Privacy-Preserving Classification of Photon-Limited Images Using Deep Learning

This paper presents a privacy-preserving image classification scheme based on deep learning and photon counting imaging. The vision transformer-based deep learning model can classify the photon-limited color images, which represents the images sparsely by using a few photons. As a result, over 90 % of photon-limited images can be classified when at least 50,000 photons are used. Since some information from the original images is lost in the photon-limited images, it is challenging to recognize the personal information in the photon-limited images. Therefore, the proposed scheme can classify the images while concealing the portion of the information in the images.