Deep Learning for Improving the Robustness of Image Encryption

In this paper, we propose a method to increase the robustness of 2D/3D optical image encryption using the dilated deep convolutional neural network (CNN). In order to solve the problem that encrypted images suffer from some attacks in practical application, we utilize a fast and effective CNN denoiser based on the principle of deep learning. The CNN improves the robustness of the algorithm by improving the resolution of the reconstructed images. Besides, CNN has a high performance against blur and occlusion attacks. We introduce the pixel scrambling method to enhance the security level of the encryption by the private key of pixel scrambling operation. The proposed method can not only realize the encryption of a two-dimensional image but also implement three-dimensional image encryption by combining the integral imaging technology. Double random phase encoding in the fractional Fourier domain is selected for experimental verification, and the results show the capability for robustness, noise immunity, and security of the proposed method.