Traffic Sign Classification for Autonomous Vehicles Using Split and Federated Learning Underlying 5G

• Published in: IEEE open journal of vehicular technology Vol. 4; pp. 877 - 892
• Main Authors: Padaria, Ali Asgar, Mehta, Aryan Alpesh, Jadav, Nilesh Kumar, Tanwar, Sudeep, Garg, Deepak, Singh, Anupam, Pau, Giovanni, Sharma, Gulshan
• Format: Journal Article
• Language: English
• Published: IEEE 2023
• Subjects: 5G mobile communication; Artificial intelligence; artificial intelligence (AI); Computational modeling; Deep learning; Federated learning; federated learning (FL); German Traffic Sign Research Benchmark (GTSRB); Real-time systems; split learning; traffic sign recognition
• ISSN: 2644-1330; 2644-1330
• DOI: 10.1109/OJVT.2023.3326286

By enabling self-driving capabilities, autonomous vehicles (AVs) are revolutionizing transportation. Artificial intelligence (AI) integrating with AVs improves their perception, decision-making, and control systems. Despite AVs' remarkable growth and acceleration, two significant challenges emerge, i.e., distributed data and computational overhead. To address these challenges, we proposed a novel architecture combining split and federated learning (FL) for traffic sign classification in AVs. Our method distributes deep learning (DL) models across multiple AV systems, allowing efficient DL model training and usage. To mitigate computational overhead on vehicular clients, split learning partitions have been applied between the vehicular clients and a server. Additionally, FL advances this approach by simultaneously addressing data decentralization concerns and constructing a strong traffic sign detection model. It achieves this by training the model individually on each client and on that individual client's data and transmitting only crucial weight parameters of the DL model; it also adds a layer of security by avoiding the direct transfer of location-revealing image data. To send model weight parameters in a timely manner, we employed the staggering benefits of a 5G network, which improves the accuracy of applied DL models in each communication round. The proposed architecture offers efficient and real-time model training and utilization for accurate traffic sign identification in AVs by harnessing the capabilities of split learning, FL, and 5G networks. The experimental findings show that our architecture is effective, with a maximum accuracy of 99.54 % on the German Traffic Sign Research Benchmark (GTSRB) dataset with five client architectures, contributing to improving AV technology and developing intelligent transportation systems. Further, we analyze the performance of a 5G network in terms of phase noise and modulation; it is depicted from the experiment that the 5G network is showing noteworthy performance than the 4G network.