A Federated Class Incremental Learning Framework for Edge Intelligence

With the advancement of deep learning, artificial intelligence has garnered significant attention in the field of edge computing. However, the characteristics of edge devices present two major challenges for the application of deep neural networks: the conflict between high accuracy and strong data privacy requirements, and the conflict between the high consumption of computing resources and the low performance of edge devices. To address these challenges, this paper proposes a federated class-incremental learning framework for edge intelligence. Based on the varying consumption of computing and storage resources, the scheme divides the training process of the complete federated class-incremental model into three stages, which are executed at different locations. The pre-training of the basic model and the aggregation of global parameters are carried out on the central cloud server with strong computing power, while analytic incremental learning is introduced on the edge server to reduce the computation and storage load during model updates. This scheme offers three advantages: continuous updating capabilities of the model, low resource consumption on edge devices, and a robust data privacy protection mechanism for edge devices. Performance analysis results from experiments demonstrate the reliability of this method.