Enhanced secure lossless image steganography using invertible neural networks

Abstract

Image steganography is a technique that embeds secret data into cover images in an imperceptible manner, ensuring that the original data can be recovered by the receiver without arousing suspicion. The key challenges currently faced by image steganography are capacity, invisibility, and security. We suggest an invertible neural network-based image steganography technique to concurrently address these three issues. To achieve better invisibility, we adopt a method that avoids the loss of information, thereby preventing ill-posed problems. The learning cost during image embedding can be reduced by only fitting part of the color channels in order to address the issue of high capacity. Additionally, we introduce the concept of a key to constrain the embedding process of the secret information, significantly enhancing the security of the hidden data. According to our experimental results, our method outperforms other image steganography algorithms on DIV2K, COCO, and ImageNet datasets, achieving perfect recovery of the secret images, its PSNR and SSIM can reach the theoretical maximum values.