Research map / Federated learning and privacy / Personalized and heterogeneous FL
Improving Federated Domain Generalization Through Dynamical Weights Calculated from Data Influences on Global Model Update
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 DI, which computes per-client aggregation weights from influence-function estimates of each source domain's data on the global model update, then adjusts them with a second term to improve federated domain generalization.
Problem
Federated domain generalization needs good aggregation weights for local models, but existing weighting schemes (FedAvg data volume, GA fairness-based) do not use how individual data points influence the global model update.
Approach
Dynamically weight local models per round using influence-function approximations of source-domain data's effect on the change in generalization (DIC), then regularize those weights with a loss-based term and normalization (DIA) so they are not dominated by a small data subset.
Key results
- DI improves average accuracy on PACS, Office-Home, and Office-31 by 1.79%, 1.61%, and 2.39% per the abstract.
- On Table 3, DI's accuracy gains exceed GA's across all datasets; the greatest DI gain is 3.31% vs GA's 1.81%.
- Ablation on PACS with FedAvg: DIC alone raises accuracy from 77.76% to 79.24%, DIA alone to 78.24%, and both together to 80.16%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DI to AM | AM · PACS | accuracy: +1.58 | Helped |
| Add DI to AM | AM · Office-31 | accuracy: +3.31 | Helped |
| Add DI to AM | AM · OfficeHome | accuracy: +1.93 | Helped |
| Add DI to FedAvg | FedAvg · PACS | accuracy: +2.25 | Helped |
| Add DI to FedAvg | FedAvg · Office-31 | accuracy: +2.63 | Helped |
| Add DI to FedAvg | FedAvg · OfficeHome | accuracy: +1.63 | Helped |
| Add DI to RSC | RSC · PACS | accuracy: +1.74 | Helped |
| Add DI to RSC | RSC · Office-31 | accuracy: +2.58 | Helped |
| Add DI to RSC | RSC · OfficeHome | accuracy: +1.50 | Helped |
| Add DI to Scaffold | Scaffold · PACS | accuracy: +1.57 | Helped |
| Add DI to Scaffold | Scaffold · Office-31 | accuracy: +1.05 | Helped |
| Add DI to Scaffold | Scaffold · OfficeHome | accuracy: +1.38 | Helped |
Limitations
- Influence computation relies on a first-order approximation and a randomly sampled core dataset, adding per-client cost.
- Gains are reported only on image-classification benchmarks with ResNet18 and a small set of FL baselines.
Experimental setup
Datasets: PACS Office-Home Office-31
Benchmarks: PACS Office-Home Office-31
Metrics: accuracy accuracy improvement variance of per-domain accuracy
Baselines: AM FedAvg RSC Scaffold GA
Builds on: FedAvg GA (federated generalization adjustment) influence functions (Koh and Liang 2017)
Closest papers by meaning
- Federated Domain Generalization with Data-free On-server Matching Gradient (ICLR 2025 · similarity 0.88)
- Benchmarking Algorithms for Federated Domain Generalization (ICLR 2024 · similarity 0.87)
- Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization (CVPR 2024 · similarity 0.87)
- Improving Generalization in Federated Learning with Model-Data Mutual Information Regularization: A Posterior Inference Approach (NeurIPS 2024 · similarity 0.87)
- Improving Generalization in Federated Learning with Highly Heterogeneous Data via Momentum-Based Stochastic Controlled Weight Averaging (ICML 2025 · similarity 0.87)
- TTA-FedDG: Leveraging Test-Time Adaptation to Address Federated Domain Generalization (AAAI 2025 · similarity 0.86)
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.