Image reversible steganography method, device and system, and storage medium

Abstract

The invention discloses an image reversible steganography method, device and system, and a storage medium, and the method comprises the steps: S1, carrying out the adversarial training of a generator of a GAN model according to a data set; s2, the trained generator synthesizes a plurality of bits of secret messages into a secret-containing image through bidirectional mapping between bit streams of the secret messages and floating-point numbers of random noise; meanwhile, lossless message extraction is carried out on the secret-containing image through a reverse generator. By adopting the technical scheme of the invention, iterative training for image steganography can be quickly completed, and the method is more efficient and safer, has larger steganography capacity and is completely reversible.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular relates to an image reversible steganography method and device, a system, and a storage medium. Background Art

[0002] Traditional image steganography is to embed secret information into digital images, making it difficult for the human eye to detect, so as to achieve the purpose of hiding information and then realize hidden storage or transmission. This technology is widely used in many fields such as copyright protection, secret communication, digital watermarking, integrity verification, etc., and provides effective technical means for copyright tracking of digital media, illegal copy detection and confidential transmission of sensitive information. In recent years, with the advancement of deep learning technology, deep generative models have begun to be used for image steganography. The current image steganography methods based on generative models have the following problems: 1. The current carrier synthesis steganography method mostly uses the GAN (Generative Adversarial Networks, GAN) model, and the current GAN-based method has a low message embedding rate. Extracting messages requires the design and training of special extractors, and it is impossible to achieve completely reversible message extraction. 2. Some methods use the structure generation model of the Invertible Neural Networks (INN) for image steganography. Although there is no need to design and train a separate message extractor, the extraction accuracy still cannot reach 100%, and the hardware and time costs of training the model are very high. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method and device, a system and a storage medium for reversible image steganography.

[0004] To achieve the above object, the present invention adopts the following technical solution:

[0005] A reversible image steganography method, comprising:

[0006] Step S1: According to the data set, the generator of the GAN model is trained adversarially;

[0007] Step S2: The trained generator synthesizes several bits of secret message into a secret image through bidirectional mapping between the bit stream of the secret message and the floating point number of random noise; at the same time, the secret image is losslessly extracted through the reverse generator.

[0008] Preferably, in step S1, a generator of the GAN model is designed based on the reversible neural network model.

[0009] Preferably, in step S2, the output and input ends of the generator of the trained GAN model are reversed to obtain an extractor; the generator and the extractor are based on a bidirectional mapping, and information steganography and message extraction are performed between the sender and the receiver through a secret channel.

[0010] The present invention also provides an image reversible steganographic device, comprising:

[0011] The training module is used to perform adversarial training on the generator of the GAN model based on the data set;

[0012] The processing module is used to synthesize several bits of secret message into a secret image through bidirectional mapping between the bit stream of the secret message and the floating point number of random noise through the trained generator; at the same time, the secret image is losslessly extracted through the reverse generator.

[0013] Preferably, the training module designs a generator of the GAN model based on the reversible neural network model.

[0014] Preferably, the processing module is used to invert the output and input ends of the generator of the trained GAN model to obtain an extractor; the generator and the extractor are based on a bidirectional mapping, and perform information steganography and message extraction between the sender and the receiver through a secret channel.

[0015] An embodiment of the present invention further provides an image reversible steganography system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes the image reversible steganography method when executed by the processor.

[0016] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the image reversible steganography method when running.

[0017] The present invention combines a generative adversarial network with a flow model, uses a reversible neural network model to construct a generator of a generative adversarial network, and then conducts adversarial training on the combined generative adversarial network on a fixed data set, thereby quickly obtaining a generator that can efficiently fit the image distribution of the data set. By utilizing the bidirectional mapping characteristics of this generator, bidirectional mapping of secret messages and random noise can be completed, and several bits of secret messages can be synthesized into a secret image; wherein, the secret image is directly subjected to lossless message extraction through the reverse generator. Different from traditional methods, the present invention can quickly complete iterative training for image steganography, is more efficient and secure, has a larger steganographic capacity, and is completely reversible. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0019] Figure 1 This is a flow chart of the image reversible steganography method according to an embodiment of the present invention;

[0020] Figure 2 Schematic diagram of the mapping from secret messages to floating point numbers;

[0021] Figure 3 is the specific structure of the reversible neural network;

[0022] Figure 4 The architecture of the M module in the reversible neural network;

[0023] Figure 5 The loss value convergence of the model during training;

[0024] Figure 6 The loss value convergence of the model during training;

[0025] Figure 7 Schematic diagram of 225 images after the generator maps the secret message. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Embodiment 1:

[0029] The embodiment of the present invention provides an image reversible steganography method, which uses the advantages of the adversarial training architecture of the GAN model and the reversibility of the INN model to solve the problem of hiding a large amount of secret messages in an image file and extracting them losslessly. To this end, the embodiment of the present invention adopts the INN structure to design the idea of ​​the GAN generator, and then uses the adversarial training method of the GAN to quickly train the generator with a reversible architecture, and finally reverses the structure of the trained generator and directly uses it for reverse message extraction, such as Figure 1 As shown, specifically including:

[0030] Step S1: According to the data set, the generator of the GAN model is trained adversarially;

[0031] Step S2: The trained generator synthesizes several bits of secret message into a secret image through bidirectional mapping between the bit stream of the secret message and the floating point number of random noise; at the same time, the secret image is losslessly extracted through the reverse generator.

[0032] As an implementation method of an embodiment of the present invention, due to the inherent characteristics of a reversible neural network, its input is random noise in the interval (-1, 1) that conforms to a normal distribution, and its output is a secret image. If INN is needed for steganography, the bit stream of the secret message must be converted into a floating point number in the interval (-1, 1) before it can be used. To this end, it is necessary to design a mapping rule from bit stream to floating point numbers. However, the storage and calculation of floating point numbers in computers have a precision range. Data that exceeds the range cannot be accurately represented in a computer, which is undesirable. For reversible steganography, the data mapping cannot be simply designed, and the accuracy problem in the calculation must be considered; otherwise, when extracting the message, the wrong message will be extracted due to the calculation error of the floating point number. Based on the fact that the error accuracy of the reversible neural network after forward and reverse operations is within the level of 10-4, for this reason, the embodiment of the present invention designs the following data mapping method, such as Figure 2 shown.

[0033] Since the embodiment of the present invention maps the secret message to the pixel value of the image through a reversible neural network. The pixel value of a single-channel image usually contains 8 bits, so this process is designed to map the 8-bit secret message to a floating point number between (-0.999, 0.999), but it must be within the range of calculation accuracy, so that the floating point number can be reversely mapped to the 8-bit secret message during the reverse operation. Specifically, when the bit stream is mapped to the floating point number, it can be calculated according to formula (1):

[0034]

[0035] Among them, z i is the floating point number after mapping, s i is an 8-bit secret message string, and ()DEC is the corresponding decimal value. i A decimal random number not exceeding 3 digits.

[0036] When the receiver performs reverse operation on the received encrypted image, the floating point number is mapped into a bit stream and calculated according to formula (2):

[0037]

[0038] Among them, z i Indicates, [.]BIN-8 indicates 8-bit binary conversion.

[0039] Furthermore, adversarial training is performed on the image dataset to obtain the generator G, and the message extractor G can also be easily obtained. -1 , G and G -1 The secret information is sent to the sender and the receiver through a secret channel to generate a stego image and extract the secret information. The secret message can be embedded and extracted according to the following algorithm.

[0040] Information steganography algorithm:

[0041] Step 1: Divide the secret bit stream S into 8-bit segments (fill 0s if there are less than 8 bits) and record them as s i , take the m×m segment and record it as S', that is, S'={s i |k∈[1,m×m]};

[0042] Step 2: Map S′ to a floating point array Z using formula (1), Z = {z i |i∈[1,m×m]};

[0043] Step 3: Input Z into the generator G to obtain a stego image Ii;

[0044] Step4: Transmission II.

[0045] After training on the data set, the sender hides the secret information in the following way:

[0046] Message extraction algorithm:

[0047] Step 1: Accept Ii;

[0048] Step 2: Input Ii into G -1 Get a floating point array

[0049] Step 3: According to formula (2), the reverse mapping is the secret message

[0050] Step 4: Put several segments Put them together to get the complete secret message.

[0051] Experimental simulation:

[0052] The experimental platform is CPU Core i7-7700, memory 16g, graphics card 3080; operating system is Windows10Professional_64 bit; software platform is python3.8.13+numpy1.23.1+PyTorch1.13.0. The database uses the MNIST dataset, which contains 60,000 handwritten digit images from 0 to 9. The size of the images in the dataset is 28×28. The secret information is a random 0 and 1 bit sequence generated by numpy.

[0053] Train the model:

[0054] Design and build a reversible neural network based on the NICE model. The specific structure is as follows: Figure 3 As shown, CL is CouplingLayer, the model is composed of 8 stacked CLs, M is a neural network, and the specific structure is as follows Figure 4 As shown in the figure, FC is fully connected and Relu is the activation function. The generative adversarial network model was trained on the MNIST dataset. The batch size of the training samples (Batchsize) was set to 64, and the optimization method was Adam. The total training time was 12 epochs (about 12125 steps). The training results are shown in the figure below. Figure 5 shown.

[0055] The model training takes 3 minutes. After the training is completed, the trained model and parameters are reversely reconstructed to obtain a generator from simple distribution data (normal distribution data) to digital image data. After the model training converges and balances, the output of the generator is sampled. The results are as follows: Figure 6 shown.

[0056] The model was trained multiple times on the MNIST dataset, and it was proven that the model can be effectively trained and converged. The average training time was 3.3 minutes. At the same time, the flow model training method was also used to train the generator alone. Under the condition of the same output image quality, the average training time was 8 minutes, which exceeded the training time of the GAN structure.

[0057] Secret message embed:

[0058] The sampled data mapped with secret information is input into the generator to obtain the stego image and detect the quality of the image generated by the generator, such as Figure 7 shown.

[0059] As long as the messages can be extracted reliably from these generated images, they can be used as stego images for message transmission. Since the NICE model is reversible, when using floating-point numbers that do not exceed the computer's calculation precision, theoretically, 100% message extraction accuracy can be achieved. The following verifies the message extraction accuracy, assuming that the channel used in the communication is a reliable lossless channel (the receiver will receive lossless stego image data), directly uses the model to read the generated image and reversely calculate the floating-point data, and then reversely map it back to the bit stream of the secret message, and finally compare it with the original data. Each time, 100 groups of generated digital images (225 images per group) are sent to the model for reverse calculation and compared with the original data. A total of 10 such experiments were conducted, and the experimental results are shown in Table 1.

[0060] Table 1

[0061] Test Group Average extraction accuracy (%) 1 100 2 100 3 100 4 100 5 100 6 100 7 100 8 100 9 100 10 100

[0062] From Table 1, it can be seen that the accuracy of extracting secret information in each test is always maintained at 100%, which shows that the present invention can completely reversibly restore the secret.

[0063] The embodiment of the present invention combines the GAN model and the INN model in the generation model, uses the reversible property of the INN to build the generator in the GAN, and uses the GAN model adversarial training method to perform game training on the model, which can efficiently complete the training of the generator in the GAN model, reduce the hardware cost and time cost of training the generation model in such methods, and at the same time, the generator has image reversible steganography capability.

[0064] The embodiments of the present invention have the following technical effects:

[0065] (1) The embodiment of the present invention is based on a combined generation model of a stream model and a GAN model and an image steganography technology framework for specific applications, and completes the bidirectional mapping between secret messages and generated images through the stream model. By utilizing the reversibility of the INN model, the problem that current technologies of this type require a separate message extractor can be overcome. By utilizing the adversarial game training method of GAN, the training cost of INN can be reduced.

[0066] (2) The bidirectional mapping between secret messages and hidden variables in the flow model can overcome the common problem that the current carrier synthesis image steganography technology based on generative models cannot perform reversible steganography.

[0067] Since there are many types of reversible neural network models, the model of the GAN architecture in the embodiment of the present invention can flexibly select the current mainstream reversible neural network model for partial replacement, such as RealNVP, FLOW, Resnet, etc. Using the method of the embodiment of the present invention, generative image steganography can still be completed.

[0068] Embodiment 2:

[0069] The embodiment of the present invention further provides an image reversible steganographic device, comprising:

[0070] The training module is used to perform adversarial training on the generator of the GAN model based on the data set;

[0071] The processing module is used to synthesize several bits of secret message into a secret image through bidirectional mapping between the bit stream of the secret message and the floating point number of random noise through the trained generator; at the same time, the secret image is losslessly extracted through the reverse generator.

[0072] As an implementation method of an embodiment of the present invention, the training module designs a generator of the GAN model based on the reversible neural network model.

[0073] As an implementation mode of an embodiment of the present invention, the processing module is used to invert the output and input ends of the generator of the trained GAN model to obtain an extractor; the generator and the extractor are based on bidirectional mapping, and perform information steganography and message extraction between the sender and the receiver through a secret channel.

[0074] Embodiment 3:

[0075] An embodiment of the present invention further provides an image reversible steganography system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes the image reversible steganography method when executed by the processor.

[0076] Embodiment 4:

[0077] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the image reversible steganography method when running.

[0078] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A reversible image steganography method, characterized in that: include: Step S1: According to the data set, the generator of the GAN model is trained adversarially; Step S2: The trained generator synthesizes several bits of secret message into a secret image through bidirectional mapping between the bit stream of the secret message and the floating point number of random noise; at the same time, the secret image is losslessly extracted through the reverse generator.

2. The image reversible steganography method according to claim 1, characterized in that: In step S1, a generator of a GAN model is designed according to a reversible neural network model.

3. The image reversible steganography method according to claim 2, characterized in that: In step S2, the output and input ends of the generator of the trained GAN model are reversed to obtain an extractor; the generator and the extractor are based on a bidirectional mapping, and information steganography and message extraction are performed between the sender and the receiver through a secret channel.

4. An image reversible steganographic device, characterized in that: include: The training module is used to perform adversarial training on the generator of the GAN model based on the data set; The processing module is used to synthesize several bits of secret message into a secret image through bidirectional mapping between the bit stream of the secret message and the floating point number of random noise through the trained generator; at the same time, the secret image is losslessly extracted through the reverse generator.

5. The image reversible steganography device according to claim 4, characterized in that: The training module designs the generator of the GAN model based on the reversible neural network model.

6. The image reversible steganography device according to claim 5, characterized in that: The processing module is used to invert the output and input ends of the generator of the trained GAN model to obtain an extractor; the generator and the extractor are based on a bidirectional mapping to perform information steganography and message extraction between the sender and the receiver through a secret channel.

7. An image reversible steganography system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the image reversible steganography method according to any one of claims 1 to 3 is executed.

8. A storage medium, characterized in that: The storage medium stores a computer program, which executes the image reversible steganography method according to any one of claims 1 to 3 when running.

Images