High-accuracy image steganography with invertible neural network and generative adversarial network

Abstract

Image steganography conceals secret messages imperceptibly within cover images. However, many existing deep learning-based image steganography methods have limitations in visual quality, payload size, and security. Specifically, they often require error correction codes for complete message extraction. In this paper, we propose a novel image steganography network architecture based on Invertible Neural Network (INN) and Generative Adversarial Network (GAN). Leveraging the reversibility of INN and the connection with the hidden network, we design three extraction networks based on DenseNet, shared weights, and unshared weights. These are respectively combined with the hidden network and discriminator network to create new network structures, effectively improving invisibility and message extraction accuracy. Furthermore, the discriminator participates in adversarial training by comparing cover and stego images in a patch-to-patch manner, thereby enhancing visual quality and security. Extensive experiments demonstrate the effectiveness of our proposed method across various aspects, including image quality, payload, extraction accuracy, and security, particularly achieving close to 100% message extraction accuracy without requiring error correction codes.