Meta Representation-Based Personalized Federated Continual Learning in Edge Computing Systems

Federated learning (FL) enables multiple clients to collaboratively train a shared global model while keeping their data local, fostering privacy-preserving innovation in sectors such as healthcare, finance, and telecommunications. Federated continual learning (FCL) extends FL by enabling clients to incrementally learn and adapt the global model over time, addressing dynamic data distributions and evolving tasks while maintaining data privacy. However, as differences in data distribution and class diversity across clients’ data streams widen, the local personalization and continuous adaptation of the global model to both previously learned tasks and current tasks in clients’ local data streams present a significant challenge. To address this challenge, we propose a meta representation based class incremental learning (Meta-RBCIL) algorithm for peronalized FCL in edge computing systems. The proposed approach meta-trains a representation network using FL, equipping it with strong generalization capabilities across diverse tasks. This network is then leveraged to adapt to each client’s current task data and exemplar memory using balanced gradients. By doing so, the method mitigates catastrophic forgetting in clients’ continual learning models and alleviates bias caused by the imbalance between old samples in exemplar memories and new class samples in current tasks. The experimental results show that our approach surpasses key baseline algorithms in terms of average task accuracy per communication round, across varying levels of task similarity in clients’ data streams. Furthermore, the personalized continual learning models produced by the Meta-RBCIL algorithm achieve the lowest average task forgetting ratio and the highest average task generalization capability, making Meta-RBCIL highly suitable for deployment in edge computing systems.