SSFL: Tackling Label Deficiency in Federated Learning via Personalized Self-Supervision

Chaoyang He, Zhengyu Yang, Erum Mushtaq, Sunwoo Lee, Mahdi Soltanolkotabi, Salman Avestimehr

AAAI Workshop, 2022

arXiv

This work proposes SSFL, a unified self-supervised and personalized Federated Learning framework that extends FedAvg to work with self-supervised methods like SimSiam and introduces Per-SSFL, a personalization algorithm that regularizes the distance between local and global representations, showing that self-supervised FL closes most of the accuracy gap with supervised FL while representation-regularized personalization outperforms other variants.

BibTeX

@article{he2021ssfl,
  title={Ssfl: Tackling label deficiency in federated learning via personalized self-supervision},
  author={He, Chaoyang and Yang, Zhengyu and Mushtaq, Erum and Lee, Sunwoo and Soltanolkotabi, Mahdi and Avestimehr, Salman},
  journal={arXiv preprint arXiv:2110.02470},
  year={2021}
}