Image Classification Using Vision Transformer for EtC Images
Genki Hamano, Shoko Imaizumi, Hitoshi Kiya · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
In this paper, we propose an image classification method for compressible encrypted images called encryption-then-compression(EtC) images, where the accuracy can be preserved without any degradation. The method focuses on the structure of the Vision Transformer(ViT). We first train a ViT model by using plain images, and then encrypt the trained ViT model with a series of encryption keys. On another front, test images are also encrypted with the same series of encryption keys as the model encryption. We employ an EtC method for image encryption, so the encrypted images have high compression performance. When we classify the encrypted images through the encrypted trained model, the classification accuracy is exactly equal to that when plain images are applied to models trained with plain images. In our experiments, the key space is further calculated in order to indicate that the encrypted images are robust against brute-force attacks.