Research map / Federated learning and privacy / Personalized and heterogeneous FL

COALA: A Practical and Vision-Centric Federated Learning Platform

ICML 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated prompt and LoRA tuning

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.

Presents COALA, a vision-centric federated learning platform and benchmark suite covering 15 CV tasks, practical data patterns (semi-supervised, unsupervised, continual, feature shift), and model-level scenarios (split learning, multiple models, PEFT), with open-source code.

Problem

Existing federated learning (FL) benchmarks and libraries focus on classification and label heterogeneity, and do not cover diverse computer vision tasks or realistic data and model configurations such as feature shift, semi-supervised learning, continual learning, split models, and multiple models

Approach

A modular FL platform that supports 15 vision tasks, multi-task FL, three levels of practical scenarios (task, data, model), and three customization degrees (configuration, component, workflow), with benchmark experiments across these scenarios mostly using FedAvg.

Key results

• Federated CIFAR-10 with Dir(0.5) over 100 clients reaches 84.08% accuracy versus 90.90% under IID, showing label shift degrades performance.
• Fed-LoRA on ViT-B/16 reaches 97.89% on CIFAR-10 versus 95.36% for Fed-Linear, with 0.0519% trainable parameters.
• On BDD100K object detection with 10 clients, IID achieves 34.45 mAP@0.5 and H-Dir(0.5) achieves 31.36 mAP@0.5.

Measured effects reported in the paper

Surprising findings

• Feature shift combined with data quantity imbalance did not significantly reduce, and sometimes slightly improved, object detection performance relative to IID.
• A larger ResNet backbone (34 to 101) did not significantly improve federated pose estimation performance on MPII.

Limitations

• Most benchmark experiments use only FedAvg and a single run configuration, limiting conclusions on algorithm comparisons.
• Several benchmark results are reported without deep analysis, with the authors noting deeper investigations are needed.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 Digits-Five Office-Caltech-10 DomainNet BDD100K Pascal VOC MPII Taskonomy FEMNIST Landmarks MSMT17

Benchmarks: CIFAR-10 CIFAR-100 Digits-Five Office-Caltech-10 DomainNet BDD100K Pascal VOC MPII Taskonomy

Metrics: accuracy mAP@0.5 mIoU PCK@0.5 precision recall

Baselines: FedAvg FedAvg-FT IFCA FedRep SemiFL FedICON TARGET MocoSFL FedMTL Fed-Linear

Builds on: FedAvg FedCV FederatedScope FedML

Closest papers by meaning

• FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.85)
• F^3OCUS - Federated Finetuning of Vision-Language Foundation Models with Optimal Client Layer Updating Strategy via Multi-objective Meta-Heuristics (CVPR 2025 · similarity 0.83)
• BLEND: Balancing Personalization vs. Generalization in Federated Vision–Language Models (NeurIPS 2026 · similarity 0.83)
• FORLA: Federated Object-centric Representation Learning with Slot Attention (NeurIPS 2025 · similarity 0.83)
• Personalized Federated Learning via Feature Distribution Adaptation (NeurIPS 2024 · similarity 0.83)
• FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning (AAAI 2024 · similarity 0.83)