An image encryption framework based on chaotic sequence combined with deep learning

Abstract Aiming to address the security issues in the storage and transmission of digital images, this paper proposes an image encryption framework combined with deep learning. Firstly, by integrating a convolutional neural network (CNNs) with the encoder of a transformer, a deep learning model named Chaos-Encoder Model (CEM) has been formed. Subsequently, the publicly standard test images used for encryption test are used as the training dataset. The training targets of this model are Lorenz hyperchaotic sequences, which contain the inherent features of the training images. After training, the plain image is fed into the model. Then the proposed model generates a new chaotic sequence based on the image’s unique characteristics. This resulting sequence will also be the key stream of the subsequent encryption algorithm. Next we conducted 0–1 tests, Lyapunov tests and randomness tests on the new chaotic sequence. The results show that it not only retains chaotic performance, but also behaves more randomly and unpredictably compared with the corresponding chaotic sequence. Finally, we apply Lorenz hyperchaotic sequences and the generated sequences of CEM to conduct an advanced multidirectional interleaved diffusion and permutation algorithm based on backtracking. Experiments on various performance indicators show that the encryption algorithm proposed in this paper can effectively resist brute-force attacks.