A plaintext-related image encryption scheme based on key generation using generative adversarial networks

Image encryption based on chaotic systems is a hot research topic in recent decade. However, the existing methods have shortcomings in key generation techniques such as insufficient randomness and limited key space. For that reason, this paper presents a new substitution-scrambling-diffusion color image encryption scheme based on key generation using generative adversarial networks (GANs). In the key generation phase, a key image associated with an original plaintext image is produced by GANs trained on the hyperchaotic Chen system for increasing the key space. The image encryption phase involves three steps: pixel-level substitution, scrambling and diffusion, which all are dependent on the produced key image. In the first step, a pixel-substitution rule is performed on each color channel for increasing the difference between the plaintext and ciphertext images. In scrambling step, the plaintext image is confused globally for reducing the correlation between neighborhood pixels by means of the randomness of the produced key. The diffusion step of the XOR operation employs the diffusion matrix from the generated key for spreading the change of each plaintext bit over as many output ciphertext bits as possible. Experimental simulations and analysis demonstrate that GANs can significantly expand the key space, the proposed image encryption scheme has effective advantages in terms of key randomness, key sensitivity and strong robustness against common security attacks.