Enhanced Privacy-Preserving Federated Class-Incremental Learning Using Attention-Based Dynamic Aggregation and Continual Learning

Federated Class-Incremental Learning (FCIL) enables distributed clients to collaboratively learn new classes without sharing raw data, thereby ensuring data privacy. However, existing FCIL methods suffer from catastrophic forgetting, inefficient client aggregation, and a challenging privacy–utility tradeoff. This paper proposes an Enhanced Privacy-Preserving Federated Class-Incremental Learning (Enhanced PP-FCIL) framework that integrates Attention-Based Dynamic Aggregation, Elastic Weight Consolidation (EWC), CoreSet Memory Selection, Herding, Rényi Differential Privacy (RDP), Bayesian Differential Privacy (BDP), and a Convolutional Neural Network (CNN) for secure and efficient incremental learning. The proposed framework adaptively aggregates client models, preserves previously acquired knowledge through parameter regularization and representative memory selection, and strengthens privacy protection while maintaining high classification performance. The model is evaluated on the CIFAR-100 dataset under sequential federated class-incremental learning tasks. Experimental evaluation is performed using incremental learning accuracy, catastrophic forgetting analysis, privacy–utility tradeoff, confusion matrix, ablation study, loss landscape visualization, computational complexity analysis, and comparison with state-of-the-art methods. The proposed framework achieves 97.5% classification accuracy, reduces the forgetting rate to 7.8%, improves adaptation to new classes by approximately 4–5%, and lowers false-positive predictions by nearly 65% compared with existing approaches. Furthermore, the framework demonstrates stable convergence, efficient computational complexity, and strong privacy guarantees, making it suitable for privacy-sensitive applications such as healthcare, intelligent IoT systems, financial services, and autonomous intelligent systems.