Privacy-Optimized Camera

Abstract: With the growing reliance on visual data, the tradeoff between privacy and utility has become a major concern. The need for ensuring personal privacy without sacrificing the value of visual data for machine learning applications has motivated the development of privacy-preserving methods. In this work, we evaluated the general framework of using optical and digital elements to protect privacy in computer vision applications. In particular, a trainable optical filter and a digital encoder were jointly optimized for privacy and utility objectives. We explore two ways of parameterizing the optical filter and compare the results with basic privacy-preserving techniques. The method was evaluated on two tasks: image classification and semantic segmentation. The results demonstrate that it can preserve task-relevant information without compromising privacy.

Website: here

Paper: here