF-D).
Limitations
- Server uses all client local-loss signals; malicious clients could inflate signals to bias aggregation.
- Convergence analysis in parameter space alongside mixing coefficients is left for future work.
- Regret results are for the mixing-coefficient decision problem, not end-to-end model convergence.
Experimental setup
Datasets: Berka MQP ISIC CelebA Reddit SpeechCommands Heart TinyImageNet
Benchmarks: Berka MQP ISIC CelebA Reddit SpeechCommands Heart TinyImageNet
Metrics: average performance worst 10% best 10% Gini coefficient accuracy parity gap
Baselines: FedAvg AFL q-FedAvg TERM FedMGDA PropFair FedProx FedAdam FedAdagrad FedYogi
Builds on: FedAvg q-FedAvg TERM PropFair Online Newton Step Exponentiated Gradient
Closest papers by meaning
- FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.89)
- Debiasing Federated Learning with Correlated Client Participation (ICLR 2025 · similarity 0.88)
- Exact and Linear Convergence for Federated Learning under Arbitrary Client Participation is Attainable (NeurIPS 2025 · similarity 0.88)
- FedAA: A Reinforcement Learning Perspective on Adaptive Aggregation for Fair and Robust Federated Learning (AAAI 2025 · similarity 0.88)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · similarity 0.88)
- Fair Federated Learning via the Proportional Veto Core (ICML 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.