Evolving Generative Adversarial Networks to improve image steganography

Images have been repeatedly used as the perfect environment to hide information through the use of steganography techniques. Whether messages, documents or even other images, the bitmap of an digital picture provides a place where hidden data can be embedded without human notice. So far, a plethora of steganography methods can be found in the state-of-the-art literature, together with steganalysis techniques, devoted to detect the presence of hidden information in files. Recent steganography techniques rely on Convolutional Neural Networks, trying to embed as information as possible while minimising visual changes in the image. Following this trend, this article tries to demonstrate that a Generative Adversarial Network (GAN) can be used to improve the ability of a spatial domain steganalysis method and to insert secret information with minimal image alteration. Through a training process, the GAN learns how to adapt an image to later introduce a message using the Least Significant Bit steganography algorithm. The results evidence that the approach is successful at avoiding detection by a state-of-the-art Deep Learning steganalysis architecture.

Automated summary

Images are commonly used for hiding information using steganography techniques. A wide range of steganography methods and steganalysis techniques are available, with recent techniques relying on Convolutional Neural Networks to minimize visual changes. This article demonstrates the use of a Generative Adversarial Network (GAN) to enhance a spatial domain steganalysis method and insert secret information with minimal image alteration. The results show that this approach successfully avoids detection by a state-of-the-art Deep Learning steganalysis architecture.