Effects of Data Hiding on Vision Transformer Classification for Encryption-then-Compression Images

A privacy-preserving image classification method using a vision transformer (ViT) model was proposed so that the encryption processes for a model and test images never affects the classification accuracy. In this paper, we assume that arbitrary information is further embedded to encrypted images and evaluate the effects of the data hiding process on the classification accuracy. In view of compression of encrypted images, an encryption-then-compression (EtC) system is employed for encryption in our method. Simulation results show that the classification accuracy could be preserved without severe degradation on both the classification accuracy and compression performance.