Research map / Federated learning and privacy / Personalized and heterogeneous FL
Data Heterogeneity and Forgotten Labels in Split Federated Learning
Joana Tirana, Dimitra Tsigkari, David Solans Noguero, Nicolas Kourtellis
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 |
|---|---|---|---|
| add Hydra multi-head aggregated part-2b heads | SFL · CIFAR-10, 80%-DL, random order, MobileNet | global accuracy: 35±4 -> 66.5±0.1 | Improved |
| add Hydra heads | SFL · CIFAR-10 MobileNet, cyclic | performance gap: decreases by up to 64.5% | Improved |
| add Hydra heads | SFL · CIFAR-10 ResNet101, cyclic | global accuracy: increase of 112.2% | Improved |
| aggregate heads at end of round (Hydra) instead of classic multi-head (ablation) | multi-head SFL · CIFAR-10, 80%-DL, MobileNet | global accuracy: 47±0.3 -> 66.5±0.1 | Improved |
| increase number of heads from 2 to 10 (ablation) | SFL+Hydra · CIFAR-10, 80%-DL, cyclic, MobileNet | global accuracy: 49±1 -> 65±0.1 | Improved |
| use fewer heads (2) than baseline SFL (ablation) | SFL+Hydra · CIFAR-10, 80%-DL, cyclic, MobileNet | global accuracy: 41.1±1 -> 49±1 | Improved |
| use larger heads (part-2b) instead of 2 layers (ablation) | SFL+Hydra · ResNet101 CIFAR-10 | accuracy and PG | Hurt |
| use larger heads (part-2b) instead of 2 layers (ablation) | SFL+Hydra · MobileNet CIFAR-10 | accuracy and PG | No clear effect |
| use cyclic processing order instead of random order at the server | SFL · CIFAR-10 etc. | performance gap, BW | Improved |
| reduce clients with same dominant label processed consecutively (phi 10 -> 1) | SFL · CIFAR-10 MobileNet 80%-DL | global accuracy: 41.1±1 -> 45±0.8 | Improved |
| use deeper cut layer | SFL · MobileNet CIFAR-10 80%-DL | accuracy and PG | Improved |
| add Hydra on Dirichlet and sharding partitions | SFL · CIFAR10 and TinyImageNet | global accuracy: 50 ±0.4 -> 72 ±0.4 (CIFAR10 sharding) | Improved |
Closest papers by meaning
- Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.90)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.90)
- MultiSFL: Towards Accurate Split Federated Learning via Multi-Model Aggregation and Knowledge Replay (AAAI 2025 · similarity 0.88)
- Towards Efficient Replay in Federated Incremental Learning (CVPR 2024 · similarity 0.87)
- Federated Orthogonal Training: Mitigating Global Catastrophic Forgetting in Continual Federated Learning (ICLR 2024 · similarity 0.87)
- Handling Spatial-Temporal Data Heterogeneity for Federated Continual Learning via Tail Anchor (CVPR 2025 · similarity 0.87)
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.
