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

Published: 28 Jan 2022, Last Modified: 13 Feb 2023ICLR 2022 SubmittedReaders: Everyone
Keywords: federated learning
Abstract: Federated Learning (FL) is transforming the ML training ecosystem from a centralized over-the-cloud setting to distributed training over edge devices in order to strengthen data privacy, reduce data migration costs, and break regulatory restrictions. An essential, but rarely studied, challenge in FL is label deficiency at the edge. This problem is even more pronounced in FL, compared to centralized training, due to the fact that FL users are often reluctant to label their private data and edge devices do not provide an ideal interface to assist with annotation. Addressing label deficiency is also further complicated in FL, due to the heterogeneous nature of the data at edge devices and the need for developing personalized models for each user. We propose a self-supervised and personalized federated learning framework, named SSFL, and a series of algorithms under this framework which work towards addressing these challenges. First, under the SSFL framework, we analyze the compatibility of various centralized self-supervised learning methods in FL setting and demonstrate that SimSiam networks performs the best with the standard FedAvg algorithm. Moreover, to address the data heterogeneity at the edge devices in this framework, we have innovated a series of algorithms that broaden existing supervised personalization algorithms into the setting of self-supervised learning including perFedAvg, Ditto, and local fine-tuning, among others. We further propose a novel personalized federated self-supervised learning algorithm, Per-SSFL, which balances personalization and consensus by carefully regulating the distance between the local and global representations of data. To provide a comprehensive comparative analysis of all proposed algorithms, we also develop a distributed training system and related evaluation protocol for SSFL. Using this training system, we conduct experiments on a synthetic non-I.I.D. dataset based on CIFAR-10, and an intrinsically non-I.I.D. dataset GLD-23K. Our findings show that the gap of evaluation accuracy between supervised learning and unsupervised learning in FL is both small and reasonable. The performance comparison indicates that representation regularization-based personalization method is able to outperform other variants. Ablation studies on SSFL are also conducted to understand the role of batch size, non-I.I.D.ness, and the evaluation protocol.
One-sentence Summary: Tackling Label Deficiency in Federated Learning via Personalized Self-Supervision
Supplementary Material: zip
9 Replies

Loading