Sensor-Driven mmWave Beam Selection in Heterogeneous Conditions Using Shared-Specific Pruning-Expanding Federated Learning

The swift progress of the Internet of Things (IoT) has resulted in integrating sensors in vehicles and smart mobile devices, enabling the continuous collection of contextual information. With the purpose of efficiently leveraging contextual information from the sensors, this study introduces SSFedSPE, a novel framework designed to tackle the difficulties associated with deep learning models in mm Wave beam selection for vehicles and smart mobile devices. By combining Shared-Specific Feature (SSF) modeling and Sensor-based Adaptive Pruning-Expanding Federated (FedSPE) learning, SSFedSPE effectively handles equipment heterogeneity, resource disparity, statistical heterogeneity, communication efficiency, and training latency. The SSF model's specific and shared blocks enable effective feature extraction and fusion, while FedSPE optimizes model pruning and expansion compatible with client resources and the performance of the model. The experimental results on the FLASH dataset show significant improvements in accuracy and efficiency, particularly in scenarios with varied sensor inputs, underscoring the robustness and adaptability of the proposed approach.