Privacy-Preserving ConvMixer Without Any Accuracy Degradation Using Compressible Encrypted Images

We propose an enhanced privacy-preserving method for image classification using ConvMixer, which is an extremely simple model that is similar in spirit to the Vision Transformer (ViT). Most privacy-preserving methods using encrypted images cause the performance of models to degrade due to the influence of encryption, but a state-of-the-art method was demonstrated to have the same classification accuracy as that of models without any encryption under the use of ViT. However, the method, in which a common secret key is assigned to each patch, is not robust enough against ciphertext-only attacks (COAs) including jigsaw puzzle solver attacks if compressible encrypted images are used. In addition, ConvMixer is less robust than ViT because there is no position embedding. To overcome this issue, we propose a novel block-wise encryption method that allows us to assign an independent key to each patch to enhance robustness against attacks. In experiments, the effectiveness of the method is verified in terms of image classification accuracy and robustness, and it is compared with conventional privacy-preserving methods using image encryption.