Invertible Mosaic Image Hiding Network for Very Large Capacity Image Steganography
1. Introduction
Image steganography refers to the technique of hiding secret images within a cover image. The resulting merged image, known as a stego image, needs to look similar to the cover image while successfully carrying the concealed information. Previous methods of image steganography have faced issues of low hiding capacity and a lack of accuracy during the image reconstruction process. Recent innovations in steganography have aimed to combat these challenges through the use of handcrafted features, with some success. However, there is still a need for steganography methods capable of very large-capacity image hiding, which can maintain the quality and detail of the secret images.
The Invertible Mosaic Image Hiding Network (InvMIHNet) is a unique approach to very large capacity image steganography, where multiple secret images can be hidden within a single cover image. By concealing a mosaic of secret images, InvMIHNet can maintain high quality and accuracy in the resulting stego image. This is enabled through the use of an Invertible Image Rescaling (IIR) module and an Invertible Image Hiding (IIH) module. This novel approach has been found to surpass existing methods in both imperceptibility and recovery accuracy.
2. The Proposed Method
The proposed Invertible Mosaic Image Hiding Network begins the steganography process with the downsampling of the secret images. These downscaled images are then arranged into a mosaic, which is treated as a single image unit and hidden within the designated cover image. The created stego image and a hidden image-agnostic component can then be recovered through the reversibility of the IIR and IIH modules. As such, even if the secret images are downsized or their information is interchanged, the essential characteristics and details can be successfully recovered.
The proposed InvMIHNet operates in a two-stage process, with the secret images and cover image first being processed separately. The cover image and amalgamated mosaic are then processed together to create the final stego image that carries all the hidden information. With the use of Invertible Neural Networks, both of these processes are reversible, ensuring accurate reconstruction of the secret images. Detailed loss function expression is also provided to ensure the optimization of the image steganography processes.
3. Experiments and Results
Extensive experiments were conducted with the InvMIHNet on a variety of images from established datasets, using Nvidia's powerful RTX 3090 GPU. Theoretical and visual comparisons with state-of-the-art methods, such as DeepMIH and ISN, showed a clear improvement in the use of InvMIHNet. Specifically, InvMIHNet outperformed existing methods in terms of image quality, hiding capacity, and recovery accuracy, all while reducing computational cost and memory consumption.
InvMIHNet was found to be able to handle the informational input of up to 16 secret images, successfully embedding and reconstructing them within a single stego image. This degree of very-large-capacity image steganography sets InvMIHNet apart from current practices, where information from multiple images is interchanged, affecting the quality of the secret images.
4. Conclusion
The proposed Invertible Mosaic Image Hiding Network provides a balanced approach to the challenge of very-large-capacity image steganography. Combining the use of Invertible Neural Networks, Image Rescaling and Image Hiding modules, InvMIHNet ensures the efficient storage and accurate recovery of up to 16 secret images within a single stego image.
Despite already surpassing existing methods in a number of areas, further studies and improvements in invertible networks will continue to refine and enhance the capabilities of the InvMIHNet.