Method and system for large-capacity image steganography and recovery based on invertible neural networks

Patent metadata

Field	Value
Publication number	US-11908037-B2
Application number	US-202117436755-A
Country	US
Kind code	B2
Filing date	Jun 18, 2021
Priority date	Feb 4, 2021
Publication date	Feb 20, 2024
Grant date	Feb 20, 2024

Abstract

The present disclosure provides a method and system for large-capacity image steganography and recovery based on an invertible neural networks. The method is intended to embed one or more hidden images into a single host image, and recover all the hidden images from a stego image. The method designs an image steganography model that supports bidirectional mapping. The model includes cascaded invertible modules containing a host branch and a hidden branch. A hidden image is embedded into a host image through forward mapping to form a stego image, and the host image and the hidden image are separated and recovered from the single stego image through reverse mapping.

First claim

What is claimed is:

1. A method for large-capacity image steganography and recovery based on an invertible neural networks, wherein an invertible neural networks model that supports bidirectional mapping is configured to complete tasks of embedding hidden images into a host image and recovering the hidden images from a stego image, and the invertible neural networks model comprises cascaded invertible modules containing a host branch and a hidden branch, the method comprising the following steps:
a first part: embedding hidden images
a. inputting the host image and the hidden image: importing one or more hidden images and a selected host image into a forward mapping input end of the invertible neural networks model as forward inputs;
b. embedding image features: embedding feature information of the hidden image into the host image through the forward mapping of the invertible neural networks; and
c. generating a stego image: taking forward output of a host branch of the invertible neural networks as the stego image, and specifying forward output of a hidden branch as a constant matrix containing no effective information;
a second part: recovering the hidden images
d. inputting the stego image: importing the stego image and the constant matrix obtained in step c into a reverse mapping input end of the invertible neural networks model as reverse inputs;
e. separating the image features: separating features of the host image and the hidden image from the stego image through the reverse mapping of the invertible neural networks; and
f. recovering the hidden image: taking reverse output of the host branch of the invertible neural networks as a recovered host image, and reverse output of the hidden branch as a recovered hidden image.

2. The method for large-capacity image steganography and recovery based on an invertible neural networks according to claim 1 , wherein all learnable parameters are shared for the forward mapping and the reverse mapping of the invertible neural networks model.

3. A method for large-capacity image steganography based on an invertible neural networks, wherein the method for large-capacity image steganography is implemented based on the invertible neural networks, the invertible neural networks comprises invertible modules, the invertible module comprises a host branch and a hidden branch, and the method for large-capacity image steganography comprises:
inputting a host image and a hidden image;
embedding feature information of the hidden image into the host image through forward mapping of the invertible neural networks;
forwardly outputting a stego image by utilizing the host branch of the invertible neural networks, wherein the stego image is an image obtained by embedding the feature information of the hidden image into the host image; and
forwardly outputting a constant matrix by utilizing the hidden branch of the invertible neural networks, wherein the constant matrix does not contain effective information of the hidden image.

4. The method for large-capacity image steganography based on an invertible neural networks according to claim 3 , wherein there are a plurality of the invertible modules in the invertible neural networks which are cascaded, and the invertible neural networks comprising the plurality of the invertible modules simultaneously completes steganography process of a plurality of the hidden images.

5. A method for large-capacity image recovery based on an invertible neural networks, wherein the method for large-capacity image recovery is applied to the method for large-capacity image steganography according to any one of claim 4 , the method for large-capacity image recovery is implemented based on the invertible neural networks, the invertible neural networks comprises invertible modules which comprises a host branch and a hidden branch, and the method for large-capacity image recovery comprises:
inputting a stego image and a constant matrix, wherein the stego image is an image obtained by embedding feature information of a hidden image into a host image; and the constant matrix does not contain effective information of the hidden image;
separating features of the hidden image and the host image from the stego image through reverse mapping of the invertible neural networks;
reversely outputting a recovered host image by utilizing a host branch of the invertible neural networks; and
reversely outputting a recovered hidden image by utilizing a hidden branch of the invertible neural networks.

6. The method for large-capacity image recovery based on an invertible neural networks according to claim 5 , wherein there are a plurality of the invertible modules in the invertible neural networks which are cascaded, and the invertible neural networks comprising the plurality of the invertible modules simultaneously completes recovery process of a plurality of the hidden images.

7. A system for large-capacity image recovery based on an invertible neural networks, wherein the system for large-capacity image recovery is applied to the system for large-capacity image steganography according to any one of claim 6 , the system for large-capacity image recovery is implemented based on the invertible neural networks, and the invertible neural networks comprises invertible modules which comprises a host branch and a hidden branch, the system for large-capacity image recovery comprising:
an input module, configured to input a stego image and a constant matrix, wherein the stego image is an image obtained by embedding feature information of a hidden image into a host image, and the constant matrix does not contain effective information of the hidden image;
an information separation module, configured to separate features of the hidden image and the host image from the stego image through reverse mapping of the invertible neural networks;
a host image output module, configured to reversely output a recovered host image by utilizing a host branch of the invertible neural networks; and
a hidden image output module, configured to reversely output a recovered hidden image by utilizing a hidden branch of the invertible neural networks.

8. The system for large-capacity image recovery based on an invertible neural networks according to claim 7 , wherein there are the plurality of the invertible modules in the invertible neural networks which are cascaded, and the invertible neural networks comprising the plurality of the invertible modules simultaneously completes recovery process of the plurality of the hidden images.

9. A system for large-capacity image recovery based on an invertible neural networks, wherein the system for large-capacity image recovery is applied to the system for large-capacity image steganography according to any one of claim 5 , the system for large-capacity image recovery is implemented based on the invertible neural networks, and the invertible neural networks comprises invertible modules which comprises a host branch and a hidden branch, the system for large-capacity image recovery comprising:
an input module, configured to input a stego image and a constant matrix, wherein the stego image is an image obtained by embedding feature information of a hidden image into a host image, and the constant matrix does not contain effective information of the hidden image;
an information separation module, configured to separate features of the hidden image and the host image from the stego image through reverse mapping of the invertible neural networks;
a host image output module, configured to reversely output a recovered host image by utilizing a host branch of the invertible neural networks; and
a hidden image output module, configured to reversely output a recovered hidden image by utiliz

Assignees

• Univ Nankai

Inventors

• Lu Shaoping
• Wang Rong
• Zhong Tao

Classifications

• G06T1/0021 Primary

  Image watermarking · CPC title

• H04N1/32347 Primary

  Reversible embedding, i.e. lossless, invertible, erasable, removable or distorsion-free embedding · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06T2201/0203

  whereby the image with embedded watermark is reverted to the original condition before embedding, e.g. lossless, distortion-free or invertible watermarking · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title