A Steganalysis Method and System Based on Multi-Source Domain Adaptation

Abstract

The present invention discloses a steganography detection method and system based on multi-source domain adaptation, which relates to the field of information security technology, and includes: receiving data of multiple source domains and a target domain, inputting the data of the multiple source domains and the target domain into a pre-established common feature extractor, and extracting the feature representations of the source domains and the feature representation of the target domain; inputting the feature representations of the source domains and the feature representation of the target domain into a pre-established feature-level mixing module for weighted mixing to obtain mixed features, obtaining target domain features, and calculating a mixed class loss based on the target domain features and the mixed features; inputting the data of each source domain and the target domain into a pre-established domain feature extractor with unshared weights, and outputting a domain alignment loss, inputting the mixed class loss and the domain alignment loss into a pre-established domain-specific classifier, and outputting a final prediction result, effectively improving the detection accuracy and generalization ability of the steganography detection model under cross-domain conditions.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and specifically to a steganography detection method and system based on multi-source domain adaptation. Background Art

[0002] With the rapid development of digital technology, multimedia carriers are increasingly widely used in information dissemination. Steganography, as a technique for secretly embedding secret information into multimedia carriers, is widely applied in fields such as secret communication, copyright protection, and data security. In particular, image steganography, which uses images widely used in digital media as carriers. However, the concealment of steganography also brings potential security risks. For example, steganographic information may be used for illegal purposes or unauthorized secret transmission. Therefore, steganalysis techniques have emerged, aiming to detect whether there is hidden secret information in multimedia carriers, thus ensuring information security and privacy protection. Early image steganalysis mainly relied on manually designed feature extraction methods to detect whether there was secret information by analyzing the statistical characteristics of digital images. With the continuous progress of steganography techniques, traditional handcrafted feature-based methods have gradually become difficult to cope with complex steganographic algorithms and diverse image contents. In recent years, the booming development of deep learning technology has brought new opportunities to steganalysis. Deep learning-based steganalysis methods significantly improve the detection performance by automatically learning image features. Existing methods have achieved remarkable detection effects in steganalysis tasks. These methods usually assume that the training data and test data come from the same distribution, that is, the source domain and the target domain are consistent. However, in practical applications, the target images may come from data sources different from the training samples, and the labeled data in the target domain is usually unavailable. This inconsistency between the training data and the test data is called carrier source mismatch (CSM). The CSM problem is very common in practical scenarios and may lead to a significant decline in the detection accuracy of steganalysis. Therefore, how to improve the detection accuracy of steganalysis models under cross-domain conditions is the key challenge for applying steganalysis techniques to practical scenarios. In recent years, researchers have conducted extensive explorations on the carrier source mismatch problem in steganalysis. The CSM problem in steganalysis is similar to the unsupervised domain adaptation (UDA) task in the field of computer vision. Therefore, some UDA methods have been introduced into the steganalysis task to alleviate the performance degradation caused by the distribution difference between the source domain and the target domain. J-Net is the first network specifically designed for the CSM problem in deep steganalysis. By integrating the joint maximum mean discrepancy (JMMD) into deep steganalysis, it minimizes the joint maximum average difference between the source and the target, thus alleviating the performance degradation caused by CSM. Based on J-Net, subsequent studies further considered the domain difference and class difference, and achieved sub-domain alignment by minimizing the local maximum mean discrepancy, further reducing the performance degradation of steganalysis caused by different carrier sources. ISNet generates an intermediate domain by designing a local mixing module, enabling the network to gradually adapt to the steganographic noise distribution in the target domain, thereby improving the discriminability on the target domain. Although existing methods have alleviated the CSM problem to a certain extent, most of them rely on data from a single source domain and perform limitedly in complex scenarios with multi-source domain data.In addition, existing methods still have deficiencies in adapting to the steganographic noise distribution in the target domain, especially when the steganographic noise is weak and difficult to distinguish. Summary of the Invention

[0003] To address the deficiencies mentioned in the above background art, the purpose of the present invention is to provide a steganography detection method and system based on multi-source domain adaptation.

[0004] In a first aspect, the purpose of the present invention can be achieved through the following technical solutions: A steganography detection method based on multi-source domain adaptation, the method comprising the following steps:

[0005] Receive data from multiple source domains and a target domain, and input the data from the multiple source domains and the target domain into a pre-established common feature extractor to obtain the feature representations of the source domains and the feature representation of the target domain;

[0006] Input the feature representations of the source domains and the feature representation of the target domain into a pre-established feature-level mixing module for weighted mixing to obtain mixed features, obtain target domain features, and calculate a mixed class loss based on the target domain features and the mixed features;

[0007] Input the data of each source domain and the target domain into a pre-established domain feature extractor with non-shared weights, output a domain alignment loss, and input the mixed class loss and the domain alignment loss into a pre-established domain-specific classifier to output a final prediction result.

[0008] In combination with the first aspect, in certain implementation manners of the first aspect, the method further comprises: The data sources of the data from the multiple source domains and the target domain are: labeled data from N source domains , where and respectively represent the data of the domain and their corresponding labels, as well as the unlabeled data of the target domain . .

[0009] In combination with the first aspect, in certain implementation manners of the first aspect, the method further comprises: The process of inputting the data from the multiple source domains and the target domain into a pre-established common feature extractor:

[0010] Pass through a common feature extractor G, which maps the data from the multiple source domains and the target domain from the original feature space to the common feature space;

[0011] The common feature extractor is composed of the first three types of convolutional layers designed in SRNet. For a batch of data from the source domain and the data of the target domain , the feature representations obtained after passing through the common feature extractor ​ and The formula for

[0012]

[0013] The feature representation obtained from each source domain will all be paired up two by two and input into the feature-level mixing module to additionally generate mixed features.

[0014] Combined with the first aspect, in some implementations of the first aspect, the method further includes: the input feature-level mixing module receives the output features from the common feature extractor and , and then calculates the mixing ratio of the target domain according to the current training epoch . The update of the mixing ratio follows the following formula:

[0015]

[0016] where are respectively the maximum values of the mixing ratio, is the parameter controlling the exponential growth rate, and current_epoch and total_epoch respectively represent the current training epoch and the total number of training epochs;

[0017] According to the calculated mixing ratio , the source domain features and the target domain features are weighted and summed according to the mixing ratio to obtain the mixed features , and its formula is:

[0018]

[0019] where represents the addition operation between matrices, denotes the th sample in this batch, and respectively represent the source domain and target domain output features from the common feature extractor and in the th sample;

[0020] In addition to feature mixing, the labels also need to be mixed. The label of the mixed feature is calculated using the same mixing ratio as the feature. Here is the prediction result of the network on the target domain, is the label of the jth sample:

[0021]

[0022] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: aligning the data of the source domain and the target domain, and the domain feature extractor F i receives features from a batch of data from the source domain and features from the target domain and maps each pair of source domain and target domain to a specific feature space, and then executes the first-stage alignment strategy to align the feature distributions of the two domains. The maximum mean discrepancy (MMD) is used for the first-stage alignment to determine whether the two distributions and are the same, based on the observed sample data. MMD defines the difference measure between distributions through the following formula:

[0023]

[0024] where represents the distance measure between two probability distributions and , H is the reproducing kernel Hilbert space (RKHS) with the feature kernel k, and are samples from different distributions respectively, represents mapping the original sample to a certain feature map in the RKHS, represents sampling from the distribution and calculating the expected value of the function , and represents sampling from the distribution and calculating the expected value of the function ; The feature kernel k represents , , represents the inner product of vectors. When and only when , and have the same distribution. MMD estimates the difference between distributions by comparing the squared distance of the empirical kernel mean embeddings. Combining with the input of the domain feature extractor layer, the formula is as follows:

[0025]

[0026] Here is an unbiased estimator of , and H is the reproducing kernel Hilbert space (RKHS) with the feature kernel k; and are the number of samples in a batch, and are the outputs of the source domain and the target domain processed by the common feature extractor and respectively represent the samples therein. An estimate of the difference between each source domain and the target domain is obtained, and according to MMD, the domain alignment loss MMD Loss of this layer is designed as follows:

[0027]

[0028] where F i represents the i-th domain feature extractor, and are the outputs of the source domain and the target domain processed by the common feature extractor. N domain feature extractors with non-shared weights map each pair of source domain and target domain data into a specific feature space, and the obtained domain alignment loss , realizes the alignment of feature distributions by minimizing during the backpropagation process.

[0029] Combined with the first aspect, in some implementation manners of the first aspect, the method further includes: inputting the processed features into a pre-established domain-specific classifier to obtain the prediction outputs of source, target, and mixed features:

[0030] Based on the alignment strategy of loss, measure the prediction differences of different classifiers on the target domain data, The form of the loss function is as follows:

[0031]

[0032] where and respectively represent the prediction probabilities of the classifiers corresponding to the source domain and branches for the -th target domain sample. By minimizing the loss, is the number of samples in a batch; finally, the mean of the prediction results of all classifiers on the target domain data is used as the final prediction result for output.

[0033] In a second aspect, to achieve the above object, the present invention discloses a steganography detection system based on multi-source domain adaptation, including:

[0034] A feature extraction module, configured to receive multiple source domain and target domain data, input the multiple source domain and target domain data into a pre-established common feature extractor, and extract the feature representations of the source domain and the feature representations of the target domain;

[0035] A feature mixing module, configured to pair and input the feature representations of the source domain and the feature representations of the target domain into a pre-established feature-level mixing module for weighted mixing to obtain mixed features;

[0036] A feature prediction module, which is configured to input the mixed features, the paired feature representations of the source domain and the target domain into a domain feature extractor to obtain processed features, and input the processed features into a pre-established domain-specific classifier, and output the prediction outputs of the source, target, and mixed features.

[0037] In another aspect of the present invention, in order to achieve the above object, a terminal device is disclosed, which includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores a computer program capable of running on the processor. When the processor loads and executes the computer program, it adopts a steganography detection method based on multi-source domain adaptation as described above.

[0038] In still another aspect of the present invention, in order to achieve the above object, a computer-readable storage medium is disclosed. The computer-readable storage medium stores a computer program. When the computer program is loaded and executed by a processor, it adopts a steganography detection method based on multi-source domain adaptation as described above.

[0039] Advantages of the present invention:

[0040] The present invention aims to solve the problem of the decline in detection performance caused by different sources of the sample datasets to be detected and the lack of label information in the actual application of image steganalysis. Through the multi-source two-stage alignment strategy and the gradually changing feature-level mixing module, the detection accuracy and generalization ability of the steganography detection model under cross-domain conditions are effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings;

[0042] Figure 1 It is a schematic flowchart of the method of the present invention;

[0043] Figure 2 It is a schematic diagram of the overall structure of the steganography detection network based on multi-source domain adaptation of the present invention;

[0044] Figure 3 It is a schematic diagram of the structure of the progressive feature fusion module of the present invention;

[0045] Figure 4 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0047] Embodiment 1:

[0048] As Figure 1 shown, a steganography detection method based on multi-source domain adaptation, the method includes the following steps:

[0049] S101: Receive data of multiple source domains and a target domain, and input the data of the multiple source domains and the target domain into a pre-established common feature extractor to obtain the feature representations of the source domains and the feature representation of the target domain;

[0050] As Figure 2 shown, Source Domain 1 to Source Domain N respectively represent labeled samples of N different domains for domain adaptation and network training. The target domain refers to unlabeled samples from the target domain. The common feature extractor is a convolutional neural network composed of the first three convolutional layers of SRNet, which is used to initially extract the features of the samples. The feature-level mixing module in the figure is specifically as Figure 3 shown, which is used to perform weighted mixing on each pair of source and target samples to generate mixed features. The domain-specific extractors in the figure are all composed of the last convolutional layer of SRNet, but are independent of each other and do not share parameters. This extractor is used to further perform convolutional processing on the features and calculate the feature alignment loss using the processed features. The domain-specific classifiers in the figure are composed of a linear classification layer, and each classifier is independent of each other and does not share parameters. It is used to predict the feature category and provide the prediction output for the calculation of the cross-entropy loss. The average in the figure refers to the operation of taking the mean of the prediction outputs of each subnet classifier and using this mean as the prediction output of the sample. The domain alignment loss in the figure is the MMD Loss mentioned above, which refers to the domain alignment loss calculated by the maximum mean discrepancy. The classifier loss refers to the classification alignment loss calculated by the L1 norm for the prediction outputs of different classifiers. The average in the figure refers to taking the average of the prediction outputs of different domain-specific classifiers as the final prediction output of the network. The carrier image and the stego image in the figure respectively represent the classification results of the network. The cross-entropy loss in the figure refers to the classification loss calculated by combining the classification result of the network with the true label using the cross-entropy loss function.

[0051] As Figure 3 shown, this figure takes the first branch of the network as an example to represent the process of feature mixing. Among them, G represents the Figure 2 common feature extractor in this and respectively represent the features of the samples from the first source domain and the target domain after being processed by the common feature extractor. Here and respectively represent the mixing weights of the target domain features and the mixing weights of the source domain features. represents the features after being mixed by weights. H1 represents the domain-specific extractor of the first branch network.

[0052] Assume that the present invention has labeled data from N source domains , where and respectively represent the data of domain and its corresponding label, as well as the unlabeled data of the target domain . After these data are input into the network, they first pass through a common feature extractor G, which maps the data from the original feature space to the common feature space. . This common feature extractor of the present invention is composed of the first three types of convolutional layers designed in SRNet. For a batch of data from the source domain

[0053] , and the data of the target domain , the feature representations obtained after passing through the common feature extractor and are given by the formula:

[0054]

[0055] The feature representation obtained for each source domain will be paired with and input into the feature-level mixing module in pairs to generate an additional mixed feature.

[0056] S102: Pair the feature representation of the source domain and the feature representation of the target domain and input them into the pre-established feature-level mixing module for weighted mixing to obtain a mixed feature;

[0057] For the progressive feature fusion module as shown in Figure 3 , fuse the feature distribution information of the samples from the source domain and the target domain to enable the network to more directly learn the steganographic traces of the target domain:

[0058] As shown in Figure 3 , for the source domain and the target For example, the core idea of the feature-level mixing module is to weight and mix the features from the source domain and the target domain in a certain proportion. This proportion changes gradually during the training process. The mixing proportion is adjusted in an exponential growth manner in each training round, and the features of the target domain gradually occupy a larger proportion in the mixed features, enabling the network to gradually transition to the feature distribution of the target domain during training. Specifically, the domain feature-level mixing module receives the output features from the common feature extractor and , and then the present invention calculates the mixing proportion of the target domain according to the current training round . The update of the mixing proportion follows the following formula:

[0059]

[0060] are the maximum values of the mixing proportion, respectively, and the value is set to 0.9. is the parameter that controls the exponential growth rate. current_epoch and total_epoch represent the current training round and the total number of training rounds, respectively.

[0061] According to the calculated mixing proportion , the source domain features and the target domain features are weighted and summed according to this proportion to obtain the mixed features , and its formula is:

[0062]

[0063] represents the addition operation between matrices, indicating weighted fusion at the feature level. represents the rd sample in this batch. and respectively represent the source domain and target domain output features from the common feature extractor and in the rd sample. In addition to feature mixing, the labels also need to be mixed accordingly. The label of the mixed features is calculated using the same mixing proportion as the features:

[0064]

[0065] Since the samples in the target domain are unlabeled, here the present invention uses the prediction result of the network on the target domain as the label of the target domain samples. The target domain features and the mixed features are sent into the subsequent network together and the cross-entropy loss is calculated according to the mixed labels As the classification loss in the hybrid domain. Among them is the predicted output of the network for the hybrid features, is the number of samples within a batch.

[0066]

[0067] Since there are significant distribution differences between different carrier sources, and it is difficult to align data from multiple domains simultaneously, which is prone to negative transfer. Therefore, this network designs N domain feature extractors for N source-target pairs, aligning each source domain with the target domain respectively. The domain feature extractor F i receives the features from a batch of data from the source domain and the features from the target domain and maps each pair of source domain and target domain into a specific feature space. Subsequently, it executes the first-stage alignment strategy to align the feature distributions of the two domains. The present invention uses the Maximum Mean Discrepancy (MMD) for the first-stage alignment. The Maximum Mean Discrepancy (MMD) is a kernel-based two-sample test method for judging whether two distributions and and are the same, based on the observed sample data. MMD defines the difference metric between distributions through the following formula:

[0068]

[0069] where, represents the distance metric between two probability distributions and and and are samples from different distributions respectively, represents mapping the original sample to a certain feature map in the RKHS, represents sampling from the distribution and calculating the expected value of the function and the expected value of the function represents sampling from the distribution and and calculating the expected value of the function The feature kernel k represents here represents the inner product of vectors. Theoretically, when and only when then and have the same distribution. In practical applications, MMD estimates the difference between distributions by comparing the squared distances of the empirical kernel mean embeddings. Combining with the input of the domain feature extractor layer, the formula is as follows:

[0070]

[0071] Here is an unbiased estimator of, where H is the Reproducing Kernel Hilbert Space (RKHS) with characteristic kernel k. and is the number of samples in a batch, and are the outputs of the source domain and the target domain processed by the common feature extractor. and represent the samples therein respectively. The estimation of the difference between each source domain and the target domain is obtained according to this formula. According to MMD, the domain alignment loss MMD Loss of this layer is designed as:

[0072]

[0073] where Fi represents the i-th domain feature extractor, and are the outputs of the source domain and the target domain processed by the common feature extractor. S103: Input the mixed features, the feature representations of the paired source domain, and the feature representations of the target domain into the domain feature extractor to obtain the processed features, and input the processed features into the pre-established domain-specific classifier to obtain the predicted outputs of the source, target, and mixed features.

[0074] N domain feature extractors without shared weights map each pair of source domain and target domain data into a specific feature space, and the obtained domain alignment loss . Feature distribution alignment is achieved by minimizing in the backpropagation process to learn feature representations that are not affected by the domain. The present invention uses the last convolutional layer of SRNet as the domain feature extractor.

[0075] After that, these features will be sent into their respective domain-specific classifiers for classification. To reduce the prediction bias caused by classifier differences, the present invention introduces a strategy based on prediction probability alignment, which improves the generalization ability of the model to target domain samples by constraining the outputs of different classifiers on the target domain.

[0076] To achieve this goal, the present invention designs an alignment strategy based on loss to measure the prediction differences of different classifiers on the target domain data. The form of this loss function is as follows:

[0077] s

[0078] where and represent the source domain and By minimizing this loss, the prediction probability of the classifier corresponding to the branch for the j-th target domain sample can significantly reduce the output deviation of different classifiers on the target domain, thus making the decision boundary more consistent.

[0079] Finally, the mean of the prediction results of all classifiers on the target domain data is used as the final prediction output. Through the two-stage alignment steganography analysis framework of the present invention, the distribution differences between different source domains and target domains can be well aligned, and more discriminative domain-invariant feature representations can be learned from multiple domains.

[0080] Finally, all losses are weighted and summed as the total loss for network training. By optimizing the network through domain adaptation, the detection accuracy on the target domain can be improved.

[0081] Experiments were conducted using the common steganography algorithm S-UNIWARD with an embedding rate of 0.4 bpp on four different datasets: BOSSbase 1.01, ALASKA#2, MIRFlickr 25k, and ImageNet. BOSSbase contains 10,000 grayscale images in PGM format with a size of 512×512. ALASKA#2 contains 80,000 images in various formats. MIRFlickr 25 contains 25,000 color JPEG images of different sizes. ImageNet contains images of different sizes from the real world in 1000 categories. Subsequently, the present invention uses B, A, M, and I to represent the four datasets. 10,000 images were randomly selected from each dataset for experiments. The network was pre-trained with 8000 pairs of images for training and 2000 pairs of images for validation. In the domain adaptation stage, 500 pairs of images were used for training and testing respectively. The batch size in the domain adaptation stage was set to 8, and the domain adaptation stage was trained for 60 rounds. It was compared with the backbone network SRNet of the present invention. First, the problem of the degradation of detection performance in the CSM case was tested through experiments. For example, B-B here represents training using the data on BOSSbase and testing on the data from the same source. And B-A represents the CSM case, that is, the result of training on BOSSbase and testing on ALASKA#2.

[0082] Table 1 Detection results of the backbone network SRNet for detecting steganographed images of the S-UNIWARD algorithm when CSM occurs

[0083]

[0084] The occurrence of the CSM situation leads to a high probability of a significant decline in performance when the network trained on the source domain is used for detection on the target domain. The solution proposed in this invention can effectively improve the cross-domain detection performance when using multiple source domains for domain adaptation training. When using two source domains, the experiments are conducted under the same scenario as shown in Table 2:

[0085] Table 2 Comparison of the detection effect of the network proposed in this paper in the case of dual source domains and the detection accuracy of the backbone network

[0086]

[0087] Here, the "best accuracy on the source domain" refers to the maximum value among the detection accuracies obtained when SRNet detects the target domain samples on the two source domains respectively. It can be seen from the experimental results that the method proposed in this invention effectively extracts the invariant features of multiple domains through operations such as feature mixing and aligning the feature distributions of multiple domains, and improves the image steganography detection accuracy of the network in the CSM scenario.

[0088] Embodiment 2: As Figure 4 shown, in order to achieve the above object, the present invention discloses a steganography detection system based on multi-source domain adaptation, including:

[0089] A feature extraction module 11, configured to receive data from multiple source domains and a target domain, input the data from the multiple source domains and the target domain into a pre-established common feature extractor, and extract the feature representations of the source domain and the target domain;

[0090] A feature mixing module 12, configured to pair and input the feature representations of the source domain and the target domain into a pre-established feature-level mixing module for weighted mixing to obtain mixed features;

[0091] A feature prediction module 13, configured to input the mixed features and the paired feature representations of the source domain and the target domain into a domain feature extractor to obtain processed features, and input the processed features into a pre-established domain-specific classifier to output the prediction outputs of the source, target, and mixed features.

[0092] Based on the same inventive concept, the present invention further provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is configured to execute the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically to load and execute one or more instructions in the computer storage medium to implement the above method.

[0093] It should be further noted that, based on the same inventive concept, the present invention further provides a computer storage medium, on which a computer program is stored, and the computer program executes the above method when run by a processor. The storage medium may be any combination of one or more computer-readable media. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may, for example, but not be limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage medium include: an electrical connection with one or more wires, 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), an optical fiber, a portable compact disk read only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, the computer-readable storage medium may be any tangible medium that contains or stores a program, and the program can be used by or combined with an instruction execution system, apparatus, or device.

[0094] In the description of this specification, the description referring to terms such as "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0095] The foregoing has shown and described the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure will have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure claimed.

Claims

1. A steganalysis detection method based on multi-source domain adaptation, characterized in that: The method comprises the following steps: Receiving multiple source domain and target domain data, inputting the multiple source domain and target domain data into a pre-established common feature extractor, and extracting feature representations of the source domain and the target domain; The process of inputting multiple source domain and target domain data into a pre-established common feature extractor: After a common feature extractor G, the common feature extractor maps multiple source domain and target domain data from the original feature space to the common feature space; The public feature extractor is composed of the first three types of convolutional layers designed in SRNet. A batch of data, and target domain data Feature representation obtained by common feature extractor and The formula is: The feature representation obtained from each source domain Metropolis and Each pair is input into the feature-level mixing module to generate additional mixed features; The input feature-level mixing module receives the source domain and target domain output features from the common feature generator and , and then calculate the mixture ratio of the target domain according to the current number of training rounds , the update of the mixing ratio follows the following formula: in, are the maximum value of the mixing ratio, It is a parameter that controls the exponential growth rate. current_epoch and total_epoch represent the current training round and the total number of training rounds respectively. According to the calculated mixing ratio , the source domain features and target domain features are weighted and summed according to the mixing ratio to obtain the mixed features , the formula is: in, represents the addition operation between matrices, Indicates the number of samples, and Represent the source domain and target domain output features from the common feature generator respectively and The samples; In addition to feature mixing, labels also need to be mixed. The labels of the mixed features Calculated using the same mixing ratio as the feature, is the prediction result of the network on the target domain, is the label of the jth sample: use As mixed domain sample labels, target domain features And the mixed features They are fed into the subsequent network and the cross entropy loss is calculated based on the mixed labels. as the classification loss for the mixed domain, where is the network’s predicted output for the mixed features, is the number of samples in a batch: ； The feature representation of the source domain and the feature representation of the target domain are paired and input into the pre-established feature-level mixing module for weighted mixing to obtain the mixed features; The mixed features and the paired feature representations of the source domain and the target domain are input into the domain feature extractor to obtain the processed features, which are then input into the pre-established domain preset classifier to obtain the predicted outputs of the source, target, and mixed features.

2. The steganalysis detection method based on multi-source domain adaptation according to claim 1, characterized in that: The data sources of the multiple source domains and target domain data are: labeled data from N source domains ,in and Represents domain The data and its corresponding labels, as well as the target domain Unlabeled data , i represents the i-th source domain.

3. The steganalysis detection method based on multi-source domain adaptation according to claim 1, characterized in that: The source domain and target domain data are aligned, and the domain feature extractor F i Accept a batch from the source domain Features and from the target domain Features , and maps each pair of source and target domains into a preset feature space, and then executes the first-stage alignment strategy to align the feature distributions of the two domains, using the maximum mean difference (MMD) to perform the first-stage alignment, thereby determining the two distributions. and Whether they are the same or not depends on the observed sample data. MMD defines the difference measure between distributions by the following formula: in, Represents two probability distributions and The distance metric between them, H is the reproducing kernel Hilbert space RKHS with characteristic kernel k, and are samples from different distributions, Indicates mapping the original sample to a feature map in RKHS, Indicates that from the distribution mid-sampling And calculate the function The expected value of Indicates that from the distribution mid-sampling And calculate the function The expected value of; the characteristic kernel k represents , represents the inner product of vectors if and only if hour, and The distribution of is the same, MMD estimates the difference between the distributions by comparing the squared distance of the empirical kernel mean embedding, combined with the input of the domain feature extractor layer, the formula is as follows: here yes The unbiased estimator of , H is the reproducing kernel Hilbert space RKHS with characteristic kernel k; and is the number of samples in a batch, and is the output of the source domain and target domain processed by the common feature device, and Represent the samples therein respectively, so as to derive an estimate of the difference between each source domain and the target domain. According to MMD, the domain alignment loss MMD Loss of this layer is designed as: Among them F i represents the i-th domain-specific feature extractor, and It is the output of the source domain and target domain processed by the common feature extractor; N domain feature extractors that do not share weights map each pair of source domain and target domain data into the feature space, and the domain alignment loss is obtained. , by minimizing To achieve feature distribution alignment.

4. The steganalysis detection method based on multi-source domain adaptation according to claim 1, characterized in that: The process of inputting the processed features into the pre-established domain preset classifier to obtain the predicted output of source, target, and mixed features: based on The loss alignment strategy measures the prediction differences between different classifiers on the target domain data. The loss function has the following form: in and Represents the source domain and The classifier corresponding to the branch is The predicted probability of target domain samples, is the number of samples in a batch; by minimizing the loss, the mean of the prediction results of all classifiers on the target domain data is output as the final prediction result.

5. A steganalysis detection system based on multi-source domain adaptation, which adopts the steganalysis detection method based on multi-source domain adaptation according to any one of claims 1 to 4, characterized in that: include: A feature extraction module is used to receive multiple source domain and target domain data, input the multiple source domain and target domain data into a pre-established common feature extractor, and extract feature representations of the source domain and the target domain; A feature mixing module is used to pair the feature representation of the source domain and the feature representation of the target domain and input them into a pre-established feature-level mixing module for weighted mixing to obtain mixed features; The feature prediction module is used to input the mixed features and the paired feature representations of the source domain and the feature representations of the target domain into the domain feature extractor to obtain processed features, and input the processed features into a pre-established domain preset classifier to output the predicted output of the source, target, and mixed features.

6. A terminal device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, the steganalysis detection method based on multi-source domain adaptation according to any one of claims 1 to 4 is adopted.

7. A computer-readable storage medium storing a computer program, wherein: When the computer program is loaded and executed by the processor, a steganalysis detection method based on multi-source domain adaptation according to any one of claims 1 to 4 is adopted.

Images