DeepClean: a robust deep learning approach for autonomous vehicle camera data privacy

Adeboye, O

Abstract

Autonomous Vehicles (AVs) generate several forms of tracking data, such as geolocation, distance, and camera data. The utility of these data, especially camera data for computer vision projects, has contributed to the advancement of high-performance self-driving applications. However, location inference attacks, which involve extracting knowledge from camera data to track and estimate user locations are potential privacy threats to AV-generated camera data. Recently, a few studies investigated privacy-preserving approaches for AV-generated camera data using powerful generative models such as Variational Auto Encoder (VAE) and Generative Adversarial Network (GAN). However, the related work considered a weak geo-localisation attack model, which leads to weak privacy protection against stronger attack models.

This study develops LIFT (Location InFerence aTtack), a robust geo-localisation technique to exploit subjects’ location privacy in a GAN-based camera dataset. LIFT’s performance is evaluated on a 200k Google Street view as a reference dataset and 500 distorted image datasets as test query data. The result obtained show that the localisation accuracy of LIFT outperforms the benchmark techniques by 20%.

To efficiently address AV camera data privacy preservation, DeepClean is proposed in this thesis. DeepClean combines VAE and private clustering to learn distinct labelled object structures of the image data in clusters. It then generates a more visual
representation of the non-private object clusters, e.g., roads, and distorts the private object areas using a private Gaussian Mixture Model (GMM) to learn distinct cluster structures of the labelled clusters. The synthetic images generated from DeepClean
guarantee privacy and resist robust location inference attacks (such as LIFT) by less than 4% localisation accuracy. The image utility level of the non-private object areas is comparable to the benchmark studies.

Citation

Adeboye, O. DeepClean: a robust deep learning approach for autonomous vehicle camera data privacy. (Thesis). University of Salford

• Thesis Type: Thesis
• Deposit Date: Apr 12, 2023
• Publicly Available Date: Apr 12, 2023
• Award Date: Oct 28, 2022