Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedTSA: A Cluster-based Two-Stage Aggregation Method for Model-heterogeneous Federated Learning
Boyu Fan, Chenrui Wu, Xiang Su, Pan HUI
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace partial-training aggregation (HeteroFL) with FedTSA two-stage DML aggregation | model-heterogeneous FL · CIFAR-100 Dir(0.3) | test accuracy (%): 53.37 -> 59.57 | Improved |
| add DML knowledge exchange via diffusion data instead of HeteroFL when pruning rate is very low (ablation) | FedTSA · CIFAR-10 IID, heterogeneity [0.6,0.8,1.0] to [0.1,0.4,0.7,1.0] | test accuracy (%): FedTSA 89.28 -> 82.37 vs HeteroFL 82.61 -> 68.21 | Improved |
| increase DML global epochs on generated data (ablation) | FedTSA · CIFAR-10 and CIFAR-100 | test accuracy | Hurt |
| add cross-entropy loss to KL loss in DML (ablation) | FedTSA · CIFAR-10 | test accuracy | Improved |
| increase synthetic images beyond 200 (ablation) | FedTSA · CIFAR-10 and CIFAR-100 | test accuracy | No clear effect |
| use fewer than 150 synthetic images (ablation) | FedTSA · CIFAR-10 and CIFAR-100 | test accuracy | Hurt |
| raise DML temperature from 5 to 20 (ablation) | FedTSA · CIFAR-10 IID | test accuracy (%): 89.28 -> 87.30 | Hurt |
| use descriptive prompt "a photo of real bird" instead of single word "bird" (ablation) | diffusion data generation · CIFAR-10 | test accuracy (%): 81.31 -> 89.28 | Improved |
| add dataset-style detail ("in CIFAR-10 style") to the prompt (ablation) | diffusion data generation · CIFAR-10 | test accuracy (%): 89.28 -> 89.26 | No clear effect |
| use diffusion-generated data instead of CGAN-generated data for DML | FedTSA · CIFAR-10 style images | image quality (qualitative) | Improved |
| exclude slow Raspberry Pi devices from FedAvg training | FedAvg · MNIST motivation study | convergence time (s): 2731 -> 1428 | Improved |
Closest papers by meaning
- An Aggregation-Free Federated Learning for Tackling Data Heterogeneity (CVPR 2024 · similarity 0.91)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.91)
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.90)
- Tackling Intertwined Data and Device Heterogeneities in Federated Learning with Unlimited Staleness (AAAI 2025 · similarity 0.89)
- Fake It Till Make It: Federated Learning with Consensus-Oriented Generation (ICLR 2024 · similarity 0.89)
- FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates (ICML 2025 · 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.
