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

TL;DR

This work shows that standard FedAvg suffers catastrophic forgetting under temporal concept drift in federated learning, demonstrated with a seasonal Fashion-MNIST drift. The authors propose a simple, server-free mitigation: client-side experience replay buffers that mix past samples with current data during local updates. A buffer-size ablation reveals a memory-accuracy trade-off, with a 50-sample-per-class buffer restoring accuracy to about 78–82% while preserving the FedAvg aggregation. The approach is deployment-friendly, requiring only client-side changes and offering practical protection against forgetting in drifting, non-IID environments.

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.

Figures (5)

• Figure 1: Global Accuracy Recovery Curve
• Figure 2: Standard FedAvg: Catastrophic Forgetting Under drift
• Figure 3: FedAvg + Experience Replay: Forgetting Mitigated
• Figure 4: Per class Robustness Profile
• Figure 5: Ablation study: Impact of replay buffer size