Beyond Fixed Rounds: Data-Free Early Stopping for Practical Federated Learning

TL;DR

This paper introduces a data-free early stopping method for federated learning that monitors task vector growth to determine optimal training duration without validation data, reducing costs and privacy risks.

Contribution

It presents the first data-free early stopping framework for federated learning, improving efficiency and privacy by eliminating the need for validation data.

Findings

Achieves comparable performance to validation-based early stopping.

Requires fewer additional rounds to reach higher accuracy.

Demonstrates effectiveness on skin lesion and blood cell classification tasks.

Abstract

Federated Learning (FL) facilitates decentralized collaborative learning without transmitting raw data. However, reliance on fixed global rounds or validation data for hyperparameter tuning hinders practical deployment by incurring high computational costs and privacy risks. To address this, we propose a data-free early stopping framework that determines the optimal stopping point by monitoring the task vector's growth rate using solely server-side parameters. The numerical results on skin lesion/blood cell classification demonstrate that our approach is comparable to validation-based early stopping across various state-of-the-art FL methods. In particular, the proposed framework requires an average of 45/12 (skin lesion/blood cell) additional rounds to achieve over 12.3%/8.9% higher performance than early stopping based on validation data. To the best of our knowledge, this is the first work to propose an data-free early stopping framework for FL methods.