Abstract

Image hiding is a field of research that focuses on covert storage and transmission techniques. It involves embedding a secret image within a container image to create a classified-carrying image that resembles a normal image. Nevertheless,the current image hiding methods based on Invertible neural networks suffer from a critical issue of information loss during the hiding process. This leads to a substantial degradation in the quality of the secret image extracted,thereby preventing the simultaneous achievement of secure transmission, high-capacity transmission, and high fidelity of the secret image in an insecure network environment. To address this issue,we introduce a novel image hiding architecture called FMIN (Fitting Models Based on Invertible Network). FMIN incorporates our innovative fitting model, which fits the loss information and generates a variable z simulating the loss information at the receiver side. This variable z is used as an input to the revealing process,enabling the high-quality extraction of multiple secret images from a single classified loaded image. Additionally,we introduce a novel decoupled training strategy aimed at enhancing the stability of image hiding model during training. Experimental results demonstrate that the image hiding method based on the proposed FMIN architecture in this paper significantly outperforms other SOTA image hiding methods for single-image and multiple-images hiding.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call