Decrypting the Chao-based Image Encryption via A Deep Autoencoding Approach

In this work, as a proof of concept, we develop a simple yet effective deep autoencoding approach to attack the chaos-based image encryption algorithm. Specifically, we first project chaos-based encrypted images into the low-dimensional feature space. In the manifold, essential information of plain images can be largely preserved. A deconvolutional generator is then utilized to regenerate decrypted images perceptually similar to plain images in the high-dimensional image space. For the first time, we show the feasibility of attacking chaos-based encryption methods in a key-independent manner, which is fundamentally different from traditional image encryption attacks. Given chaos-based encrypted images, a well-trained decryption model can automatically reconstruct plain images with high visual fidelity. In both static-key and dynamic-key experiments, we successfully attack several chaos-based algorithms on the MNIST dataset and show that decrypted images are perceptually recognizable to both human eyes and intelligent models. Importantly, this study demonstrates the potential of employing deep learning approaches to crack image encryption algorithms.