Key Generation Method Based on Generative Adversarial Network and Its Application in Low-light-level Image Encryption

Generative adversarial network (GAN) is a deep learning model, which is one of the most promising methods for unsupervised learning in complex distributed environment in recent years. The random numbers generated by GAN and chaotic systems have many similar advantages, such as randomness and sensitivity. But at the same time, the random numbers generated by GAN have the features that the random numbers generated by chaotic systems do not have, such as non-reproducibility.The random number generated by GAN training shows its rapidity and higher security in the encryption of low-light-level images with low signal-to-noise ratio and few gray levels. GAN is introduced into the random key generation. The pseudo-random number generated by quantum cellular neural network (QCNN) system is used as the training set of GAN, and a completely random key pool is obtained through GAN to learn and train the random key generated by hyperchaotic system. For the application of the proposed key generation method in low-light-level image encryption, a new low-light-level image encryption algorithm is proposed, which uses a 2D-pointer related to the plaintext to randomly select two phase masks in the key pool to achieve the security of low-light-level images. The results show that the encryption key generated by the learning key generation scheme can pass all the random tests of The National Institute of Standards and Technology, and the scheme can resist differential attacks by cryptanalysts, known/chosen plaintext attacks, and various statistical analyses. The performance comparison with other similar algorithms also shows the superiority of the proposed cryptosystem. © 2022, Editorial Board of Acta Armamentarii. All right reserved.