GAN-INN-Steg: a simple and accurate method for image steganography

Abstract

We propose an efficient and accurate image steganography method that combines generative adversarial networks (GAN) with reversible neural network models. In this approach, the reversible neural network model is used to construct the generator of the GAN. Then, this combined GAN is trained using adversarial training on a fixed dataset, allowing for the rapid acquisition of a generator that can efficiently fit the image distribution of the dataset. Leveraging the bidirectional mapping capability of this generator, we can achieve bidirectional mapping between input noise and forged images. By designing a mapping method for secret messages and noise, we can synthesize several bits of secret messages into a stego image, which can then undergo direct and lossless message extraction using the reverse generator. Unlike other steganography methods, this model can quickly complete iterative training for image steganography, offering enhanced efficiency, security, substantial steganographic capacity, and complete reversibility. We have experimentally validated this method and conducted theoretical analysis and experimental testing on the quality, security, reversibility, and steganographic capacity of the stego images.