Visual Security Evaluation of Learnable Image Encryption Methods against Ciphertext-only Attacks

摘要

Various methods for protecting visual information have been proposed for privacy-preserving deep neural networks (DNNs). In contrast, methods of attack against such protection methods have been simultaneously developed. In this paper, we evaluate state-of-the-art visual-protection methods for privacy-preserving DNNs in terms of visual security against ciphertext-only attacks (COAs). We focus on the brute-force attack, feature reconstruction attack (FR-Attack), inverse transformation attack (ITN-Attack), and GAN-based attack (GAN-Attack), which have been proposed to reconstruct the visual information of plain images from visually protected images. The details of the various attacks are first summarized, and the visual security of the protection methods is then evaluated. Experimental results demonstrate that most of the protection methods, including pixelwise encryption, are not robust enough against GAN-Attack, while a few are robust enough against it.