# ProtoFSSL: Federated Semi-Supervised Learning with Prototype-based Consistency Regularization

**Authors:** Woojung Kim, Keondo Park, Kihyuk Sohn, Raphael Shu, Hyung-Sin Kim  
**arXiv:** [2205.13921](https://arxiv.org/abs/2205.13921)  
**Date:** 2022-05-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.13921v1/

## Key points

- ProtoFSSL is a novel approach to Federated Semi-Supervised Learning (FSSL) that addresses limited labeled data on client devices.
- FSSL techniques leverage unlabeled data during model training and ProtoFSSL is based on prototypical networks.
- Clients collaborate in ProtoFSSL by sharing knowledge through lightweight prototypes to maintain consistency and improve overall model accuracy.
- ProtoFSSL uses accurate pseudo-labels generated by each client based on shared prototypes when computing loss on unlabeled data, enhancing the learning process.
- Prototype-based knowledge sharing in ProtoFSSL significantly reduces communication and computation costs compared to traditional FSSL methods relying on weight sharing.
- Experimental results show that ProtoFSSL outperforms recent FSSL methods such as FixMatch, FedRGD, and FedMatch both with and without knowledge sharing, achieving comparable performance to fully supervised FL methods on the SVHN dataset.

## Summary

ProtoFSSL is a novel approach to Federated Semi-Supervised Learning (FSSL) that addresses the challenge of limited labeled data on client devices. With the rise of edge computing power, Federated Learning (FL) has become a popular method for training models without compromising user privacy. However, most existing studies in this field assume that data on client devices are fully labeled, which is often not the case in practice. In response to this limitation, FSSL techniques have been developed to leverage unlabeled data during model training. ProtoFSSL stands out as an innovative FSSL approach based on prototypical networks. In ProtoFSSL, clients collaborate by sharing knowledge through lightweight prototypes, ensuring that local models do not diverge during training. This inter-client knowledge sharing mechanism is crucial for maintaining consistency and improving overall model accuracy. One key feature of ProtoFSSL is its use of accurate pseudo-labels generated by each client based on shared prototypes when computing loss on unlabeled data. These pseudo-labels, combined with labeled data, provide valuable training signals for local prototypes, enhancing the learning process and leading to improved model performance. Compared to traditional FSSL methods that rely on weight sharing, ProtoFSSL's prototype-based knowledge sharing significantly reduces communication and computation costs. This efficiency enables more frequent knowledge exchange between multiple clients, ultimately resulting in higher accuracy levels across various datasets. Experimental results demonstrate that ProtoFSSL outperforms recent FSSL methods such as FixMatch, FedRGD, and FedMatch both with and without knowledge sharing. On the SVHN dataset specifically, ProtoFSSL achieves comparable performance to fully supervised FL methods, highlighting its effectiveness in leveraging unlabeled data for improved model accuracy.