Generating Synthetic Vehicle Data Using Decentralized Generative Adversarial Networks

Abstract

Ensuring the privacy of personal data is crucial in the era of big data, especially in the transportation industry where sensitive data needs to be processed to develop intelligent vehicle technologies. In particular, collecting and analyzing anomalous data is essential for improving vehicle safety and performance, but accessing such data is often difficult and costly. To address this problem, we propose a novel approach to generating synthetic anomalous data using Generative Adversarial Networks (GANs) and Federated Learning (FL). The proposed learning strategy is decentralized, utilizing data generated by each vehicle to locally train individual discriminators. These discriminators then share only the loss values and weights with a centralized generator. Consequently, the GAN model is trained without exchanging raw data, thereby ensuring the privacy of personal information. Our approach involves a Convolutional Neural Network (CNN)-based architecture for both the generator and discriminator, with the generator residing on the server and a separate discriminator at each vehicle. This design reduces the computational demand on edge devices and enables us to train the GANs using FL. We experiment with different FL strategies and find that the best performer favored the least forgiving discriminator considering data from a pool of vehicles. Our results demonstrate the feasibility of using FL with CNN-based GANs to generate synthetic time-series data for training models in a privacy-preserving manner. This approach has potential applications in the transportation industry, particularly in the context of intelligent vehicles and automated driving systems.