Family-based continual learning for multi-domain pattern analysis in federated frameworks with GCN and ViT

Continual Learning (CL) and Federated Learning (FL) integration have attracted a lot of interest in dynamic and decentralized areas where client data distributions show a lot of unpredictability, such as industrial imaging, satellite images, medical imaging, and robotic vision. Catastrophic forgetting, non-IID data, and the requirement for effective model updates across clients with restricted data privacy are issues that traditional FL approaches find difficult to handle. These limits impede the development of robust models that can generalize across a range of applications and adjust to changing data and resource limitations. In this research, we offer a novel framework for FL (FedCL) that combines Graph Convolutional Networks (GCNs) and Vision Transformers (ViTs) with Family-based CL (FCL). Our approach reduces catastrophic forgetting and allows the model to be dynamically adjusted to various client data distributions by introducing a hierarchical, three-tiered model architecture made up of the Parent Model (Learning Model), Grandparent Model (Stable Model), and Child Model (Plastic Model). The system utilizes the power of GCN for capturing structural links in patient data and ViT's self-attention mechanism for fast feature extraction, assuring stable performance across varied datasets. Knowledge Distillation Loss (KDL) and surrogate ratios are added to the model to improve learning and facilitate efficient information transfer. We assess our proposed approach on several benchmark datasets, such as FashionMNIST, MedMNIST, and DigitMNIST, and validate it using the MVTeC AD and Vision dataset under several criteria, including F1-score (97.0 %), accuracy (97.6 %), precision (97.2 %), Learning Performance (LP - 97.3 %), and Anomaly Identification Performance (AIP - 96.5 %). Our findings show that the suggested FCL framework considerably lowers catastrophic forgetting across domains with different data properties while outperforming conventional FL techniques in terms of model adaptability, data privacy preservation, and computational efficiency. The suggested approach offers a viable path forward for the development of federated CL in intricate, practical applications.