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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Federated Learning (FL) faces challenges with unpredictable data distributions and catastrophic forgetting in dynamic, decentralized applications.
• Traditional FL struggles with non-IID data, privacy concerns, and adapting models to diverse client data and resource constraints.

Purpose of the Study:

• To develop a novel framework, FedCL, that integrates Graph Convolutional Networks (GCNs) and Vision Transformers (ViTs) with Family-based Continual Learning (FCL).
• To address catastrophic forgetting and enhance model adaptability in FL systems with heterogeneous data distributions and privacy requirements.

Main Methods:

• Implemented a hierarchical, three-tiered model architecture (Parent, Grandparent, Child models) for dynamic adjustment to client data.
• Utilized GCNs for structural data links and ViTs for efficient feature extraction, incorporating Knowledge Distillation Loss (KDL) and surrogate ratios.
• Evaluated on benchmark datasets (FashionMNIST, MedMNIST, DigitMNIST) and real-world applications (MVTeC AD, Vision dataset).

Main Results:

• Achieved high performance metrics: F1-score (97.0%), accuracy (97.6%), precision (97.2%), Learning Performance (LP - 97.3%), and Anomaly Identification Performance (AIP - 96.5%).
• Demonstrated significant reduction in catastrophic forgetting across diverse data distributions.
• Outperformed conventional FL techniques in model adaptability, data privacy, and computational efficiency.

Conclusions:

• The proposed FCL framework offers a robust solution for federated continual learning in complex, real-world applications.
• FedCL provides a viable path for developing adaptable and privacy-preserving FL models in dynamic environments.
• The framework effectively balances model stability and plasticity for improved performance on heterogeneous data.