RDP-FedGAN: A Rényi-Differential Privacy Empowered Federated Learning GAN in Smart Parking

Abstract

Leveraging smart parking systems can effectively improve urban traffic congestion and parking efficiency. Due to the widespread application of smart sensors, vehicle-mounted devices and Internet technology, the data collected and analyzed by smart parking systems has increased exponentially. Moreover, sensitive information can have the risk of leaking user privacy such as vehicle location, dwell time and real-time environment. In response to these challenges, we present a rényi-differential privacy empowered federated learning generative adversarial networks (RDP-FedGAN) in smart parking. Specifically, we first design a FedGAN model, where GAN can generate synthetic data to train the model in conjunction with the federated learning framework without sharing the original data. Additionally, we propose an adaptive Gaussian noise dynamic adjustment strategy and give rigorous mathematical proof that the proposed mechanism can flexibly adjust the trade-off between privacy and data utility while satisfying rényi differential privacy. Finally, extensive experiments executed on PyTorch under four datasets and various benchmarks with baseline comparison demonstrate the validity of our scheme. Specifically, Our scheme can not only effectively enhance the performance of collaborative models but also strike a flexible and effective balance between privacy preserving and data utility.