Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedLF: Layer-Wise Fair Federated Learning
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 FedLF, a layer-wise fair federated learning algorithm that computes a common descent direction avoiding model-level and layer-level gradient conflicts, using a cosine-based fair-driven objective and handling absent clients' estimated gradients.
Problem
Federated learning global models are unfair, performing well for some clients and poorly for others; existing proactive fair-FL methods only mitigate model-level gradient conflicts and leave layer-level conflicts and improvement bias unaddressed.
Approach
Formulate FL as a multi-objective problem with an added static fair-driven objective, -cos(1, F(w)), and compute a fair direction layer by layer via a multiple-gradient-descent variant (LMGDA) so that no layer's update conflicts with any client's layer gradient; also estimates absent clients' gradients from recent history.
Key results
- On CIFAR-100 Pat-1, FedLF's average test accuracy is reported as increasing by 47.1% over FedMDFG.
- On CIFAR-10 Pat-1 (Table 2), FedLF reaches 0.752 mean accuracy with fairness indicator 0.086, versus 0.723 (0.158) for FedMDFG, with zero conflicting clients at model and all layers.
- FedLF attains the top mean accuracy in all nine FMNIST/CIFAR-10/CIFAR-100 settings of Table 1 (e.g., 0.892 on FMNIST Dir(0.1)), outperforming FedMDFG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FedLF instead of FedMDFG | federated learning models · CIFAR-100 Pat-1 | mean test accuracy: .278 -> .409 | Helped |
| replace layer-wise fair direction with simple aggregated direction (ablation) | FedLF · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .259 | Hurt |
| combine all layers to compute a model-level fair direction (ablation) | FedLF · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .378 | Hurt |
| split gradients by parameter fragments instead of layers (ablation) | FedLF · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .384 | Hurt |
| remove the fair-driven objective (ablation) | FedLF · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .377 | Hurt |
| replace cosine fair-driven objective with FedMDFG dynamic objective (ablation) | FedLF · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .389 | Hurt |
| do not account for absent-client fairness (ablation) | FedLF · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .229 | Hurt |
| consider only clients online in the last communication round (ablation) | FedLF absent-client handling · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .321 | Hurt |
| consider historical gradients of every absent client regardless of absence duration (ablation) | FedLF absent-client handling · CIFAR-100 Pat-2 | mean test accuracy: .418 -> .391 | Hurt |
Surprising findings
- Replacing the cosine fair-driven objective with FedMDFG's dynamic objective degrades both accuracy and fairness, indicating the static objective is preferable.
Limitations
- Convergence proof assumes all clients are online; absent-client handling relies on estimated historical gradients.
- Layers with fewer parameters than clients may need merging, adding complexity to the layer-wise procedure.
Experimental setup
Datasets: Fashion MNIST CIFAR-10 CIFAR-100
Benchmarks: Fashion MNIST CIFAR-10 CIFAR-100
Metrics: mean test accuracy across clients fairness indicator (arccos cosine between accuracy vector an worst/best 5% client accuracy average number of conflicting clients per layer
Baselines: FedAvg FedProx qFedAvg AFL Ditto FedFV DRFL FedFa FedGini FedCKA FedMGDA+ FedMDFG
Builds on: MGDA FedMDFG FedFV FedMGDA+
Closest papers by meaning
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.90)
- Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity (CVPR 2024 · similarity 0.90)
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.89)
- Multi-Dimensional Fair Federated Learning (AAAI 2024 · similarity 0.89)
- Pursuing Overall Welfare in Federated Learning through Sequential Decision Making (ICML 2024 · similarity 0.89)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 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.