Dynamic Adaptive Federated Learning for mmWave Sector Selection

TL;DR

The paper tackles the latency and privacy challenges of mmWave sector selection in high-mobility autonomous-vehicle networks by proposing sigla, a dynamic adaptive federated learning framework that combines layer-wise selective aggregation with environment-aware clustering. Sigla uses a layer sensitivity analysis to transmit only informative network layers and employs intra- and inter-cluster aggregation to tailor models to non-IID data while maintaining generalization. Evaluated on a real-world multi-modal dataset, sigla achieves about a 6.76% accuracy gain over a strong baseline (FLASH-and-Prune) and substantially reduces inference time (~84%) and model size (~52%), approaching centralized performance without sharing raw sensor data. The approach significantly enhances beam sector selection reliability and scalability in dynamic vehicular environments, offering a practical path toward faster, privacy-preserving mmWave communications in autonomous networks.

Abstract

Beamforming techniques use massive antenna arrays to formulate narrow Line-of-Sight signal sectors to address the increased signal attenuation in millimeter Wave (mmWave). However, traditional sector selection schemes involve extensive searches for the highest signal-strength sector, introducing extra latency and communication overhead. This paper introduces a dynamic layer-wise and clustering-based federated learning (FL) algorithm for beam sector selection in autonomous vehicle networks called enhanced Dynamic Adaptive FL (eDAFL). The algorithm detects and selects the most important layers of a machine learning model for aggregation in the FL process, significantly reducing network overhead and failure risks. eDAFL also considers intra-cluster and inter-cluster approaches to reduce overfitting and increase the abstraction level. We evaluate eDAFL on a real-world multi-modal dataset, demonstrating improved model accuracy by approximately 6.76% compared to existing methods, while reducing inference time by 84.04% and model size by up to 52.20%.

Figures (5)

• Figure 1: fl-based Sector Selection Scenario
• Figure 2: sigla components and interactions
• Figure 3: Accuracy Results for the tested algorithms
• Figure 4: Final accuracy, Convergence time, and Sector Selection Latency for the evaluated algorithms
• Figure 5: Evaluation Results for the tested algorithms