FCLIT: Block-wise Federated Continual Learning for Computing Power Measurement under Dynamic Tasks

In the Computing Power Network (CPN), calculating the computing power is essential before executing distributed tasks. Considering that various tasks have different requirements of the computing power, it is important to classify various tasks accurately in the computing power measurement stage. Compared with calculating the computing power in a centralized mode, applying the Federated Learning (FL) framework to CPN can not only reduce the response delay, but also protect the privacy of client data, as Artificial Intelligence (AI) models are deployed close to the distributed computing power demand sides. However, considering that types of tasks increase dynamically over time, when training AI models with conventional FL algorithms in dynamic data scenarios, AI models are biased to identify current tasks and gradually lose the capability to classify old tasks, leading to the catastrophic forgetting phenomenon. To this end, we propose a novel Federated Continual Learning mechanism for Incremental Tasks (FCLIT) method to identify whole-process data, which splits the AI model into the encoder block and the decoder block. At the local side, FCLIT assesses the AI model to be structurally extensible and stores old encoders in order to characterize old data. At the global side, FCLIT executes a block-wise model aggregation strategy based on the data distribution similarity. We verify the feasibility and effectiveness of FCLIT in multi-task-incremental scenarios under various datasets. Evaluation results demonstrate that, compared with existing works, the proposed FCLIT can effectively improve the model accuracy on classifying old tasks, and keep the capability to classify new tasks in the meanwhile.