Article Preview
Top1. Introduction
Steganography is a technology in the field of information hiding. It is defined as concealing a file, message, image, or video within another carrier medium. In this paper, we focus on steganography in image, one of the most commonly used medium. In an image steganography case (Abadi & Andersen, 2016), Alice encodes a secret message in an image and send it through the public channel to Bob, who can decode the image and extract secret message from it. But Eve can’t distinguish whether the image Alice sent contains secret or not. Eve’s detecting stego image is called steganalysis, a technology against steganography.
Generally, security, robustness, and capacity are three metrics to measure a steganography algorithm. Security means the difficulty for a steganalysis algorithm to detect an stego image with certain message. Robustness is the ability of the hidden message to remain undamaged even if the stego media undergoes transformation, sharpening, linear and non-linear filtering, scaling, blurring, cropping and various other techniques. Capacity is the average amount of information hidden in a unit, which is usually measured by bit-per-pixel.
There are a wide variety of steganography algorithms for different applications. At the very beginning, least significant bit(LSB) method changes the least significant bit of a pixel to hide secret message. Variants of LSB methods are described in (Wolfgang & Delp, 1996). Although cover image using LSB methods can’t be found visually, it can be easily detected by analyzing the statistics of image. To solve this problem, steganography algorithms such as HUGO (Pevný et al., 2010), HILL (Li et al., 2014), WOW (Holub & Fridrich, 2012) and S-UNIWARD (Holub et al., 2014) hide secret information by minimizing a distortion metric.
With rapid development of deep convolution neural networks, generative adversarial network(GAN) (Goodfellow et al., 2014) was proposed and is widely used in computer vision. There are lots of works that focus on improving GAN, especially on stabilizing the training process. WGAN (Arjovsky et al., 2017) introduced Wasserstein distance that can accurately measure the distance between two different distributions instead of Kullback–Leibler(KL) divergence. Later on, Gulrajani et al. (2017) used gradient penalty instead of weight clipping in WGAN. Spectral normalization (Miyato et al., 2018) further improves the performance. Jolicoeur-Martineau (2018) used relativistic discriminator not only to judge whether a sample is real or not, but also to increase the probability that generated data are real. Except improvement in discriminator, Heusel et al. (2017) proposed different learning rate for discriminator and generator is also useful. In our work, many techniques are adopted to stabilize the training process. GAN is also introduced to steganography these years. As a pioneer, Volkhonskiy et al. (2017) proposed SGAN. They used GAN to generate some images as cover, and hid secret in the generated images. Based on their work, Shi et al. (2017) used WGAN to improve the quality of cover images, decreasing the artifacts. Conservatively, ASDL-GAN (Yang et al., 2018), which used neural network to generate a distortion metric, was proposed. Recently, Rehman et al. (2017) proposed the end-to-end training CNN, an encode-decode structure. This kind of algorithm is a blind steganography method in which receiver is unnecessary to have access to the original image, so it is automatic and very convenient to use in practice. With such advantages, the encode-decode structure has been widely used ever since. A deep steganography method that can hide an image to another image is proposed by Baluja (2017). Based on his work, Dong et al. (2018) introduced a network that hid a grey-scale image to a colorful image with GAN. However, the secret images revealed by their work have low quality, suffering from severe noise.