Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedST: Federated Style Transfer Learning for Non-IID Image Segmentation
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 FedST, a federated image segmentation method that uses a conditional diffusion model to share style features across clients via label-conditioned style transfer, in two variants (separate and join), to mitigate Non-IID data.
Problem
Non-IID image domains across federated clients (e.g., different microscopy, CT/MRI modalities, face races or ages) degrade federated segmentation accuracy and convergence.
Approach
Train a label-conditioned diffusion-based style transfer generator so each client can synthesize cross-domain images from other clients' style while keeping structure (segmentation label) private; FedST-separate keeps per-client generators in a shared style store, FedST-join trains one unified generator with a domain-vector-controlled module via FedAvg.
Key results
- On the race-split face segmentation task, FedST-separate reaches Dice 0.9693 versus 0.9117 for FedDc (best baseline).
- On the material microscopy task, FedST-separate achieves VI 0.2978 versus 0.4022 for FedDc, lower is better.
- On the age-split AAF task with 79 clients, FedST-separate reaches Dice 0.8560 vs 0.8102 for FedAvg (MPA 0.8763 vs 0.8362 for FedDc).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Material microscopic images | VI: 0.4022 -> 0.2978 | Helped |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Material microscopic images | Dice: 0.7400 -> 0.7982 | Helped |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Material microscopic images | MPA: 0.9367 -> 0.9246 | Hurt |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Medical images (CT/MRI) | Dice: 0.5758 -> 0.9233 | Helped |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Medical images (CT/MRI) | MPA: 0.9117 -> 0.9693 | Helped |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Face images, different races | Dice: 0.5936 -> 0.9046 | Helped |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Face images, different races | MPA: 0.1392 -> 0.5962 | Helped |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Face images, different ages | Dice: 0.7979 -> 0.8560 | Helped |
| Add FedST-separate style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Face images, different ages | MPA: 0.8362 -> 0.8763 | Helped |
| Add FedST-join style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Material microscopic images | VI: 0.4022 -> 0.2984 | Helped |
| Add FedST-join style transfer and cross-domain synthetic training | FedAvg-based U-Net segmentation · Medical images (CT/MRI) | Dice: 0.5758 -> 0.9221 | Helped |
| Use FedST-join instead of FedST-separate (aggregate the generator) (ablation) | FedST · Medical images (CT/MRI) | Dice: 0.9233 -> 0.9221 | No clear effect |
Surprising findings
- FedST-join gives comparable accuracy to FedST-separate while having lower communication cost only when the client count exceeds roughly 100 given 50 rounds.
Limitations
- Quality of synthetic cross-domain images is limited; authors note roughly learned translations.
- Shared style generators may pose information leakage risk, which is argued rather than measured.
- Age-task setup uses pre-training on CelebAMask-HQ for all methods, complicating comparison.
Experimental setup
Datasets: Austenite/pure iron microscopy images (in-house) CHAOS (CT/MRI liver) CelebAMask-HQ All-Age-Faces (AAF)
Benchmarks: material microscopy segmentation (in-house) CHAOS CT/MRI liver segmentation CelebAMask-HQ race-split face segmentation AAF age-split face segmentation
Metrics: VI Dice MPA
Baselines: FedAvg FedProx FedDyn FedDc
Builds on: FedAvg SR3 image-to-image diffusion conditional diffusion models FedProx FedDyn
Closest papers by meaning
- Federated Black-Box Adaptation for Semantic Segmentation (NeurIPS 2024 · similarity 0.87)
- FedA3I: Annotation Quality-Aware Aggregation for Federated Medical Image Segmentation against Heterogeneous Annotation Noise (AAAI 2024 · similarity 0.85)
- FedBiP: Heterogeneous One-Shot Federated Learning with Personalized Latent Diffusion Models (CVPR 2025 · similarity 0.85)
- FedRTS: Federated Robust Pruning via Combinatorial Thompson Sampling (NeurIPS 2025 · similarity 0.84)
- FedNE: Surrogate-Assisted Federated Neighbor Embedding for Dimensionality Reduction (NeurIPS 2024 · similarity 0.84)
- CLIP-Guided Federated Learning on Heterogeneity and Long-Tailed Data (AAAI 2024 · similarity 0.84)
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.