SIA-GAN: Scrambling Inversion Attack Using Generative Adversarial Network

Abstract

This paper presents a scrambling inversion attack using a generative adversarial network (SIA-GAN). This method aims to evaluate the privacy protection level achieved by image scrambling method. For privacy-preserving machine learning, scrambled images are often used to protect visual information, assuming that searching the scramble parameters is highly difficult for an attacker due to the application of complex image scrambling operations. However, the security of such methods has not been thoroughly investigated. SIA-GAN learns the mapping between pairs of scrambled images and original images, then attempts to invert image scrambling. Therefore, the attacker is assumed to have real images whose domain is the same as that of scrambled images. Experimental results demonstrate that scrambled images cannot be recovered if block shuffling is applied as a scrambling operation. The experimental code of SIA-GAN is publicly available after the paper is accepted.

Privacy-Preserving Image Classification Using Isotropic Network

“…In contrast, recently, novel attack methods for restoring visual information have been proposed that use deep neural networks such as in [9], [19]. Accordingly, we consider the feature reconstruction attack (FR-Attack) [18] that exploits the local properties of an encrypted image to reconstruct visual information from encrypted images.…”

“…Accordingly, we consider the feature reconstruction attack (FR-Attack) [18] that exploits the local properties of an encrypted image to reconstruct visual information from encrypted images. Furthermore, with a synthetic dataset and encrypted images, the adversary may carry out a GAN-based attack (GAN-Attack) [19]. As the distribution of the dataset is known, we also consider that the adversary may prepare exact pairs of plain images and encrypted ones with different multiple keys to learn a transformation model, i.e., the inverse transformation network attack (ITN-Attack) [9].…”

“…We performed the FR-Attack [18], GAN-Attack [19], and ITN-Attack [9] to evaluate the robustness of the EtC images. The results are summarized in Fig.…”

StyleGAN Encoder-Based Attack for Block Scrambled Face Images

“…Similarly, another recent neural network-based attack utilizes a generative adversarial network (GAN) [10]. This GAN-based method aims to reconstruct visual information by using the hinge loss.…”

“…The previous attacks exploit the correlation between pixels [8], [9], or deep neural networks [10], [11] to recover visual information from encrypted images. EtC images are robust against such attacks.…”