# Federated Multi-Target Domain Adaptation

**Authors:** Chun-Han Yao, Boqing Gong, Yin Cui, Hang Qi, Yukun Zhu, Ming-Hsuan Yang  
**arXiv:** [2108.07792](https://arxiv.org/abs/2108.07792)  
**Date:** 2021-08-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.07792v1/

## Key points

- Federated learning methods preserve privacy while training machine learning models on distributed user data.
- The authors propose a new task called Federated Multi-Target Domain Adaptation (FMTDA) to adapt a model trained on centralized server data to multiple target domains represented by distributed client data.
- The DualAdapt method is introduced as an effective approach for FMTDA, achieving high accuracy in image classification and semantic segmentation tasks while minimizing communication cost and computational resources on client devices.
- Ablation studies show that introducing local and global classifiers, along with self-training and a GMM weighting mechanism, improves accuracy without significantly increasing computation overhead.
- This research contributes to advancing federated learning methods for privacy-preserving machine learning on distributed user data.

## Summary

Federated learning methods have emerged as a solution to train machine learning models on distributed user data while preserving privacy. To tackle the practical scenario where the distributed client data is unlabeled and only a centralized labeled dataset is available on the server, the authors propose a new task called Federated Multi-Target Domain Adaptation (FMTDA). This task takes into account both server-client and inter-client domain shifts and aims to adapt a model trained on the centralized server data to multiple target domains represented by the distributed client data. The authors analyze the performance of existing domain adaptation methods in this setting and introduce an effective method called DualAdapt. The proposed DualAdapt method achieves high accuracy in image classification and semantic segmentation tasks while minimizing communication cost and computational resources required on client devices. Ablation studies are conducted on a Digit-Five experiment which show that introducing local and global classifiers leads to an accuracy gain with reduced computational cost and communication overhead compared to Fed-MCD. Self-training further improves accuracy without significantly increasing computation overhead. The inclusion of a GMM weighting mechanism allows achieving accuracy close to an oracle model with minimal additional parameters. This work is closely related to centralized multi-target domain adaptation (DA) but does not consider the unique challenges of federated settings with unlabeled client data. In conclusion, this study presents an approach for Federated Multi-Target Domain Adaptation (FMTDA) that addresses these challenges while achieving high accuracy, incurring minimal communication cost, and requiring low computational resources on client devices. This research contributes to advancing federated learning methods for privacy-preserving machine learning on distributed user data.