Abstract
Image security in visual data is growing in importance due to the increasing use of digital images in many applications. This paper describes a novel image encrypting system using deep generative models with a self-attention mechanism to improve encrypting speed, decrypting speed, or both. Based on the CycleGAN models, the encrypting generator encrypts images while a dedicated decryption network decrypts them properly. The self-attention module improves the dispersibility of visual data to capture global dependencies among images at variable scales to achieve a highly secure image transformation with high reconstruction accuracy. Performance analysis has been done using two color image sets, brain MRI images to detect brain tumors, and images to detect skin cancer. Analysis indicates high security (Entropy = 7.9996, NPCR ≈ 99.99%), high reconstruction accuracy (SSIM ≈ 0.99, PSNR > 40 dB), and high resistance to differential (UACI ≈ 33.46) and occlusion attacks to images. This paper opens up novel avenues in applying deep models to visual cryptography schemes.