Optical Image Encryption Vulnerability Detection by Using Generative Adversarial Networks

Although the rapid development of science and technology makes information transmission more convenient, it also brings about information leakage. Therefore, information needs to be encrypted. Among many encryption methods, optical encryption based on computer generated hologram (CGH) is very popular for its parallelism and low cost. Generally, unauthorized receivers are unable to realize decoding without the encryption rule, security key or decryption instrument. However, the vulnerability and security of CGH based optical encryption has not been verified effectively. In this paper, a new type of Deep Learning (DL) framework based on Generative Adversarial Networks (GAN) is proposed to restore unknown plaintexts. Unauthorized users can extract information from encrypted images by the learning model without a direct retrieval of the encryption rule or security key. Meanwhile, a classification network is added into GAN to improve its classification accuracy. Experimental results show that the convolutional neural network (CNN) is robust to the encryption rule. The applicability of this network for different optical encryption algorithms is also verified. The proposed method is demonstrated to be effective and feasible to the cryptoanalysis of CGH based optical encryption systems, which makes an effective contribution to the the improvement of optical encryption technology.