Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Learning Over Images: Vertical Decompositions and Pre-Trained Backbones Are Difficult to Beat
Erdong Hu, Yuxin Tang, Anastasios Kyrillidis, Chris Jermaine
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 |
|---|---|---|---|
| use frozen pre-trained backbones (ResNet101+DenseNet121 features) with an MLP instead of full backpropagation through a ResNet18 (ablation) | federated image classification · Stanford Cars i.i.d. | final accuracy within budget: 2.8 -> 66.0 (MOON) / 65.7 (FedProx) | Improved |
| use FedFull (full ResNet18 training) instead of pre-trained backbone under 100 teraFLOP budget (ablation) | FedProx · all six datasets, i.i.d. | success reaching target accuracy: FAIL on every dataset | Hurt |
| choose among FedAvg, FedProx, MOON, FedNova, IST when only final accuracy matters on i.i.d. data | federated learning methods · i.i.d., 6 datasets | final accuracy: Birds 73.5-74.8 | No clear effect |
| use IST (vertical decomposition) instead of FedAvg on i.i.d. data | federated MLP head · CUB-200-2011 i.i.d. | GFLOPs to threshold: 904 -> 11 | Improved |
| use IST instead of FedAvg on i.i.d. data | federated MLP head · CUB-200-2011 i.i.d. | communication to threshold (GB): 9 -> 3 | Improved |
| add proximal term to FedAvg (FedProx) | FedAvg · i.i.d. and skewed | FLOPs to threshold: 904 -> 53 GFLOPs on Birds i.i.d. | Improved |
| add proximal term to IST (ISTProx) | IST · i.i.d. and skewed | final accuracy and cost: 74.6 -> 74.8 on Birds i.i.d. | No clear effect |
| use IST instead of FedAvg under extreme label skew on CIFAR-100 | federated learning · CIFAR-100 skewed | final accuracy: 61.4 -> 49.4 | Hurt |
| use FedAvg instead of IST on Stanford Cars under skew | federated learning · Stanford Cars skewed | final accuracy: 50.8 -> 56.4 | Improved |
| increase active-site participation from 2% to 6% (1000 sites) for IST/FedAdam (ablation) | IST, FedAdam · skewed, 1000 sites | accuracy change: FedAdam +12.8 on Birds; IST +1.2 on Birds | Improved |
| increase participation from 10% to 30% (100 sites) (ablation) | FedAvg, FedProx, MOON, FedNova · skewed, 100 sites | accuracy change: FedAvg -0.5 Birds, +4 CIFAR | Mixed |
| increase number of local iterations and compare methods in FL descent direction similarity to centralized training | IST vs FedAvg/FedProx · Flowers skewed, 20 epochs | cosine similarity to central descent direction: 0.01 -> 0.035 | Improved |
Closest papers by meaning
- Guiding The Last Layer in Federated Learning with Pre-Trained Models (NeurIPS 2023 · similarity 0.89)
- Recurrent Early Exits for Federated Learning with Heterogeneous Clients (ICML 2024 · similarity 0.87)
- FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data (ICLR 2024 · similarity 0.86)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.86)
- Vertical Federated Learning with Missing Features During Training and Inference (ICLR 2025 · similarity 0.85)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.85)
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.
