Privacy-Preserving Machine Learning Models for Autonomous Vehicle Data Analysis

Abstract

The ability to motivate secure communication protocols that maintain both data privacy and training model owner safety has drawn a lot of attention to PPML design [1]. However, this area is relatively new, and most of the work is centered around supervised learning, where the server can easily identify each data point. In cooperative learning, less attention is given to privacy protection, but the cooperation model can potentially exploit the privacy requirements of the various players involved. This fact becomes more important when we combine the standard cooperative learning model with a setup with several possible sources of heterogeneity between the data of the various devices. This has led to research directions that are often considered in isolation from each other. However, due to the fast-paced nature of these technologies, the reality is that these aspects are likely to be influential in combined form during the next-generation technologies.

Autonomous vehicles (AVs) have tremendous potential to change future commuting and transportation [2]. Cars possess the ability to utilize sensor data to anticipate and mitigate traffic accidents, driver errors and poor road conditions. But for these tasks to be carried out effectively, the AVs must be equipped with robust learning algorithms to interpret sensor data. However, these algorithms are also put at risk because of privacy concerns. People are more wary of sharing personal data now than they were a few years ago. Research has shown that people are more accepting of partial data sharing concepts than full data sharing.