A training method for steganalyzer based on automatic virtual data augmentation

Abstract

The present invention discloses a training method for a steganalyzer based on automatic virtual data enhancement, comprising: performing image data enhancement on a batch sample image set using an enhancement network to obtain a probability distribution of enhanced noise; sampling the probability distribution of the enhanced noise using a differentiable sampler to obtain enhanced noise and an augmented image set; processing the batch sample image set and the augmented image set respectively using a steganalyzer, calculating a first loss value using a loss function of the steganalyzer, and updating the parameters of the steganalyzer by performing gradient back propagation using a gradient descent method; processing the augmented image set respectively using the steganalyzer and a category retainer after parameter update, and processing the enhanced noise using a stabilizer; calculating a second loss value using the loss function of the enhancement network, and updating the parameters of the enhancement network by performing gradient back propagation using a gradient descent method; and iteratively performing the above operations to obtain a trained steganalyzer.

Description

Technical Field

[0001] The present invention relates to the technical field of steganalysis, and in particular to a training method for a steganalyzer based on automatic virtual data enhancement, a recognition method for a steganalyzer based on automatic virtual data enhancement, an electronic device, and a storage medium. Background Art

[0002] Steganalysis is a defensive technique against steganography. Early steganalysis methods relied on artificial heuristics to extract features and then performed classification based on machine learning classifiers. In fact, digital image steganalysis can be viewed as a binary classification problem. Inspired by the tremendous success of deep learning in classification tasks in computer vision, the focus of steganalysis has shifted to deep learning. In recent years, the performance of deep learning steganalysis has rapidly improved, significantly surpassing that of steganalysis based on handcrafted features. However, deep learning models are data-driven. In the real world, steganalysts may not have access to sufficient data. When data is limited, models are prone to overfitting. Data augmentation is an effective method to mitigate overfitting and improve the performance of deep learning models.

[0003] Many data augmentation methods for steganalysis have been proposed. These methods can be divided into carrier augmentation and virtual augmentation. Carrier augmentation involves introducing or creating more carrier images and then using the same steganography algorithm to generate corresponding secret images, forming more carrier-secret pairs. This is performed before training the steganalyzer. Representative methods for carrier augmentation include adding datasets with similar distributions, pixels-off, and DPAA. Virtual augmentation methods perform some transformation on the carrier and secret images during training. These methods include rotation and flipping, BitMix, and CutOut. They can be flexibly inserted into the end-to-end training process.

[0004] However, existing deep learning-based virtual enhancement methods for steganalysis are all heuristically designed and rely on expert knowledge. In contrast, many automatic data augmentation methods have been proposed in the field of computer vision, but they cannot be directly applied to steganalysis because the two tasks focus on different things. Computer vision (CV) tasks focus on semantic information, while steganalysis focuses on subtle stegographic signals primarily distributed in image texture regions. These automatic enhancement methods rarely alter semantic information but can easily eliminate stegographic signals. Summary of the Invention

[0005] In view of the above problems, the present invention provides a training method for a steganalyzer based on automatic virtual data enhancement, an identification method for a steganalyzer based on automatic virtual data enhancement, an electronic device, and a storage medium, in order to solve at least one of the above problems.

[0006] According to a first aspect of the present invention, a method for training a steganalyzer based on automatic virtual data augmentation is provided, comprising:

[0007] Operation 1: randomly sampling the training sample image set according to the preset sampling number to obtain a batch sample image set;

[0008] Operation 2: Use the enhancement network to perform image data enhancement on the batch sample image set to obtain the probability distribution of enhanced noise;

[0009] Operation three, using a differentiable sampler to sample the probability distribution of the enhanced noise to obtain the enhanced noise, and according to the enhanced noise, the batch sample image set and the preset hyperparameters of the enhanced noise amplitude, an augmented image set is obtained;

[0010] Operation 4: using a steganalyzer to process the batch sample image set and the augmented image set respectively to obtain a first processing result and a second processing result, wherein the steganalyzer is constructed based on a deep graph neural network;

[0011] Operation 5: Calculate a first loss value of the first processing result and the second processing result using the loss function of the steganalyzer, and update the parameters of the steganalyzer by performing gradient backpropagation using a gradient descent method based on the first loss value;

[0012] Operation six, using the steganalyzer and the class preserver with updated parameters to process the augmented image set respectively to obtain a third processing result, and using the stabilizer to process the enhanced noise to obtain a fourth processing result, wherein the class preserver and the steganalyzer have the same network structure but different parameters;

[0013] Operation seven, using the loss function of the enhanced network to calculate a second loss value of the third processing result and the fourth processing result, and based on the second loss value, performing gradient back propagation by a gradient descent method to update the parameters of the enhanced network;

[0014] Iterate operations 1 to 7 until the parameters of the steganalyzer meet the preset convergence conditions, and obtain a trained steganalyzer.

[0015] According to an embodiment of the present invention, the method of performing image data enhancement on a batch sample image set using an enhancement network to obtain a probability distribution of enhanced noise includes:

[0016] The image data in the batch sample image set is input into the enhanced network for downsampling convolution processing, and the obtained convolution feature vector is used as the input vector of the next round of downsampling convolution processing operation;

[0017] According to the preset number of downsampling times, the input operation and the downsampling convolution processing operation are iteratively performed to obtain multiple convolution feature vectors with different feature dimensions;

[0018] Selecting a convolution feature vector with a first preset feature dimension from a plurality of convolution feature vectors with different feature dimensions and inputting the convolution feature vector into the enhancement network for upsampling deconvolution processing, concatenating the obtained deconvolution feature vector with a convolution feature vector having the same feature dimension as the deconvolution feature vector to obtain a fused feature vector, and using the fused feature vector as the input vector for the next round of upsampling deconvolution processing;

[0019] According to the preset number of downsampling times, the input operation, upsampling deconvolution processing operation and vector splicing operation are iteratively performed to obtain multiple fused feature vectors with different feature dimensions;

[0020] A fused feature vector with a second preset feature dimension is selected from a plurality of fused feature vectors with different feature dimensions, and the selected fused feature vector is processed successively using a saturation activation function and a non-saturation activation function to obtain a probability distribution of enhanced noise.

[0021] According to an embodiment of the present invention, the above-mentioned sampling of the probability distribution of enhanced noise by using a differentiable sampler to obtain enhanced noise, and obtaining an augmented image set according to the enhanced noise, a batch sample image set, and a preset hyperparameter of the enhanced noise amplitude includes:

[0022] According to the Gumbel-Softmax method, the relationship between the enhanced noise probability distribution and the enhanced noise is obtained;

[0023] According to the relationship between the enhanced noise probability distribution and the enhanced noise, a differentiable sampler is used to randomly sample the enhanced noise probability distribution to obtain the enhanced noise;

[0024] The enhanced noise is vector-multiplied by a preset hyperparameter of the enhanced noise amplitude, and a vector operation is performed on the multiplied vector and the image data in the batch sample image set to obtain the augmented image set.

[0025] According to an embodiment of the present invention, the relationship between the enhanced noise probability distribution and the enhanced noise is determined by formula (1):

[0026] (1),

[0027] in, represents enhanced noise, and denote the probabilities of additive noise and subtractive noise respectively, represents the probability of not adding noise, Determined by formula (2):

[0028] (2),

[0029] in, Represents a uniform distribution in the range 0 to 1.

[0030] According to an embodiment of the present invention, the loss function of the stabilizer is determined by formula (3):

[0031] (3),

[0032] in, represents the enhanced noise; is a hyperparameter representing the desired amount of noise enhancement.

[0033] According to an embodiment of the present invention, the loss function of the enhanced network includes a loss function of a steganalyzer, a loss function of a category retainer, and a loss function of a stabilizer.

[0034] According to an embodiment of the present invention, the above-mentioned training method of the steganalyzer based on automatic virtual data enhancement further includes:

[0035] When there are multiple enhancement networks, according to the number of enhancement networks, respectively using the enhancement networks and the differentiable samplers corresponding to the enhancement networks to perform an enhancement noise probability distribution acquisition operation and an enhancement noise acquisition operation to obtain an enhancement noise set;

[0036] During each round of steganalyzer training, the following operations are performed:

[0037] Randomly generate a first random number and a second random number according to the number of enhanced networks;

[0038] Selecting an enhanced network and enhanced noise according to a first random number;

[0039] Selecting a preset hyperparameter of the enhanced noise amplitude according to the second random number;

[0040] The steganalyzer is trained according to the hyperparameters of the selected enhanced network, the selected enhanced noise, and the selected preset enhanced noise amplitude.

[0041] According to a second aspect of the present invention, there is provided a recognition method for a steganalyzer based on automatic virtual data enhancement, comprising:

[0042] Inputting a target image to be identified into a trained steganalyzer to obtain steganalysis determination information of the target image to be identified, wherein the trained steganalyzer is trained according to the above-mentioned training method for the steganalyzer based on automatic virtual data enhancement;

[0043] When the steganalysis determination information of the target image to be identified includes steganalysis information, the target image to be identified is determined to be a secret image.

[0044] According to a third aspect of the present invention, there is provided an electronic device, comprising:

[0045] one or more processors;

[0046] a storage device for storing one or more programs,

[0047] When one or more programs are executed by one or more processors, the one or more processors execute a training method of a steganalyzer based on automatic virtual data enhancement and a recognition method of a steganalyzer based on automatic virtual data enhancement.

[0048] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which executable instructions are stored. When the instructions are executed by a processor, the processor executes a training method for a steganalyzer based on automatic virtual data enhancement and a recognition method for a steganalyzer based on automatic virtual data enhancement.

[0049] The above-mentioned steganalyzer training method based on automatic virtual data enhancement provided by the present invention improves the generalization of the trained steganalyzer through rich training sample image data; the steganalyzer obtained by the above-mentioned training method greatly improves the efficiency and accuracy in identifying image steganographic information. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 2 is a schematic diagram of a principle framework of a steganalyzer training method based on automatic virtual data enhancement according to an embodiment of the present invention;

[0051] Figure 2 is a flow chart of a method for training a steganalyzer based on automatic virtual data enhancement according to an embodiment of the present invention;

[0052] Figure 3 is a flow chart of obtaining a probability distribution of enhanced noise according to an embodiment of the present invention;

[0053] Figure 4 is a schematic diagram of a process for enhancing an output image of a network according to an embodiment of the present invention;

[0054] Figure 5 is a schematic diagram of an enhanced image obtained by using an enhancement network according to an embodiment of the present invention.

[0055] Figure 6 is a flowchart of obtaining an augmented image set according to an embodiment of the present invention;

[0056] Figure 7The block diagram schematically shows an electronic device suitable for implementing a training method of a steganalyzer based on automatic virtual data enhancement and a recognition method of a steganalyzer based on automatic virtual data enhancement according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0058] The purpose of this paper is to provide an automatic virtual data augmentation scheme for deep image steganalysis. Based on a deep augmentation network and using an adversarial augmentation training strategy, the scheme automatically augments data during training to improve the generalization ability of deep learning steganalyzers.

[0059] Figure 1 3 is a schematic diagram of the principle framework of a steganalyzer training method based on automatic virtual data enhancement according to an embodiment of the present invention.

[0060] For steganalysis, existing virtual data augmentation methods are all heuristic. In contrast, many automatic data augmentation methods have been proposed in computer vision. However, due to their different task focus, these methods cannot be directly applied to steganalysis. Inspired by these methods and taking into account the characteristics of the steganalysis task, this paper proposes an automatic virtual data augmentation method for steganalysis, called AAS.

[0061] like Figure 1 As shown in the figure, the principle framework of AAS consists of an enhancement network, a differentiable sampler, a "category preservation" module, a "stabilizer" loss function, and a steganalyzer. AAS achieves data enhancement by intelligently adding noise to the original carrier image and the steganalyzer image. This enhancement method is implemented by the enhancement network and the differentiable sampler. Specifically, the enhancement network generates a probability distribution of the enhanced noise. , the differentiable sampler samples the probability distribution to obtain enhanced noise . The entire enhancement process is differentiable, so the enhancement network can be optimized using gradient descent. The enhancement network is adversarially trained with the steganalyzer to generate diverse enhanced data. At the same time, the "category preservation" module prevents the enhancement network from causing excessive changes to the original data distribution by maintaining the predicted classes of the enhanced carrier image and the secret image. Since the enhancement network is adversarially trained with the steganalyzer, the training process is unstable. For this reason, the present invention introduces a "stabilizer" loss function. The "stabilizer" loss function constrains the amount of enhancement noise, making the training process more stable. In AAS, there is no need to make any changes to the structure of the steganalyzer, and AAS is compatible with various spatial domain image steganalysis networks.

[0062] Figure 2 4 is a flowchart of a method for training a steganalyzer based on automatic virtual data enhancement according to an embodiment of the present invention.

[0063] like Figure 2 As shown, the above-mentioned training method of a steganalyzer based on automatic virtual data enhancement includes operations S210 to S280.

[0064] In operation S210 , the training sample image set is randomly sampled according to a preset sampling quantity to obtain a batch sample image set.

[0065] The training sample image set includes image data and label data corresponding to the image data. The trained steganalyzer obtained by the steganalyzer training method provided by the present invention is primarily used to perform steganalysis on image data to determine whether the image data contains steganographic information, that is, to determine whether the image data is an ordinary image or a secret image.

[0066] In operation S220, the enhancement network is used to perform image data enhancement on the batch sample image set to obtain a probability distribution of enhanced noise.

[0067] Figure 3 is a flow chart of obtaining a probability distribution of enhanced noise according to an embodiment of the present invention.

[0068] like Figure 3 As shown, the above-mentioned process of performing image data enhancement on a batch sample image set using an enhancement network to obtain a probability distribution of enhanced noise includes operations S310 to S350.

[0069] In operation S310, the image data in the batch sample image set is input into the enhancement network for downsampling convolution processing, and the obtained convolution feature vector is used as the input vector of the next round of downsampling convolution processing operation.

[0070] In operation S320, the input operation and the downsampling convolution processing operation are iteratively performed according to a preset number of downsampling times to obtain a plurality of convolution feature vectors with different feature dimensions.

[0071] In operation S330, a convolution feature vector with a first preset feature dimension is selected from a plurality of convolution feature vectors with different feature dimensions and input into the enhanced network for upsampling deconvolution processing, and the obtained deconvolution feature vector is vector-concatenated with a convolution feature vector with the same feature dimension as the deconvolution feature vector to obtain a fused feature vector, and the fused feature vector is used as the input vector for the next round of upsampling deconvolution processing operation.

[0072] In operation S340 , the input operation, the upsampling deconvolution processing operation, and the vector concatenation operation are iteratively performed according to a preset number of downsampling times to obtain a plurality of fused feature vectors with different feature dimensions.

[0073] In operation S350, a fused feature vector having a second preset feature dimension is selected from a plurality of fused feature vectors having different feature dimensions, and the selected fused feature vector is processed successively using a saturation activation function and a non-saturation activation function to obtain a probability distribution of enhanced noise.

[0074] The following further describes in detail the enhancement network provided by the present invention and the process of image data processing by the enhancement network in conjunction with Table 1.

[0075] In AAS, the original data is input into the augmented network In the probability distribution of generated augmented noise The process can be expressed as formula (4):

[0076] (4),

[0077] The structure of the enhanced network is shown in Table 1. The enhanced network adopts the U-Net structure. The U-Net consists of 8 pairs of convolution and deconvolution modules. The convolution module contains a stride of 2 and a kernel size of The deconvolution module consists of a convolution layer with a stride of 2 and a kernel size of The deconvolution layer, a batch normalization layer and an activation function ReLU are used to short-circuit the feature maps of the same size and then concatenate them into a new tensor to be input to the next deconvolution module. Finally, the probability of enhancing the noise is From formula (5), we can get:

[0078] (5),

[0079] in, is the output of the last deconvolution module.

[0080] Table 1 Structure of the enhanced network

[0081]

[0082] Figure 4 2 is a schematic diagram of a process for enhancing an output image of a network according to an embodiment of the present invention.

[0083] like Figure 4 As shown, Figure 4 (a) is the original image, which is the input of the enhancement network. Figure 4 (b)- Figure 4 (e) is the output of the enhanced network, which represents the probability distribution of the enhanced noise. Figure 4 (b) Output of the enhanced network after full AAS training. Figure 4 (c)- Figure 4 (e) Output of the enhanced network with adversarial training removed, the “category preservation” module removed, and the “stabilizer” loss function removed, respectively. Figure 4 (b)- Figure 4 In (e), the higher the brightness, the higher the probability.

[0084] Figure 5 is a schematic diagram of an enhanced image obtained by using an enhancement network according to an embodiment of the present invention.

[0085] like Figure 5 As shown, Figure 5 (a) is "1.pgm" from BOSSBase. Figure 5 (b) Augmented image generated by full AAS. Figure 5 (c)- Figure 5 (e) Enhanced images generated by AAS after removing adversarial training, removing the “category preservation” module, and removing the “stabilizer” loss function, respectively.

[0086] In operation S230, a differentiable sampler is used to sample the probability distribution of the enhanced noise to obtain the enhanced noise, and an augmented image set is obtained according to the enhanced noise, the batch sample image set, and a preset hyperparameter of the enhanced noise amplitude.

[0087] Figure 6 is a flowchart of obtaining an augmented image set according to an embodiment of the present invention.

[0088] like Figure 6 As shown, the above-mentioned sampling of the probability distribution of enhanced noise by using a differentiable sampler to obtain enhanced noise, and obtaining an augmented image set according to the enhanced noise, a batch sample image set and a preset hyperparameter of the enhanced noise amplitude includes operations S610 to S630.

[0089] In operation S610, a relationship between enhanced noise probability distribution and enhanced noise is obtained according to a Gumbel-Softmax method.

[0090] Gumbel-Softmax represents a reparameterized method.

[0091] In operation S620, based on the relationship between the enhanced noise probability distribution and the enhanced noise, a differentiable sampler is used to randomly sample the enhanced noise probability distribution to obtain the enhanced noise.

[0092] In operation S630 , vector multiplication is performed on the enhanced noise and a preset hyperparameter of the enhanced noise amplitude, and a vector operation is performed on the multiplied vector and the image data in the batch sample image set to obtain an augmented image set.

[0093] In AAS, the sampler Probability distribution Sampling is performed to obtain augmented noise The process can be expressed by formula (6):

[0094] (6),

[0095] Sampling is a non-differentiable process with random properties, but this method requires all calculations to be differentiable in order to optimize the enhanced network. To this end, the present invention uses the Gumbel-softmax technique to propose a differentiable sampler.

[0096] According to an embodiment of the present invention, the relationship between the enhanced noise probability distribution and the enhanced noise is determined by formula (1):

[0097] (1),

[0098] in, represents enhanced noise, and denote the probabilities of additive noise and subtractive noise respectively, represents the probability of not adding noise, Determined by formula (2):

[0099] (2),

[0100] in, Represents a uniform distribution in the range 0 to 1.

[0101] Using the Gumbel-softmax technique, enhanced noise is obtained The explicit expression of , the corresponding probability is In this way, the present invention converts the randomness required in the sampling process into , so that when back propagating, we can get Since the carrier image and the dense image are very close, it is difficult for the enhancement network to distinguish the subtle differences between them. Therefore, the output of the enhanced network is obtained by inputting the carrier image or the dense image into the enhanced network. Almost the same. Although The two images are very similar, but due to the randomness of the sampling, the augmented noise is not identical. That is, the distribution pattern of the augmented noise is similar, but the actual location of the augmented noise is different. The randomness of the sampling further increases the diversity of the data.

[0102] In operation S240 , the batch sample image set and the augmented image set are processed separately using a steganalyzer to obtain a first processing result and a second processing result, wherein the steganalyzer is constructed based on a deep graph neural network.

[0103] In operation S250, a loss function of the steganalyzer is used to calculate loss values of the first processing result and the second processing result, and according to the loss values, a gradient back propagation method is performed through a gradient descent method to update the parameters of the steganalyzer.

[0104] In operation S260, the augmented image set is processed respectively using the steganalyzer and the category retainer with updated parameters to obtain a third processing result, and the enhanced noise is processed using the stabilizer to obtain a fourth processing result, wherein the category retainer and the steganalyzer have the same network structure but different parameters.

[0105] In operation S270, loss values of the third processing result and the fourth processing result are calculated using the loss function of the enhanced network, and according to the loss values, gradient back propagation is performed by a gradient descent method to update the parameters of the enhanced network.

[0106] In operation S280 , operations S210 to S270 are iterated until the parameters of the steganalyzer meet a preset convergence condition, thereby obtaining a trained steganalyzer.

[0107] The above-mentioned steganalyzer training method based on automatic virtual data enhancement provided by the present invention improves the generalization of the trained steganalyzer through rich training sample image data; the steganalyzer obtained by the above-mentioned training method greatly improves the efficiency and accuracy in identifying image steganographic information.

[0108] According to an embodiment of the present invention, the loss function of the stabilizer is determined by formula (3):

[0109] (3),

[0110] is a hyperparameter representing the desired amount of noise enhancement.

[0111] Since the training process of AAS is similar to that of generative adversarial networks and inherits its training instability, without a "stabilizer" loss function, the enhanced noise may converge to different values during different training processes. If the amount of enhanced noise converges to a large value, it will erase a large amount of steganographic signals. Conversely, converging to a smaller value will not achieve the effect of increasing data diversity. Therefore, it is necessary to design a "stabilizer" loss function involved in the present invention. to limit the amount of enhanced noise.

[0112] According to an embodiment of the present invention, the loss function of the enhanced network includes a loss function of a steganalyzer, a loss function of a category retainer, and a loss function of a stabilizer.

[0113] The above “category preservation” module is used to maintain the predicted classes of the enhanced carrier image and the dense image to prevent the enhanced network from changing the original data distribution too much. and steganalyzers They have the same structure but different parameters. The “category preserving” module is pre-trained and its parameters are fixed throughout the training process. In contrast, the parameters θ of the steganalyzer are updated during the training process.

[0114] The training process of AAS mainly involves adversarial training between the enhanced network and the steganalyzer. At the same time, it is constrained by the "category preservation" module and the "stabilizer" loss function. Based on the above principles and constraints, the loss function of the enhanced network is As shown in formula (7):

[0115] (7),

[0116] in, represents the set of training sample images, is the cross entropy function, Represents a steganalyzer. For ease of presentation, the data labels are omitted. To augment the data, To augment the noise. 、 and The relationship is shown in formula (8) to formula (10):

[0117] (8),

[0118] (9),

[0119] (10),

[0120] in is a hyperparameter representing the magnitude of the enhanced noise.

[0121] When training the enhanced network, the steganalyzer is fixed, and vice versa. The two networks are updated alternately. When training the steganalyzer, the original data is and augmented data Enter the steganalyzer.

[0122] The loss function of the steganalyzer is shown in formula (11):

[0123] (11),

[0124] The overall optimization objective of AAS is shown in formula (12):

[0125] (12),

[0126] in, Represents a category retainer.

[0127] According to an embodiment of the present invention, the above-mentioned training method of the steganalyzer based on automatic virtual data enhancement further includes:

[0128] When there are multiple enhancement networks, according to the number of enhancement networks, respectively using the enhancement networks and the differentiable samplers corresponding to the enhancement networks to perform an enhancement noise probability distribution acquisition operation and an enhancement noise acquisition operation to obtain an enhancement noise set;

[0129] During each round of steganalyzer training, the following operations are performed:

[0130] Randomly generate a first random number and a second random number according to the number of enhanced networks;

[0131] Selecting an enhanced network and enhanced noise according to a first random number;

[0132] Selecting a preset hyperparameter of the enhanced noise amplitude according to the second random number;

[0133] The steganalyzer is trained according to the hyperparameters of the selected enhanced network, the selected enhanced noise, and the selected preset enhanced noise amplitude.

[0134] The steganalyzer can be combined with multiple augmentation networks to further improve the performance of the steganalyzer (called AAS-M). In AAS-M, different augmentation networks add different patterns of augmentation noise, which further enriches the data.

[0135] According to a second aspect of the present invention, there is provided a recognition method for a steganalyzer based on automatic virtual data enhancement, comprising:

[0136] Inputting a target image to be identified into a trained steganalyzer to obtain steganalysis determination information of the target image to be identified, wherein the trained steganalyzer is trained according to the above-mentioned training method for the steganalyzer based on automatic virtual data enhancement;

[0137] When the steganalysis determination information of the target image to be identified includes steganalysis information, the target image to be identified is determined to be a secret image.

[0138] In order to better illustrate the actual effect of the steganalyzer trained by the above-mentioned steganalyzer training method provided by the present invention, the present invention sets up a security evaluation experiment for AAS and AAS-M, where AAS means that during the steganalyzer training stage, it has only one enhanced network, while AASM means that during the steganalyzer training stage, it has multiple enhanced networks.

[0139] First, we evaluate the security of AAS. We selected three steganalyzers (CovNet, SRNet, and YedNet) for evaluation. The steganalyzers were trained using four enhancement methods: Baseline (rotation and flipping), BitMix, CutOut, and AAS. We tested three steganography algorithms (S-UNIWRAD, HILL, and MiPOD) at four payload sizes (0.1 bpp, 0.2 bpp, 0.3 bpp, and 0.4 bpp). The accuracy on the test set is shown in Table 2.

[0140] Table 2 Performance comparison of AAS and other existing methods

[0141]

[0142] From Table 2, we can conclude that: (1) AAS outperforms previous methods in all test cases. The average improvements of AAS, BitMix, and CutOut over the baseline are 3.03%, 1.24%, and 0.77%, respectively. The method proposed in this paper (AAS) has obvious advantages; (2) The best case of AAS is to detect HILL with 0.2 bpp using YedNet. The improvement compared with the baseline is 9.04%. The worst case of AAS is to detect HILL with 0.1 bpp using SRNet. The improvement compared with the baseline is 0.65%. Even in the worst case, the method of this paper is still better than the previous methods; (3) The average improvement rate of AAS over the baseline on CovNet, YedNet, and SRNet is 2.14%, 4.04%, and 2.89%, respectively. The improvement of AAS is opposite to the performance of the base network. This indicates that the more sophisticated the steganalysis network itself is, the less room there is for enhancing its performance. (4) When detecting S-UNIWARD, HILL, and MiPOD, AAS achieves an average improvement of 3.05%, 3.27%, and 2.75% over the baseline, respectively. AAS also achieves similar improvements when detecting different steganographic algorithms. (5) At 0.1 bpp, 0.2 bpp, 0.3 bpp, and 0.4 bpp payloads, AAS achieves an average improvement of 2.43%, 4.13%, 2.46%, and 3.09% over the baseline, respectively. AAS performs best at a payload of 0.2 bpp.

[0143] Next, we conduct an AASM security evaluation. AASM jointly trains a steganalyzer with multiple augmentation networks. The hyperparameters are set as follows: five augmentation networks are used. The noise number list is set to [100, 256, 400, 700, 1000], and the noise amplitude list is set to [16, 32, 64, 128, 255]. The above experiments have shown that AAS achieves roughly the same improvement in detecting S-UNIWARD, HILL, and MiPOD. Therefore, this article only explores the effectiveness of the three steganalyzers in detecting S-UNIWARD at 0.1 bpp, 0.2 bpp, 0.3 bpp, and 0.4 bpp payloads. The accuracy on the test set is shown in Table 3.

[0144] Table 3 Performance of AAS-M

[0145]

[0146] Finally, a comparison between AAS and AASM is presented. The advantages of AASM over AAS are mainly reflected in the following aspects: (1) AAS-M outperforms AAS, with an average improvement of 0.61%; (2) Compared with AAS, AAS-M improves CovNet, YedNet, and SRNet by 0.98%, 0.40%, and 0.45% on average, respectively, with CovNet achieving the largest improvement; (3) At 0.1 bpp, 0.2 bpp, 0.3 bpp, and 0.4 bpp, AAS-M improves AAS by 0.53%, 0.47%, 0.84%, and 0.60% on average, with the largest improvement at 0.3 bpp.

[0147] Figure 7 The block diagram schematically shows an electronic device suitable for implementing a training method of a steganalyzer based on automatic virtual data enhancement and a recognition method of a steganalyzer based on automatic virtual data enhancement according to an embodiment of the present invention.

[0148] like Figure 7 As shown, an electronic device 700 according to an embodiment of the present invention includes a processor 701, which can perform various appropriate actions and processes based on programs stored in a read-only memory (ROM) 702 or programs loaded from a storage unit 708 into a random access memory (RAM) 903. The processor 701 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or related chipsets and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 701 may also include onboard memory for caching purposes. The processor 701 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0149] Various programs and data required for the operation of the electronic device 700 are stored in the RAM 703. The processor 701, ROM 702, and RAM 703 are connected to each other via a bus 704. The processor 701 performs various operations according to the method flow of the embodiment of the present invention by executing the programs in the ROM 702 and / or RAM 703. It should be noted that the programs may also be stored in one or more memories other than the ROM 702 and RAM 703. The processor 701 may also perform various operations according to the method flow of the embodiment of the present invention by executing the programs stored in the one or more memories.

[0150] According to an embodiment of the present invention, electronic device 700 may further include an input / output (I / O) interface 905, which is also connected to bus 704. Electronic device 700 may also include one or more of the following components connected to I / O interface 705: an input section 706 including a keyboard, mouse, etc.; an output section 707 including devices such as a cathode ray tube (CRT), liquid crystal display (LCD), and speakers; a storage section 708 including a hard disk; and a communication section 709 including a network interface card such as a LAN card or modem. Communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to I / O interface 705 as needed. Removable media 711, such as a magnetic disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed in drive 710 as needed, so that computer programs read from the removable media can be installed into storage section 708 as needed.

[0151] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments, or may exist independently and not incorporated into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0152] According to an embodiment of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, and may include, for example, but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present invention, a computer-readable storage medium may include the ROM 702 and / or RAM 703 described above, and / or one or more memories other than ROM 702 and RAM 703.

[0153] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the above-mentioned module, program segment, or a part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0154] Those skilled in the art will appreciate that various combinations and / or combinations of features described in the various embodiments and / or claims of the present invention may be made, even if such combinations and / or combinations are not explicitly described in the present invention. In particular, various combinations and / or combinations of features described in the various embodiments and / or claims of the present invention may be made, without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0155] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for training a steganalyzer based on automatic virtual data augmentation, comprising: Operation 1: randomly sampling the training sample image set according to the preset sampling number to obtain a batch sample image set; Operation 2: performing image data enhancement on the batch sample image set using an enhancement network to obtain a probability distribution of enhanced noise; Operation three: using a differentiable sampler to sample the probability distribution of the enhanced noise to obtain enhanced noise, and obtaining an augmented image set according to the enhanced noise, the batch sample image set, and a preset hyperparameter of the enhanced noise amplitude; Operation 4: using the steganalyzer to process the batch sample image set and the augmented image set respectively to obtain a first processing result and a second processing result, wherein the steganalyzer is constructed based on a deep graph neural network; Operation 5: Calculate a first loss value of the first processing result and the second processing result using a loss function of the steganalyzer, and update the parameters of the steganalyzer by performing gradient backpropagation using a gradient descent method based on the first loss value; Operation six, using the steganalyzer and the class preserver with updated parameters to process the augmented image set respectively to obtain a third processing result, and using a stabilizer to process the enhanced noise to obtain a fourth processing result, wherein the class preserver and the steganalyzer have the same network structure but different parameters; Operation seven: calculating a second loss value of the third processing result and the fourth processing result using the loss function of the enhanced network, and updating the parameters of the enhanced network by performing gradient back propagation using a gradient descent method based on the second loss value; Iteratively performing operations 1 to 7 until the parameters of the steganalyzer meet a preset convergence condition, thereby obtaining a trained steganalyzer; The step of sampling the probability distribution of the enhanced noise using a differentiable sampler to obtain the enhanced noise, and obtaining the augmented image set according to the enhanced noise, the batch sample image set, and a preset hyperparameter of the enhanced noise amplitude includes: According to the Gumbel-Softmax method, a relationship between the enhanced noise probability distribution and the enhanced noise is obtained; According to the relationship between the enhanced noise probability distribution and the enhanced noise, using the differentiable sampler to randomly sample the enhanced noise probability distribution to obtain the enhanced noise; Performing vector multiplication of the enhanced noise and the preset hyperparameter of the enhanced noise amplitude, and performing vector operation on the multiplied vector and the image data in the batch sample image set to obtain the augmented image set; The relationship between the enhanced noise probability distribution and the enhanced noise is determined by formula (1): （1）， in, represents the enhanced noise, and denote the probabilities of additive noise and subtractive noise respectively, represents the probability of not adding noise, Determined by formula (2): （2）， in, Represents a uniform distribution in the range 0 to 1.

2. The method according to claim 1, wherein The step of performing image data enhancement on the batch sample image set using the enhancement network to obtain a probability distribution of enhanced noise includes: Inputting the image data in the batch sample image set into the enhancement network for downsampling convolution processing, and using the obtained convolution feature vector as the input vector of the next round of downsampling convolution processing operation; According to the preset number of downsampling times, the input operation and the downsampling convolution processing operation are iteratively performed to obtain multiple convolution feature vectors with different feature dimensions; Selecting a convolution feature vector with a first preset feature dimension from the multiple convolution feature vectors with different feature dimensions and inputting it into the enhancement network for upsampling deconvolution processing, concatenating the obtained deconvolution feature vector with a convolution feature vector having the same feature dimension as the deconvolution feature vector to obtain a fused feature vector, and using the fused feature vector as the input vector for the next round of upsampling deconvolution processing; According to the preset number of downsampling times, the input operation, upsampling deconvolution processing operation and vector splicing operation are iteratively performed to obtain multiple fused feature vectors with different feature dimensions; A fused feature vector with a second preset feature dimension is selected from the multiple fused feature vectors with different feature dimensions, and the selected fused feature vector is processed successively using a saturation activation function and a non-saturation activation function to obtain the probability distribution of the enhanced noise.

3. The method according to claim 1, wherein The loss function of the stabilizer is determined by formula (3): （3）， in, represents the enhanced noise; is a hyperparameter representing the desired amount of noise enhancement.

4. The method according to claim 1, wherein The loss function of the enhanced network includes a loss function of a steganalyzer, a loss function of a category retainer, and a loss function of a stabilizer.

5. The method according to claim 1, further comprising: In the case where there are multiple enhancement networks, performing an enhancement noise probability distribution acquisition operation and an enhancement noise acquisition operation using the enhancement networks and the differentiable samplers corresponding to the enhancement networks according to the number of the enhancement networks, to obtain an enhancement noise set; During each round of steganalyzer training, the following operations are performed: Randomly generate a first random number and a second random number according to the number of the enhanced networks; Selecting an enhanced network and enhanced noise according to the first random number; Selecting a preset hyperparameter of enhanced noise amplitude according to the second random number; The steganalyzer is trained according to the hyperparameters of the selected enhanced network, the selected enhanced noise, and the selected preset enhanced noise amplitude.

6. A recognition method for a steganalyzer based on automatic virtual data enhancement, comprising: Inputting a target image to be identified into a trained steganalyzer to obtain steganalysis determination information of the target image to be identified, wherein the trained steganalyzer is trained according to any one of the methods of claims 1 to 5; When the steganographic determination information of the target image to be identified includes steganographic information, the target image to be identified is determined to be a secret image.

7. An electronic device comprising: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors are enabled to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, causes the processor to perform the method according to any one of claims 1 to 6.

Images