Privacy Preserving Cameras Using Diffractive Optical Elements and Optical Neural Networks

Abstract

Due to the growing use of cameras in our daily lives, there is an increasing concern about individual privacy. While cameras are incredibly useful for gathering information, they also have the potential to invade our personal lives. This has led to a growing interest in creating privacy-preserving cameras that only capture task-related features which cannot be used to learn identifying information about people. We design an end-to-end system that has the twofold goal of preserving privacy while maintaining good performance on a downstream task. Our system contains an optical component which extracts features while protecting privacy, as well as a neural network component which performs the downstream task. We jointly optimize the optical component and the downstream neural network to achieve the best downstream performance while obscuring the scene images as much as possible. We first experiment with the use of diffractive optical elements (DOE) as our privacy-preserving optical component. We evaluate this approach on several different downstream tasks, including image classification, human pose estimation, and action detection. However, we find significant degradation in the performance on all downstream tasks, and we also find that the amount of privacy offered by the DOE is insufficient. After reassessing and deciding that using a DOE was not a viable solution for privacy-preserving cameras, we instead transition to using an optical neural network as our privacy-preserving optical component. Although we did not have time to finish designing and evaluating this method, we believe that it is a promising way to achieve good downstream performance while preserving privacy.