Client-Side Adaptation to Concept Drift in Federated Learning
Finn Saile, Thomas Julius, Dominik Kaaser, Stefan Schulte · 2024
Federated Learning is a paradigm at the intersection of Machine Learning and Distributed Computing. The fundamental idea is that multiple agents collaborate to train a common model without sharing their local data. In this work, we propose an algorithm that passively adapts to concept drift, the evolution of data over time. Our algorithm is deployed on the client side. The main idea is to use a dynamic learning rate at each client which adapts automatically in presence of concept drift. To calculate the learning rate we utilizes the individual training loss. We present an evaluation based on multiple datasets and two types of concept drift, sudden drift and incremental drift. Our empirical analysis shows that the algorithm mitigates the impact of concept drift on the model accuracy and reduces the recovery time required to regain the original level of accuracy.