mmWave Beam Selection in Heterogeneous Conditions Using Transfer Federated Learning With a Shared-Specific Feature Model

초록

Modern vehicles are equipped with various sensors, including global positioning system (GPS), cameras, and light detection and ranging (LiDAR) to facilitate the continual acquisition of contextual data. This study introduces a new system called transfer federated learning with a shared-specific feature model (TFedSSF), which is designed to utilize one or more available sensor inputs to enhance the efficiency of beam selection in communication networks, thereby eliminating the need for exhaustive searches across all potential beams. Our proposed system also effectively addresses challenges related to practical deployment, including the statistical heterogeneity of local data, disparities in storage resources, and communication overhead while maintaining user privacy. The TFedSSF comprises a shared-specific feature (SSF) model and a transfer-federated (TFed) learning strategy. The SSF model, a modality-agnostic deep learning framework, accommodates any combination of these sensor inputs and extracts their comprehensive features to improve Top-1 accuracy. The TFed strategy trains the SSF model collaboratively across storage-heterogeneous clients and adapts it to each client's local environment. This approach preserves user privacy and minimizes communication overhead during the transmission of the deep learning model between the base station (BS) and client devices. The experimental results on the challenging real-world FLASH dataset demonstrate notable enhancements in accuracy and efficiency, particularly in contexts with diverse sensor inputs, highlighting the robustness and adaptability of the proposed method. Specifically, the Top-1 accuracies for the beam selection of TFedSSF are respectively 29.27%, 67.54%, and 82.60% for single-sensor clients equipped with GPS, Image, and LiDAR. Moreover, they are 67.60%, 82.96%, 86.37%, and 86.37% for clients with the following combinations of sensors: GPS & Image, GPS & Lidar, Image & Lidar, and GPS & Image & Lidar, respectively.

키워드