Federated Learning Under Temporal Drift -- Mitigating Catastrophic Forgetting via Experience Replay

TL;DR

This paper addresses the challenge of temporal concept drift in federated learning by introducing client-side experience replay, which effectively prevents catastrophic forgetting and maintains high accuracy without altering server aggregation.

Contribution

It proposes a simple, effective client-side experience replay method to mitigate catastrophic forgetting under seasonal data drift in federated learning.

Findings

Experience replay restores accuracy from 28% to over 78%.

Buffer size impacts the memory-accuracy trade-off.

Method requires no changes to server aggregation.

Abstract

Federated Learning struggles under temporal concept drift where client data distributions shift over time. We demonstrate that standard FedAvg suffers catastrophic forgetting under seasonal drift on Fashion-MNIST, with accuracy dropping from 74% to 28%. We propose client-side experience replay, where each client maintains a small buffer of past samples mixed with current data during local training. This simple approach requires no changes to server aggregation. Experiments show that a 50-sample-per-class buffer restores performance to 78-82%, effectively preventing forgetting. Our ablation study reveals a clear memory-accuracy trade-off as buffer size increases.