Energy-Efficient Resource Allocation for Real-Time Video Stream Processing in 6G Networks Using Hierarchical Federated Learning Models

Abstract

This paper addresses the critical challenge of energy-efficient resource allocation for ultra-low latency video stream processing in emerging 6G networks. Traditional centralized Federated Learning (FL) imposes excessive communication overhead due to large model uploads from distributed edge devices to remote cloud servers. We propose a Hierarchical Federated Learning (HFL) framework that introduces an intermediate aggregation layer at Small Cell Base Stations (SBS), reducing long-distance communication and localizing traffic patterns. The framework jointly optimizes power allocation and subcarrier assignment using a Lagrangian dual decomposition approach. Evaluation on UCF101 and LVIS datasets demonstrates that HFL reduces total energy consumption by 32% compared to FedAvg, decreases processing latency by 21%, and achieves 94.2% peak inference accuracy. We introduce two novel evaluation metrics: Energy-Efficiency Index (EEI) and Latency-Accuracy Trade-off Ratio (LATR), providing comprehensive performance characterization. Results validate the scalability of our approach for green AI in 6G ecosystems, making it practical for bandwidth-constrained and energy-limited edge environments.