Abstract: To address the challenges of horizontal and vertical catastrophic forgetting in resource-constrained federated continual learning(FCL)environments, a modular-based adaptive weighted federated continual learning(MAWFCL)method was proposed, which effectively addressed the difficulties of model knowledge retention and task adaptability caused by continuously evolving tasks. Personalized models were constructed by combining composable base parameter modules with adaptive control parameters to achieve adaptation to specific tasks. A module similarity-based reuse mechanism was introduced to enhance the efficiency of knowledge reuse. A parameter capacity-aware precision forgetting strategy was incorporated to prune low-contribution modules and maintain a compact model structure. An adaptive aggregation algorithm based on parameter distance was designed to alleviate knowledge conflicts during global model aggregation. Experimental results showed that MAWFCL method outperformed existing methods in terms of accuracy, catastrophic forgetting mitigation, and communication efficiency. On the CIFAR-100 dataset, MAWFCL method improved test accuracy over federated generative replay learning(FedGReL)and prompt-based dual knowledge transfer(Powder)by 10.93 percentage points and 10.17 percentage points, respectively, demonstrating significant advantages in complex tasks.

Key words: federated continual learning, parameter isolation, neural network, transfer learning, adaptive algorithm