Co-designed pre-capture privacy optics for computer vision

A metasurface-based pre-capture privacy optical system is jointly optimized with a binary classification computer vision (CV) task. The pre-capture privacy module degrades the quality of images before they are captured and saved to the device's memory, making these obscured inputs the only available data for both privacy-breaching attackers and the CV task model. The objective of attaining a trade-off of preserving privacy and CV performance that favors the latter was accomplished as a demonstration of the joint optimization co-design scheme proposed in this work. Specifically, when comparing the maximal performance of the attacker and CV task models attained on their respective tasks when taking unobscured versus obscured inputs, the former model's performance drastically drops from 75.4% accuracy to 17.5% accuracy, while the latter's decreases from 95.7% average precision (AP) to 67.0% AP.