Resilient and Communication-Efficient Federated Learning for Scalable Heterogeneous IoT Networks on Grassmann Manifolds

Abstract

The rapid growth of mobile devices and IoT technologies has led to an unprecedented surge in edge data generation, shifting the paradigm from centralized data centers to decentralized networks. This trend has fueled edge intelligence, where AI and ML algorithms are deployed directly on user equipment and local data sources. Federated Learning (FL) enables devices to train models locally, sharing only model updates with a central server, preserving privacy while reducing communication costs. However, IoT FL faces challenges due to heterogeneous, high-dimensional, and non-stationary data, compounded by resource constraints and dynamic environments. This thesis addresses these challenges by enhancing communication efficiency, improving intrusion detection, and modeling dynamic IoT data. We introduce a Federated PCA framework that profiles normal and abnormal device behaviors using gradient-based optimization on Grassmann manifolds within a consensus ADMM scheme. This approach enables computationally efficient, real-time anomaly detection while ensuring privacy and low resource usage. Building on this, Federated Koopman Learning (FedKoop) integrates Koopman operator theory with manifold optimization to capture spatio-temporal dependencies in non-stationary IoT data, achieving accurate forecasting with minimal computational and memory demands. FedKooL applies this framework to wireless traffic prediction, leveraging low-rank modeling and advanced optimization to improve predictive accuracy and operational efficiency. Empirical results validate these methods, demonstrating robust, scalable, and privacy-preserving solutions for dynamic, resource-constrained IoT networks.