Federated Learning Under Temporal Drift -- Mitigating Catastrophic Forgetting via Experience Replay
This addresses forgetting issues for federated learning systems with drifting data, but it is incremental as it builds on existing experience replay methods.
The paper tackles catastrophic forgetting in Federated Learning under temporal concept drift, showing that standard FedAvg drops accuracy from 74% to 28% on Fashion-MNIST, and proposes client-side experience replay with a 50-sample-per-class buffer to restore performance to 78-82%.
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.