AFcl: Asynchronous federated continual learning with mobile devices across edges

When mobile devices enter a new environment, federated learning requires them to relearn, which incurs significant energy consumption. Existing continual federated learning methods do not effectively save energy. To improve the effectiveness of federated continual learning and reduce energy consumption, we propose a cross-edge asynchronous federated continual learning algorithm. Firstly, we design a continual learning approach based on the multi-head attention mechanism to retain generalized knowledge from local models and reduce unnecessary relearning. Secondly, we present an asynchronous federated learning scheme to lower the overhead of mobile devices in the edge federated learning system. This scheme introduces a staleness factor to reflect the local model’s staleness. Our aggregation scheme, considering the staleness factor, allows partial device participation in model updates without waiting for all clients, thus cutting down latency caused by waiting. The contributing clients are assigned different contribution weights, and we devise a momentum-based adaptive update method for these weights. Simulation experiments demonstrate that our scheme can significantly reduce communication costs, maintain model accuracy, and boost the overall efficiency of the federated learning system.