AI Based Real-Time Privacy-Aware Camera Data Processing in Autonomous Vehicles

Abstract

The three V's, namely volume, velocity and variety of sensor data are ubiquitous for decision-making in autonomous self-driving vehicles. The sensor data contain information about living and non-living entities in the neighbourhood of the moving vehicle. While identifying these objects are essential, details such as human faces, vehicle number plates, building names, etc., are not necessary for decision-making. Thus, we consider the following issues related to data collection, 1) the problem of data privacy, and 2) the problem of misuse of data by an adversary having unauthorized access. This paper proposes a method that first locates private objects (non-essential for decision-making) from frames captured by cameras installed on self-driving cars and then augments it with noise and blurring effects to make them unrecognizable. The performance results show that a combination of blurring and noise can hide private data while retaining information essential for the car to drive. Also, as the proposed approach processes within the limits of the interframe capture time, it is feasible for use in real-time. Moreover, results show that the proposed method can defend the adversarial attacks for the reconstruction of image frames from a given augmented frame.