FedPLC: Federated Learning with Dynamic Cluster Adaptation for Concept Drift on Non-IID Data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Decentralized federated learning (FL) in Internet of Things (IoT) faces challenges from non-independent and identically distributed (Non-IID) data.
• Concept drift further degrades model convergence and generalization by introducing temporal uncertainty.
• Existing FL methods struggle to differentiate data heterogeneity from dynamic shifts, limiting adaptability.

Purpose of the Study:

• To propose FedPLC, a novel FL framework designed to address Non-IID data and concept drift in IoT environments.
• To enhance model robustness and fine-grained adaptation for large-scale IoT/edge client populations.

Main Methods:

• Introduces Prototype-Anchored Representation Learning (PARL) to stabilize representations against noise and shifts by aligning sample embeddings with class prototypes.
• Implements Label-wise Dynamic Community Adaptation (LDCA) for fine-grained, label-level reorganization of classifier heads, enabling rapid personalization and drift-aware evolution.
• Explicitly disentangles static Non-IID heterogeneity from temporal concept drift.

Main Results:

• FedPLC demonstrates superior performance compared to state-of-the-art methods in both abrupt and incremental concept drift scenarios.
• Experimental results on Fashion-MNIST, CIFAR-10, and SVHN datasets validate the framework's effectiveness.
• Achieves robust and fine-grained adaptation for large-scale IoT/edge client populations.

Conclusions:

• FedPLC effectively tackles the dual challenges of Non-IID data and concept drift in federated learning for IoT.
• The proposed PARL and LDCA mechanisms offer a robust solution for dynamic and heterogeneous edge environments.
• Enables improved model generalization and convergence in real-world, evolving IoT deployments.