Adaptive Federated Continual Learning for IoT-Cloud-Edge Systems

This work presents FedAdapt, an adaptive federated continual learning framework designed for IoT-Cloud-Edge systems with limited and heterogeneous resources. FedAdapt enables edge devices to collaboratively train a shared model without sharing raw data, preserving privacy while supporting continual learning over evolving tasks. The framework dynamically offloads selected neural network layers to more powerful servers and uses reinforcement learning–based scheduling to handle device and network heterogeneity efficiently. By integrating gradients from past and current tasks, FedAdapt mitigates catastrophic forgetting and negative knowledge transfer. Experimental results on real edge devices demonstrate significant reductions in training time and communication overhead while maintaining high model accuracy, making FedAdapt a practical and scalable solution for real-world edge intelligence.