CN116112685A - Image steganography method based on diffusion probability model - Google Patents

Image steganography method based on diffusion probability model Download PDF

Info

Publication number
CN116112685A
CN116112685A CN202310121388.XA CN202310121388A CN116112685A CN 116112685 A CN116112685 A CN 116112685A CN 202310121388 A CN202310121388 A CN 202310121388A CN 116112685 A CN116112685 A CN 116112685A
Authority
CN
China
Prior art keywords
message
image
probability model
diffusion
diffusion probability
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310121388.XA
Other languages
Chinese (zh)
Inventor
刘佳
张卓
罗鹏
柯彦
张敏情
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Engineering University of Chinese Peoples Armed Police Force
Original Assignee
Engineering University of Chinese Peoples Armed Police Force
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Engineering University of Chinese Peoples Armed Police Force filed Critical Engineering University of Chinese Peoples Armed Police Force
Priority to CN202310121388.XA priority Critical patent/CN116112685A/en
Publication of CN116112685A publication Critical patent/CN116112685A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/46—Embedding additional information in the video signal during the compression process
    • H04N19/467—Embedding additional information in the video signal during the compression process characterised by the embedded information being invisible, e.g. watermarking
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
    • H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
    • H04N19/146—Data rate or code amount at the encoder output
    • H04N19/149—Data rate or code amount at the encoder output by estimating the code amount by means of a model, e.g. mathematical model or statistical model

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Algebra (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Image Processing (AREA)

Abstract

The invention provides an image steganography method based on a diffusion probability model, which comprises the following steps of: s1, training a diffusion probability model to generate a carrier image set; s2, training a message extractor to establish a message-carrier set mapping; s3, using the stage 1, hiding the secret message; s4, using the stage 2, extracting the secret information ， The invention can combine the message extractor with the convolutional neural network structure on the basis of any generator of the diffusion probability model which is directly driven by Gaussian noise to generate the image, so that the generated image can contain secret messages; in addition, by virtue of the gradual iteration of the diffusion probability model, the generated image is free from distortion caused by reconstruction, and the stability of training has higher image quality than the generator based on the generation countermeasure network structure.

Description

Image steganography method based on diffusion probability model
Technical Field
The invention relates to the technical field of deep learning, in particular to an image steganography method based on a diffusion probability model.
Background
Conventional image steganography techniques typically implement embedding of secret messages by modifying the original carrier method, which includes least significant bit substitution or steganography based on a distortion cost function, etc., which typically causes changes to the statistical properties of the image, thereby causing attacks by steganography analysts.
With the advent of deep learning techniques, researchers are currently working on implementing steganography using some method of deep learning in order to improve the concealment of steganography. The most important of the deep learning image steganography technology is to improve the implementation means or performance of the image steganography method by means of a deep neural network model. Such methods can be generally classified into three categories, one is to construct some false original carriers using generative models, and then embed the message using conventional steganography techniques. Another class is to use a generative model to generate modified policies for message embedding using the distortion cost function in conventional modification-based policies. Both of these approaches utilize generative models to improve upon traditional carrier modification-based steganography. The third type of method is to construct a dense carrier directly using deep learning, which can be classified into reference-made steganography and no-reference-made steganography depending on whether an explicit original carrier is required. The reference generation type steganography generally uses a deep neural network to input a message and a certain original carrier as a network, and the neural network outputs a carrier containing secret by means of strong coding capability of the neural network. In order to achieve extraction of the message, it is often also necessary to achieve extraction of the message by means of a message extractor. More recently, techniques have also emerged that use a reversible neural network to implement steganography, combining message encoding and message extraction into a model, thereby avoiding the process of separately training a message extractor. The non-reference generation type steganography mainly depends on the generation capability of a depth generation model, a noise or a message is input to a generator, the generator directly generates a secret-containing carrier, the secret-containing carrier in the method only depends on the generator and does not depend on a specific original carrier, the image steganography method Based on the depth generation model mainly depends on the development of the depth generation model, researchers currently realize the image generation type steganography Based on a generation countermeasure network (Generative Adversarial Networks (GAN)), and some researchers also use the generation type steganography Based on a Flow-Based model. The GAN model is considered to have the characteristics of potential unstable training and small generation difference due to its resistance training characteristics, so that the image steganography technology based on the GAN model also has the above disadvantages, in particular, the generated image is not high in quality, and meanwhile, the message extraction accuracy is low. The flow model must use a specially designed model structure to construct the reversible transformations, which is relatively complex to implement. Recently, a generative model-diffusion probability model (Diffusion Probabilistic Models) has emerged in the generative model field. The diffusion probability model is inspired by unbalanced thermodynamics, and firstly defines a diffusion process based on Markov chains, namely slowly adding random noise into data until the data becomes Gaussian noise positions, and then iteratively and gradually learning an inverse diffusion process to gradually transform the noise into required data samples. The model regards the generation process as a step-by-step iterative denoising process, so that only one simple neural network model is utilized to simulate Gaussian noise at a time, and the stability of model training is ensured.
In order to further improve the generation efficiency and quality of the generated secret carrier, an image steganography method based on a diffusion probability model is provided on the basis of the diffusion probability model.
Disclosure of Invention
In view of the foregoing, it is desirable to provide an image steganography method based on a diffusion probability model to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial choice.
The technical scheme of the embodiment of the invention is realized as follows:
an image steganography method based on a diffusion probability model comprises the following steps:
s1, training a diffusion probability model to generate a carrier image set;
s2, training a message extractor to establish a message-carrier set mapping;
s3, using the stage 1, hiding the secret message;
s4, extracting secret information by using the step 2.
Further preferred is: in the step S1, the training phase 1 includes the following steps:
s11: diffusion probability model G using a large number of real image samples θ Training to obtain an image generator G θ Wherein θ is a model parameter;
s12: using trained diffusion probability models G θ Iterating for T times to generate an image I g ；
S13: repeating the step in S12 to generate a large number of carrier image sets { I } g }。
Further preferred is: in S2, training phase 2 comprises the steps of:
s21: given a secret message M in Initializing a message extraction network
Figure BDA0004080048350000031
For message extraction, wherein->
Figure BDA0004080048350000032
Representing parameters of the extraction network E.
S22, utilizing the message extraction network E to generate a carrier image I in the step S21 stego Extracting message M out Calculate the extracted secret message M out With secret message M to be embedded in Message loss L between M Wherein;
Figure BDA0004080048350000033
the message extractor parameters are optimized by minimizing message loss.
Further preferred is: in S3, the use phase 1 includes the following steps:
s31, after S1 and S2 are completed, generating a new generated image I by using a diffusion probability model g The image satisfies the mapping relation of the message-carrier, and can be used as the image I containing the secret stego 。
Further preferred is: in S4, the use phase 2 includes the following steps:
s41, inputting the dense image obtained in S3 into a trained message extraction network E (I stego ) Output message M out I.e. a secret message.
Further preferred is: the diffusion probability model is a diffusion probability model, and the specific process of generating the steganographic image by the DPM model is also divided into two stages: the model training stage is divided into a forward diffusion process and a reverse diffusion process, model parameters are trained in the reverse diffusion process, and the diffusion process is a process of gradually adding noise to an image.
Further preferred is: the data sets used for pre-training the diffusion probability model are CIFAR10 and CelebA-HQ, wherein the CIFAR10 is a small image data set for pattern recognition, which is created by Alex Krizhevsky and Ilya Sutskever, and comprises 10 kinds of RGB color pictures, each picture has the size of 32X 32 pixels, and the data set comprises 50000 training images and 10000 test images; celebA-HQ, the data set is a high quality version of CelebA, comprising 3 ten thousand images with 1024×1024 resolution, is compressed to 256×256 size when training the diffusion probability model in order to increase model operation efficiency.
Further preferred is: the U-Net structure for estimating the parameters of a probabilistic model that follows a double-flow convolution structure while introducing residual connections.
Further preferred is: the message extraction network, the decoder contains 7 convn-bn-relu blocks with 64 filters and the last cconvn-bn-relu block with L filters, while pooling is performed in all spatial dimensions, the final linear layer yielding the predicted message Mout.
By adopting the technical scheme, the embodiment of the invention has the following advantages:
the invention can combine the message extractor with the convolutional neural network structure on the basis of any generator of the diffusion probability model which is directly driven by Gaussian noise to generate the image, so that the generated image can contain secret messages; in addition, by virtue of the gradual iteration of the diffusion probability model, the generated image is free from distortion caused by reconstruction, and the stability of training has higher image quality than the generator based on the generation countermeasure network structure.
The foregoing summary is for the purpose of the specification only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the present invention will become apparent by reference to the drawings and the following detailed description.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are required in the embodiments or the technical descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a frame diagram of an image steganography method based on a diffusion probability model according to an embodiment of the present invention;
FIG. 2 is a block diagram of a noise estimation network in a diffusion probability model according to an embodiment of the present invention;
FIG. 3 is a block diagram of a message extractor according to an embodiment of the present invention;
FIG. 4 is a dense-containing image generated by a diffusion probability model in accordance with an embodiment of the present invention.
Detailed Description
Hereinafter, only certain exemplary embodiments are briefly described. As will be recognized by those of skill in the pertinent art, the described embodiments may be modified in various different ways without departing from the spirit or scope of the present invention. Accordingly, the drawings and description are to be regarded as illustrative in nature and not as restrictive.
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
Examples
As shown in fig. 1, an embodiment of the present invention provides an image steganography method based on a diffusion probability model, including the following steps:
s1, training a diffusion probability model to generate a carrier image set;
s2, training a message extractor to establish a message-carrier set mapping;
s3, using the stage 1, hiding the secret message;
s4, extracting secret information by using the step 2.
Further preferred is: in the step S1, the training phase 1 includes the following steps:
s11: diffusion probability model G using a large number of real image samples θ Training to obtain an image generator G θ Wherein θ is a model parameter;
s12: using trained diffusion probability models G θ Iterating for T times to generate an image I g ；
S13: repeating the step in S12 to generate a large number of carrier image sets { I } g }。
Further preferred is: in S2, training phase 2 comprises the steps of:
s21: given a secret message M in Initializing a message extraction network
Figure BDA0004080048350000051
For message extraction, wherein->
Figure BDA0004080048350000052
Representing parameters of the extraction network E.
S22, utilizing the message extraction network E to generate a carrier image I in the step S21 stego Extracting message M out Calculate the extracted secret message M out With secret message M to be embedded in Message loss L between M Wherein;
Figure BDA0004080048350000061
the message extractor parameters are optimized by minimizing message loss.
Further specific is: in S3, the use phase 1 includes the following steps:
s31, after S1 and S2 are completed, generating a new generated image I by using a diffusion probability model g The image satisfies the mapping relation of the message-carrier, and can be used as the image I containing the secret stego 。
Further specific is: in S4, the use phase 2 includes the following steps:
s41, inputting the dense image obtained in S3 into a trained message extraction network E (I stego ) Output message M out I.e. a secret message.
Further, the diffusion probability model is a diffusion probability model (diffusion probabilistic models), and as shown in reference to fig. 2, the specific process of generating a steganographic image by the DPM model is also divided into two stages: a model training phase and an image sampling phase. The model training stage is divided into a forward diffusion process (noise adding) and a reverse diffusion process (noise removing) and training model parameters in the reverse diffusion process. Specifically, the diffusion process is a process of gradually denoising an image, which is a Markov process, and can be expressed as q (x t |x t-1 ) I.e. given x t-1 Under the condition of x t Obeying means
Figure BDA0004080048350000062
Variance is beta t Normal distribution of I. The core goal is to obtain a picture from any noise picture by sampling once in the back diffusion process to achieve the purpose of picture generation, but q (x) is difficult to estimate t-1 |x t ) Therefore, we need to estimate a distribution p using neural networks θ (x t-1 |x t ) To approximate q (x) t-1 |x t ) In the DPM model, noise at time t is predicted by using a self-encoder (Auto-encoder) of U-Net structure. In the training stage, real image x is originally utilized 0 One gaussian noise e estimation model parameter e θ . And in the sampling stage, after model training is finished, iteration is performed through T steps to iteratively generate a carrier image.
Specifically, the data sets used for pre-training the diffusion probability model are CIFAR10 and CelebA-HQ, wherein CIFAR10 is a small image data set for pattern recognition created by Alex Krizhevsky and Ilya Sutskevver, the data set comprises 10 kinds of RGB color pictures, each picture has a size of 32×32 pixels, and the data set comprises 50000 training images and 10000 test images. CelebA-HQ, the data set is a high quality version of CelebA, comprising 3 ten thousand images with 1024×1024 resolution, is compressed to 256×256 size when training the diffusion probability model in order to increase model operation efficiency.
Further, the U-Net structure for estimating the probability model parameters, as shown in FIG. 3, follows a double-flow (downward and downward + right) convolution structure, while introducing residual connections; first, the architecture combines downsampling and upsampling such that the internal portions of the network run on a larger spatial scale, improving computational efficiency. Second, the architecture employs remote jumpers, such that every kth layer provides a direct input for the (K-K) th layer, where K is the total number of layers in the network. The network is divided into six layers of sequences, most of which are separated by downsampling or upsampling.
Still further, the message extraction network, as shown in fig. 4, the decoder contains 7 convn-bn-relu (convolutional-batch normalization-linear rectification active) blocks with 64 filters and the last cconvn-bn-relu block with L filters. Pooling (sampling) performed in all spatial dimensions simultaneously, the final linear layer (lxl) yields the predicted message Mout.
The invention can combine the message extractor with the convolutional neural network structure on the basis of any generator of the diffusion probability model which is directly driven by Gaussian noise to generate the image, so that the generated image can contain secret messages; in addition, by virtue of the gradual iteration of the diffusion probability model, the generated image is free from distortion caused by reconstruction, and the stability of training has higher image quality than the generator based on the generation countermeasure network structure.
The foregoing is merely illustrative of the present invention, and the present invention is not limited thereto, and any person skilled in the art will readily recognize that various changes and substitutions are possible within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (9)

1. An image steganography method based on a diffusion probability model is characterized by comprising the following steps of:
s1, training a diffusion probability model to generate a carrier image set;
s2, training a message extractor to establish a message-carrier set mapping;
s3, using the stage 1, hiding the secret message;
s4, extracting secret information by using the step 2.
2. The diffusion probability model-based image steganography method of claim 1, characterized in that: in the step S1, the training phase 1 includes the following steps:
s11: diffusion probability model G using a large number of real image samples θ Training to obtain an image generator G θ Wherein θ is a model parameter;
s12: using trained diffusion probability models G θ Iterating for T times to generate an image I g ；
S13: repeating the step in S12 to generate a large number of carrier image sets { I } g }。
3. The diffusion probability model-based image steganography method of claim 1, characterized in that: in S2, training phase 2 comprises the steps of:
s21: given a secret message M in Initializing a message extraction network
Figure FDA0004080048340000011
For message extraction, wherein->
Figure FDA0004080048340000012
Representing parameters of the extraction network E.
S22, utilizing the message extraction network E to generate a carrier image I in the step S21 stego Extracting message M out Calculate the extracted secret message M out With secret message M to be embedded in Message loss L between M Wherein;
Figure FDA0004080048340000013
the message extractor parameters are optimized by minimizing message loss.
4. The diffusion probability model-based image steganography method of claim 1, characterized in that: in S3, the use phase 1 includes the following steps:
s31, after S1 and S2 are completed, generating a new generated image I by using a diffusion probability model g The image satisfies the mapping relation of the message-carrier, and can be used as the image I containing the secret stego 。
5. The diffusion probability model-based image steganography method of claim 1, characterized in that: in S4, the use phase 2 includes the following steps:
s41, inputting the dense image obtained in the S3 into trainingE in good message extraction network, and output message M out I.e. a secret message.
6. The diffusion probability model-based image steganography method of claim 1, characterized in that: the diffusion probability model is a diffusion probability model, and the specific process of generating the steganographic image by the DPM model is also divided into two stages: the model training stage is divided into a forward diffusion process and a reverse diffusion process, model parameters are trained in the reverse diffusion process, and the diffusion process is a process of gradually adding noise to an image.
7. The diffusion probability model-based image steganography method of claim 1, characterized in that: the data sets used for pre-training the diffusion probability model are CIFAR10 and CelebA-HQ, wherein the CIFAR10 is a small image data set for pattern recognition, which is created by Alex Krizhevsky and Ilya Sutskever, and comprises 10 kinds of RGB color pictures, each picture has the size of 32X 32 pixels, and the data set comprises 50000 training images and 10000 test images; celebA-HQ, the data set is a high quality version of CelebA, comprising 3 ten thousand images with 1024×1024 resolution, is compressed to 256×256 size when training the diffusion probability model in order to increase model operation efficiency.
8. The diffusion probability model-based image steganography method of claim 1, characterized in that: the U-Net structure for estimating the parameters of a probabilistic model that follows a double-flow convolution structure while introducing residual connections.
9. The diffusion probability model-based image steganography method of claim 1, characterized in that: the message extraction network, the decoder contains 7 convn-bn-relu blocks with 64 filters and the last cconvn-bn-relu block with L filters, while pooling is performed in all spatial dimensions, the final linear layer yielding the predicted message Mout.
CN202310121388.XA 2023-02-16 2023-02-16 Image steganography method based on diffusion probability model Pending CN116112685A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310121388.XA CN116112685A (en) 2023-02-16 2023-02-16 Image steganography method based on diffusion probability model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310121388.XA CN116112685A (en) 2023-02-16 2023-02-16 Image steganography method based on diffusion probability model

Publications (1)

Publication Number Publication Date
CN116112685A true CN116112685A (en) 2023-05-12

Family

ID=86255845

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310121388.XA Pending CN116112685A (en) 2023-02-16 2023-02-16 Image steganography method based on diffusion probability model

Country Status (1)

Country Link
CN (1) CN116112685A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116456037A (en) * 2023-06-16 2023-07-18 南京信息工程大学 Diffusion model-based generated image steganography method

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116456037A (en) * 2023-06-16 2023-07-18 南京信息工程大学 Diffusion model-based generated image steganography method
CN116456037B (en) * 2023-06-16 2023-08-22 南京信息工程大学 Diffusion model-based generated image steganography method

Similar Documents

Publication Publication Date Title
CN108122197B (en) Image super-resolution reconstruction method based on deep learning
CN105631807B (en) The single-frame image super-resolution reconstruction method chosen based on sparse domain
CN112288632B (en) Single image super-resolution method and system based on simplified ESRGAN
CN115660931A (en) Robust watermarking method based on Transformer and denoising diffusion model
CN116645260B (en) Digital watermark attack method based on conditional diffusion model
CN116091288A (en) Diffusion model-based image steganography method
CN116681584A (en) Multistage diffusion image super-resolution algorithm
He et al. SRDRL: A blind super-resolution framework with degradation reconstruction loss
CN113538246A (en) Remote sensing image super-resolution reconstruction method based on unsupervised multi-stage fusion network
CN116112685A (en) Image steganography method based on diffusion probability model
CN111489305B (en) Image enhancement method based on reinforcement learning
CN114820380B (en) Spatial domain steganographic carrier image enhancement method based on content self-adaption disturbance resistance
CN115713462A (en) Super-resolution model training method, image recognition method, device and equipment
CN115345768A (en) Robust watermark attack method and system based on neural network
CN109993701B (en) Depth map super-resolution reconstruction method based on pyramid structure
CN116456037A (en) Diffusion model-based generated image steganography method
Xu et al. A compact neural network-based algorithm for robust image watermarking
Liu et al. Facial image inpainting using multi-level generative network
CN113628090B (en) Anti-interference message steganography and extraction method, system, computer equipment and terminal
Xiao et al. Effective PRNU extraction via densely connected hierarchical network
Huang et al. CLSR: cross-layer interaction pyramid super-resolution network
US20230186608A1 (en) Method, device, and computer program product for video processing
CN114648436A (en) Screen shot resistant text image watermark embedding and extracting method based on deep learning
CN111275603B (en) Security image steganography method based on style conversion and electronic device
Zhou et al. Disentangled Feature Distillation for Light Field Super-Resolution with Degradations

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination