A lightweight privacy-preserving scheme using pixel block mixing for facial image classification in deep learning

Abstract

The training of state-of-the-art deep learning models generally requires significant high-quality data, including personal and sensitive data. To ensure privacy of the sensitive data used in training deep learning models, many methods have been designed by the research community. However, it has been observed that many privacy-preserving approaches for image-based deep learning model training incur significant time in the processing of images and/or have low accuracy on trained models when a large number of images is used for classification tasks. Hence, in this paper, we propose a lightweight and efficient approach to preserve image privacy while maintaining the availability of the training set. Specifically, we design the pixel block mixing algorithm for facial image classification privacy preservation in deep learning. Experimental findings show that the models trained by the mixed training set generated by the proposed algorithm maintain their availability. The comparison results of the structural similarity index measure between images in the new training set and the original training set show that our scheme preserves image privacy. Our evaluations also reveal that the data augmentation can be applied to the mixed training set to improve the training effectiveness. We also demonstrate it is computationally challenging for attackers to restore the mixed training set to the original one.