Federated Continual Learning with Bounded Forgetting via Diffusion-Based Generative Replay in Edge Computing

Federated continual learning (FCL) aims to train a shared model incrementally on evolving tasks while respecting data privacy across distributed edge devices. In edge settings, historical task data for replay is often unavailable; moreover, as the model is continuously updated, features of newly introduced classes encroach on the feature space of older classes, distorting decision boundaries and exacerbating catastrophic forgetting. We propose DGR-FCL, an FCL framework that bounds forgetting via server-side, diffusion-based generative replay. The generator is trained with margin constraints and batch-normalization alignment to faithfully recover previously learned distributions. On each client, we introduce supervised contrastive learning with synthetic negatives to separate old and new classes, and an importance-weighted feature-distillation strategy to constrain representation drift along feature directions critical for old tasks. Theoretically, an integral probability metric (IPM) analysis shows that margin preservation on replay data ensures bounded forgetting. Empirically, DGR-FCL achieves consistent 4–5% accuracy gains over state-of-the-art baselines on both independent and identically distributed (IID) and non-IID benchmarks, offering a robust solution to the stability–plasticity trade-off in decentralized, continuously evolving environments.