Holo-FAFL: Cross-Modal Feature Augmentation and Latency-Aware Asynchronous Federated Learning for Holographic IoT Systems

Le Sun, Yulin Wang, Zhiguo Qu, Jinliang Liu, Jing Tian, Yimin Yu, Ghulam Muhammad, Mahmoud Ragab

Abstract: Federated Learning (FL) is a revolutionary distributed machine learning paradigm that addresses the challenge of data silos while preserving privacy. However, when deployed in heterogeneous environments, especially in Holographic IoT systems, FL faces significant challenges such as data heterogeneity, device heterogeneity, and cross-modal distribution shifts. To overcome these limitations, this article proposes Holographic Feature Augmentation-based Federated Learning (Holo-FAFL), an adaptive asynchronous federated learning framework. The framework integrates cross-modal statistical alignment with latency-aware optimization. The Cross-modal Feature Augmentation (CFA) module in Holo-FAFL eliminates statistical discrepancies between physical sensor data and virtual 3D features through channel-wise mean-variance alignment, t-distribution modulation, and adversarial sample generation. The Adaptive Data Update (ADU) module suppresses stale updates from high-latency devices and dynamically adjusts buffer capacities based on rendering workloads. Additionally, the framework adopts Virtual Risk Minimization (VRM) through virtual feature augmentation to enhance generalization by 12.7% accuracy under extreme non-independent and identically distributed (Non-IID) conditions.

External IDs:dblp:journals/tce/SunWQLTYMR25