Title: FlashbackCL: Mitigating Temporal Forgetting in Federated Learning

Abstract

As Federated Learning (FL) involving foundation and edge models expands into environments where client data distributions evolve over time, current mitigation strategies face a significant limitation: they presuppose that each client’s data distribution remains static. Flashback, currently the most effective recent approach against cross-client (spatial) forgetting, relies on monotonically accumulating per-class label counts as a proxy for knowledge. However, this proxy becomes misaligned when temporal distribution shifts occur, effectively anchoring the global model to obsolete class balances. To address this, we formally define temporal forgetting in FL using a per-phase metric that isolates it from protocol-level noise. We introduce Flashback Continual Learning (FlashbackCL), a seamless extension of Flashback that incorporates three key innovations: (i) label counts that decay over time; (ii) a device-aware replay buffer utilizing Class-Balanced Reservoir Sampling (CBRS); and (iii) server-side active coreset curation applied to the public distillation set.

Experimental results demonstrate that FlashbackCL yields a relative improvement of 6.9% to 10.0% compared to the standard Flashback method on CIFAR-10, using 50 clients and three controlled modes of temporal shift. Furthermore, it reduces temporal forgetting by as much as 68%. An ablation study involving five variants highlights that CBRS replay is the pivotal component of this improvement. Additionally, FlashbackCL enhances Flashback’s performance by 3.5 points on stationary CIFAR-100, indicating that class-balanced replay not only mitigates temporal shifts but also regularizes spatial heterogeneity.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC