A neural network steganalysis detection method and system based on parameter distribution amplification

Abstract

This invention discloses a neural network steganalysis detection method based on parameter distribution amplification, comprising: iteratively amplifying the parameter distribution of the original neural network model to be detected using stochastic gradient descent according to a preset loss function until a preset number of iterations is reached, to obtain a fine-tuned neural network model to be detected; calculating the statistical moment vectors of the parameters of the original neural network model to be detected and the statistical moment vectors of the parameters of the fine-tuned neural network model to be detected, respectively, to obtain the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected; fusing the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected, to obtain fused features; performing logistic regression binary classification on the fused features, and outputting the steganalysis determination information of the original neural network model to be detected.

Description

Technical Field

[0001] This invention relates to the field of neural network steganalysis technology, and in particular to a neural network steganalysis method, system, electronic device, and storage medium based on parameter distribution amplification. Background Technology

[0002] Steganography is a covert information hiding technique primarily used in secure communications, but it undoubtedly creates opportunities for criminals. Network steganography mainly utilizes redundant fields and vulnerabilities in protocol rules to transmit secret information, exhibiting dynamism and enhanced concealment.

[0003] Over the past decade, numerous deep neural networks (DNNs) have been proposed and have made significant progress in many fields, such as computer vision and natural language processing. The parameter scale is considered one of the reasons why current DNNs outperform previous models. However, some work has also found that DNNs exhibit significant redundancy, meaning that only a small subset of parameters plays a crucial role in the specific task, while variations in the remaining parameters have minimal impact on performance. This phenomenon suggests that DNNs are ideal carriers for steganography, aiming to embed secret data with minimal suspicion from detectors. Currently, three neural network steganography methods exist: LSB encoding (least significant bit), COR (Correlation Encoding Attack), and SGN (Sign Encoding Attack). Subsequent methods like StegoNet and EvilModel improved upon LSB encoding by adaptively selecting the embedding region. Since neural network steganography can be used for malicious purposes, such as injecting malware and stealing sensitive data, steganography detection techniques targeting neural network steganography are needed.

[0004] To those skilled in the art, interfering with the LSB (Legacy Sequence Buffer) of the parameters can disable LSB-based encoding methods without affecting the performance of the original model. In contrast, COR (Concurrent Orientation) and SGN (Secret Generic Array) are more difficult to defend against, as they can establish a mapping between parameters and secret data. Obviously, this mapping will inevitably lead to changes in the parameter distribution. The first work on neural network steganalysis detection detected malicious models by measuring the statistical differences in model parameters. However, malicious users can reduce the embedding rate of the neural network to evade detection by this method. Due to the large number of parameters, even at low embedding rates, they can still effectively embed malware or steal sensitive data. Therefore, there is an urgent need for detection methods that can effectively detect malicious neural network models at low embedding rates. Summary of the Invention

[0005] In view of the above problems, the present invention provides a neural network steganalysis detection method, system, electronic device and storage medium based on parameter distribution amplification, in order to solve at least one of the above problems.

[0006] According to a first aspect of the present invention, a neural network steganalysis detection method based on parameter distribution amplification is provided, comprising:

[0007] Based on the preset loss function, the parameter distribution of the original neural network model to be detected is iteratively amplified using the stochastic gradient descent method until the preset number of iterations is reached, resulting in a fine-tuned neural network model to be detected.

[0008] The statistical moment vectors of the parameters of the original neural network model to be detected and the statistical moment vectors of the parameters of the fine-tuned neural network model to be detected are calculated respectively to obtain the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected.

[0009] Feature fusion is performed on the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected to obtain fused features;

[0010] Logistic regression binary classification is performed on the fused features to output the steganalysis determination information of the original neural network model to be detected, wherein the steganalysis determination information is used to characterize whether the original neural network model to be detected has embedded encrypted data.

[0011] According to an embodiment of the present invention, the preset loss function includes a cross-entropy loss function without applied steganographic constraints.

[0012] According to an embodiment of the present invention, the above-described method of iteratively amplifying the parameter distribution of the original neural network model to be detected using a stochastic gradient descent method based on a preset loss function until a preset number of iterations is reached, to obtain a fine-tuned neural network model to be detected, includes:

[0013] The sample set is randomly divided into a training sample set and a test sample set. The sample set includes the dataset and a label set that corresponds one-to-one with the dataset.

[0014] The dataset from the training sample set is input into the original neural network model to be detected, and the label set output by the model is obtained.

[0015] The stochastic gradient descent method is used to process the label set in the training sample set and the label set output by the model. Based on the processing results, the parameters of the original neural network model to be detected are updated using a preset loss function to obtain new model parameters.

[0016] A new neural network model is constructed using the new model parameters, and the parameters of the new neural network model are fine-tuned using a set of test samples to obtain the fine-tuned neural network model to be tested.

[0017] The process involves iteratively performing label set acquisition, stochastic gradient descent processing, parameter update, model building, and fine-tuning operations until the preset number of iterations is met, resulting in a fine-tuned neural network model to be detected.

[0018] According to embodiments of the present invention, the original neural network model to be detected includes a COR-based neural network model and an SGN-based neural network model.

[0019] According to an embodiment of the present invention, the fine-tuned neural network model to be detected differs from the original neural network model to be detected in terms of parameter distribution.

[0020] According to an embodiment of the present invention, the above-described logistic regression binary classification processing of the fused features to output the steganalysis information of the original neural network model to be detected includes:

[0021] If the original neural network model to be detected contains steganographic information, the original neural network model to be detected is determined to be a malicious neural network model.

[0022] If the original neural network model to be tested does not contain steganographic information, the original neural network model to be tested is determined to be a benign neural network model.

[0023] According to an embodiment of the present invention, the feature fusion of the feature vector of the original neural network model to be detected and the feature vector of the fine-tuned neural network model to be detected, to obtain the fused features, includes:

[0024] The feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected are used to calculate the vector difference to obtain the fused features.

[0025] According to a second aspect of the present invention, a neural network steganalysis detection system based on parameter distribution amplification is provided, comprising:

[0026] The difference amplification module is used to iteratively amplify the parameter distribution of the original neural network model to be detected using the stochastic gradient descent method according to the preset loss function until the preset number of iterations is reached, so as to obtain the fine-tuned neural network model to be detected.

[0027] The feature extraction module is used to calculate the statistical moment vectors of the parameters of the original neural network model to be detected and the statistical moment vectors of the parameters of the fine-tuned neural network model to be detected, respectively, to obtain the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected.

[0028] The feature fusion module is used to fuse the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected to obtain fused features;

[0029] The classification module is used to perform logistic regression binary classification on the fused features and output the steganalysis determination information of the original neural network model to be detected, wherein the steganalysis determination information is used to characterize whether the original neural network model to be detected has embedded encrypted data.

[0030] According to a third aspect of the present invention, an electronic device is provided, comprising:

[0031] One or more processors;

[0032] Storage device for storing one or more programs.

[0033] Specifically, when one or more programs are executed by one or more processors, the one or more processors execute a neural network steganalysis detection method based on parameter distribution amplification.

[0034] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform a neural network steganalysis detection method based on parameter distribution amplification.

[0035] The neural network steganography detection method based on parameter distribution amplification provided by this invention utilizes the sensitivity of malicious neural network model parameters to fine-tuning without steganography constraints. In data theft and malicious code detection scenarios, it can detect malicious models with low embedding rates that hide sensitive data or malicious code. It can efficiently detect malicious neural network models with steganography information and thus enable the safe use of secure neural network models without steganography information. Attached Figure Description

[0036] Figure 1 This is a schematic diagram illustrating the principle of a low embedding rate neural network steganalysis detection method based on parameter distribution amplification according to an embodiment of the present invention.

[0037] Figure 2 This is a flowchart of a neural network steganalysis detection method based on parameter distribution amplification according to an embodiment of the present invention;

[0038] Figure 3 This is a flowchart illustrating the acquisition of a fine-tuned neural network model to be detected according to an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the structure of a neural network steganalysis detection system based on parameter distribution amplification according to an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of the framework of a low embedding rate neural network steganalysis detection method based on parameter distribution amplification according to an embodiment of the present invention;

[0041] Figure 6 A block diagram of an electronic device suitable for implementing a neural network steganalysis method based on parameter distribution amplification according to an embodiment of the present invention is shown schematically. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0043] Network steganalysis is the inverse process of network steganography. In recent years, research on steganalysis has largely focused on general steganalysis, using raw and hidden data to train a classifier to detect various steganalysis methods. Convolutional neural networks, with their unique convolutional processing capabilities, can directly and automatically extract high-level features for further analysis without complex data preprocessing, while exponentially reducing computational complexity.

[0044] For steganalysis of neural network models, this invention provides a neural network steganalysis detection method, system, electronic device, and storage medium based on parameter distribution amplification. It allows for fine-tuning of the neural network model without steganalysis constraints, using a relatively high learning rate. The parameter distribution of the fine-tuned malicious model will differ significantly from the original malicious model. In contrast, the parameter distribution of the benign model changes very little before and after fine-tuning. This amplifies the statistical differences in parameters between the malicious and benign models. Then, the statistical moments of the model parameters before and after fine-tuning are calculated and fused as features, and input into a logistic regression classifier to determine the detection result.

[0045] As can be seen from the above technical solution provided by the present invention, the above solution designs features based on the sensitivity of malicious models and benign models to fine-tuning operations without steganography constraints at relatively high learning rates, making it easier to detect malicious neural network models with low embedding rates.

[0046] Figure 1 This is a schematic diagram illustrating the principle of a low embedding rate neural network steganalysis detection method based on parameter distribution amplification according to an embodiment of the present invention.

[0047] According to the training principles of neural networks, the distribution of model parameters is closely related to the loss function. Both COR and SGN steganography methods add steganographic constraints to the loss function during training to embed data, causing the parameter distributions of malicious and benign models to differ. However, as... Figure 1As shown in (a), the parameter distribution differences between malicious and benign models at low embedding rates may be small. This is because the number of parameters mapped to the secret data is reduced during training.

[0048] Given the trainability of neural networks, fine-tuning techniques are used to amplify the difference in parameter distribution between malicious and benign models. Since the specific data hidden by the steganologist in reality is unknown, it is difficult to impose the same malicious constraints on the loss function to fine-tune the malicious model. Therefore, the fine-tuning process below does not include malicious constraints, and the learning rate decays once per 0.1 factor during iteration.

[0049] As mentioned above, the loss functions for malicious and benign models are different. If these models are fine-tuned without malicious constraints, the parameter distribution change of the malicious model will be greater than that of the benign model; this is a key technical aspect of this invention. However, as... Figure 1 As shown in (b), directly using the last learning rate from the original training process as the initial learning rate for fine-tuning may cause these models to remain at a local optimum. Figure 1 (c) and Figure 1 As shown in (d), increasing the initial learning rate of the fine-tuning by 10 times causes the model to escape local optima, thereby amplifying the difference in parameter distribution between the fine-tuned malicious model and the benign model.

[0050] Figure 2 This is a flowchart of a neural network steganalysis detection method based on parameter distribution amplification according to an embodiment of the present invention.

[0051] like Figure 2 As shown, the above-mentioned neural network steganalysis detection method based on parameter distribution amplification includes operations S210 to S240.

[0052] In operation S210, according to the preset loss function, the parameter distribution of the original neural network model to be detected is iteratively amplified using the stochastic gradient descent method until the preset number of iterations is reached, resulting in the fine-tuned neural network model to be detected.

[0053] The aforementioned parameter distribution amplification refers to fine-tuning the parameter distribution of benign neural network models and malicious neural network models to change in different directions.

[0054] The original neural network models to be detected include neural network models for processing image data (such as image classification networks) or neural network models for processing text data.

[0055] In operation S220, the statistical moment vectors of the parameters of the original neural network model to be detected and the statistical moment vectors of the parameters of the fine-tuned neural network model to be detected are calculated respectively, so as to obtain the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected.

[0056] In operation S230, the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected are fused to obtain fused features.

[0057] In operation S240, logistic regression binary classification is performed on the fused features to output the steganalysis determination information of the original neural network model to be detected, wherein the steganalysis determination information is used to characterize whether the original neural network model to be detected has embedded encrypted data.

[0058] According to an embodiment of the present invention, the above-described logistic regression binary classification processing of the fused features to output the steganalysis information of the original neural network model to be detected includes:

[0059] If the original neural network model to be detected contains steganographic information, the original neural network model to be detected is determined to be a malicious neural network model.

[0060] If the original neural network model to be tested does not contain steganographic information, the original neural network model to be tested is determined to be a benign neural network model.

[0061] The aforementioned benign neural network model is a safe neural network model.

[0062] The neural network steganography detection method based on parameter distribution amplification provided by this invention utilizes the sensitivity of malicious neural network model parameters to fine-tuning without steganography constraints. In data theft and malicious code detection scenarios, it can detect malicious models with low embedding rates that hide sensitive data or malicious code. It can efficiently detect malicious neural network models with steganography information and thus enable the safe use of secure neural network models without steganography information.

[0063] According to an embodiment of the present invention, the preset loss function includes a cross-entropy loss function without applied steganographic constraints.

[0064] Figure 3 This is a flowchart of obtaining the fine-tuned neural network model to be detected according to an embodiment of the present invention.

[0065] like Figure 3As shown, the above-mentioned method iteratively amplifies the parameter distribution of the original neural network model to be detected using the stochastic gradient descent method according to the preset loss function until the preset number of iterations is reached, resulting in the fine-tuned neural network model to be detected, including operations S310 to S350.

[0066] In operation S310, the sample set is randomly divided into a training sample set and a test sample set. The sample set includes the dataset and a label set that corresponds one-to-one with the dataset.

[0067] In operation S320, the dataset from the training sample set is input into the original neural network model to be detected, and the label set output by the model is obtained.

[0068] In operation S330, the stochastic gradient descent method is used to process the label set in the training sample set and the label set output by the model. Based on the processing results, the parameters of the original neural network model to be detected are updated using a preset loss function to obtain new model parameters.

[0069] In operation S340, a new neural network model is constructed using the new model parameters, and the parameters of the new neural network model are fine-tuned using the test sample set to obtain the fine-tuned neural network model to be detected.

[0070] In operation S350, the label set acquisition operation, stochastic gradient descent processing operation, parameter update operation, model construction and fine-tuning operation are performed iteratively until the preset number of iterations is met, and the fine-tuned neural network model to be detected is obtained.

[0071] In operation S350, the above operations S310 to S340 are performed iteratively.

[0072] According to embodiments of the present invention, the original neural network model to be detected includes a COR-based neural network model and an SGN-based neural network model.

[0073] According to an embodiment of the present invention, the fine-tuned neural network model to be detected differs from the original neural network model to be detected in terms of parameter distribution.

[0074] According to an embodiment of the present invention, the above-mentioned feature fusion of the feature vector of the original neural network model to be detected and the feature vector of the fine-tuned neural network model to be detected to obtain fused features includes: calculating the vector difference between the feature vector of the original neural network model to be detected and the feature vector of the fine-tuned neural network model to be detected to obtain fused features.

[0075] Figure 4 This is a schematic diagram of the structure of a neural network steganalysis detection system based on parameter distribution amplification according to an embodiment of the present invention.

[0076] like Figure 4 As shown, the above-described neural network steganalysis detection system 400 based on parameter distribution amplification includes a difference amplification module 410, a feature extraction module 420, a feature fusion module 430, and a classification module 440.

[0077] The difference amplification module 410 is used to iteratively amplify the parameter distribution of the original neural network model to be detected using the stochastic gradient descent method according to the preset loss function until the preset number of iterations is reached, so as to obtain the fine-tuned neural network model to be detected.

[0078] The feature extraction module 420 is used to calculate the statistical moment vectors of the parameters of the original neural network model to be detected and the statistical moment vectors of the parameters of the fine-tuned neural network model to be detected, respectively, to obtain the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected.

[0079] The feature fusion module 430 is used to fuse the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected to obtain fused features.

[0080] The classification module 440 is used to perform logistic regression binary classification on the fused features and output the steganalysis information of the original neural network model to be detected.

[0081] Figure 5 This is a schematic diagram of the framework of a low embedding rate neural network steganalysis detection method based on parameter distribution amplification according to an embodiment of the present invention.

[0082] The following is combined Figure 5 The methods and systems provided by the present invention will be described in further detail below.

[0083] To facilitate the description of the embodiments of the present invention, the original model to be detected and the fine-tuned model to be detected are respectively referred to as... and Malicious models are called and A benign model is called and . and It is the size of The dataset and its corresponding label set. The parameter is from arrive The mapping. Denotes the set of labeled data points used for training, where It's its size. This represents the set of labeled data points used for testing, where It's its size.

[0084] like Figure 5 The overview shown is of the Low Embedding Rate Neural Network Stegia Detection Method (PSDAS) based on parameter distribution amplification. The entire framework consists of four modules, mainly including: difference amplification module, feature extraction module, feature fusion module, and classification module.

[0085] In the difference amplification module, the parameters of the original model to be detected are fine-tuned to change their distribution. The calculation method of the parameters of the original model to be detected is shown in formula (1):

[0086] (1),

[0087] in, and These represent the original model parameters and the fine-tuned model parameters, respectively. It is the cross-entropy loss function without imposing steganography constraints; Used to select the parameter that minimizes the value of the loss function.

[0088] This invention uses stochastic gradient descent (SGD) as the optimization function. The initial learning rate during fine-tuning is 10 times the last learning rate used during the original model training process. As iterations proceed, the differences in parameter distributions between the malicious and benign models are amplified. After reaching a specified number of iterations (10 for COR and 40 for SGN), the fine-tuned test model is obtained.

[0089] In the feature extraction module, statistical moments are used... The distribution of parameters is used as a feature to describe the distribution of parameters. It is a statistical moment of order j For example, a typical statistical moment is In mathematics, this is called expectation. Setting COR... Set up SGN .

[0090] Taking COR as an example, the statistical moment vector of a single-layer parameter is expressed by formulas (2) and (3):

[0091] (2),

[0092] (3),

[0093] in Represents the first... layer.

[0094] Calculate the statistical moment vector of each layer parameter. All vectors constitute the features of a model, which can be expressed by formulas (4) and (5):

[0095] (4),

[0096] (5),

[0097] in It represents the number of parameter layers in the model under test.

[0098] In the feature fusion module, the final features are obtained by fusing the features of the original and fine-tuned model under test. After a certain fine-tuning, the differences in parameter distribution are amplified, and the two features are fused using formula (6):

[0099] (6),

[0100] in This represents the feature difference between the original test model and the fine-tuned test model.

[0101] In the classification module, The input will be used to perform binary classification with a logistic regression classifier to determine whether the model under test is malicious or benign.

[0102] To better illustrate the advantages of the technical solution provided by the present invention, an experimental setup is set up below to verify the method provided by the present invention.

[0103] First, the experimental setup was conducted. This invention uses two typical image classification networks, VGG16 and ResNet34, as steganographic covers. CIFAR10 was used as the dataset for training all models. Specifically, 100 training iterations were performed using the cross-entropy loss function and SGD optimization function, with an initial learning rate of 0.1, decaying to 0.01 after 60 iterations. For malicious models, these settings were consistent with the training process for benign models, except for the steganographic constraints. The steganographic regularization coefficients for COR and SGN were set to 1.0 and 50.0, respectively. The embedding rates were set to 0.01, 0.03, and 0.05, representing the utilization rate of parameters available for embedding secret data. The embedded data consisted of scrambled and encrypted images from CIFAR10, with scrambling performed pixel-wise and bit-wise for COR and SGN, respectively. For all steganographic methods and settings, 100 benign models and 100 malicious models were trained, and the model with the best performance during the iteration process was selected to form the steganographic dataset.

[0104] For the metric, the average accuracy of 5-fold cross-validation using a binary classifier is used to evaluate the detection performance of the steganalysis method. A baseline method is used for comparison with the first neural network steganalysis method.

[0105] Secondly, the experimental effects of the present invention will be further explained in conjunction with Tables 1 to 4.

[0106] As shown in Tables 1 and 2, the experimental results demonstrate that the detection accuracy of the proposed method for COR steganography malicious models with low embedding rates is significantly higher than that of the baseline method under all experimental settings.

[0107] Table 1. Detection accuracy of the steganalysis detection method provided by this invention for the COR steganalysis VGG16 malicious model with low embedding rate.

[0108]

[0109] Table 2 shows the accuracy of the steganalysis detection method provided by this invention for detecting COR steganalysis ResNet34 malicious models with low embedding rates.

[0110]

[0111] As shown in Tables 3 and 4, the experimental results demonstrate that the steganography detection method provided by this invention achieves significantly higher detection accuracy than the baseline method for SGN steganography VGG16 and ResNet34 malicious models with low embedding rates under all experimental settings.

[0112] Table 3. Detection accuracy of the steganalysis detection method provided by this invention for the SGN steganalysis VGG16 malicious model with low embedding rate.

[0113]

[0114] Table 4 shows the accuracy of the steganography detection method provided by this invention for detecting SGN steganography ResNet34 malicious models with low embedding rates.

[0115]

[0116] Figure 6 A block diagram of an electronic device suitable for implementing a neural network steganalysis method based on parameter distribution amplification according to an embodiment of the present invention is shown schematically.

[0117] like Figure 6As shown, an electronic device 600 according to an embodiment of the present invention includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 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.

[0118] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in said one or more memories.

[0119] According to an embodiment of the present invention, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the I / O interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0120] 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 it may exist independently and not assembled 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.

[0121] According to embodiments of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, a computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0122] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0123] Those skilled in the art will understand that the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments and / or claims of the present invention can be combined or combined in various ways 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.

[0124] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely 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 within the protection scope of the present invention.

Claims

1. A neural network steganalysis detection method based on parameter distribution amplification, comprising: According to a preset loss function, the original neural network model to be detected is iteratively amplified in terms of parameter distribution using stochastic gradient descent until a preset number of iterations is reached, resulting in a fine-tuned neural network model to be detected. This includes: randomly dividing the sample set into a training sample set and a test sample set, wherein the sample set includes a dataset and a label set corresponding to each dataset; inputting the dataset from the training sample set into the original neural network model to be detected to obtain the label set output by the model; processing the label set in the training sample set and the label set output by the model using the stochastic gradient descent method, and based on the processing result, using the unadjusted... The parameters of the original neural network model to be detected are updated using the cross-entropy loss function with constraints to obtain new model parameters. The initial learning rate in the fine-tuning is 10 times the last learning rate during the training of the original neural network model to be detected. A new neural network model is constructed using the new model parameters, and the parameters of the new neural network model are fine-tuned using the test sample set to obtain the fine-tuned neural network model to be detected. The label set acquisition operation, stochastic gradient descent processing operation, parameter update operation, model construction and fine-tuning operation are iteratively performed until the preset number of iterations is met, resulting in the fine-tuned neural network model to be detected. The statistical moment vectors of the parameters of the original neural network model to be detected and the statistical moment vectors of the parameters of the fine-tuned neural network model to be detected are calculated respectively to obtain the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected. The feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected are fused to obtain fused features; The fused features are subjected to logistic regression binary classification to output the steganalysis determination information of the original neural network model to be detected, wherein the steganalysis determination information is used to characterize whether the original neural network model to be detected has embedded encrypted data.

2. The method according to claim 1, wherein, The original neural network models to be detected include a COR-based neural network model and an SGN-based neural network model.

3. The method according to claim 1, wherein, The fine-tuned neural network model to be detected differs from the original neural network model to be detected in terms of parameter distribution.

4. The method according to claim 1, wherein, The step of performing logistic regression binary classification on the fused features to output the steganalysis determination information of the original neural network model to be detected includes: If the original neural network model to be detected contains steganographic information, the original neural network model to be detected is determined to be a malicious neural network model. If the original neural network model to be detected does not contain steganographic information, the original neural network model to be detected is determined to be a benign neural network model.

5. The method according to claim 1, wherein, The feature fusion of the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected, to obtain the fused features, includes: The feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected are used to calculate the vector difference to obtain the fused features.

6. A neural network steganalysis detection system based on parameter distribution amplification, comprising: The difference amplification module is used to iteratively amplify the parameter distribution of the original neural network model to be detected using the stochastic gradient descent method according to the preset loss function until the preset number of iterations is reached, so as to obtain the fine-tuned neural network model to be detected. The feature extraction module is used to calculate the statistical moment vectors of the parameters of the original neural network model to be detected and the statistical moment vectors of the parameters of the fine-tuned neural network model to be detected, respectively, to obtain the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected. The feature fusion module is used to fuse the feature vectors of the original neural network model to be detected and the feature vectors of the fine-tuned neural network model to be detected to obtain fused features. The classification module is used to perform logistic regression binary classification on the fused features and output the steganalysis determination information of the original neural network model to be detected, wherein the steganalysis determination information is used to characterize whether the original neural network model to be detected has embedded encrypted data. Specifically, based on a preset loss function, the original neural network model to be detected is iteratively amplified in terms of parameter distribution using stochastic gradient descent until a preset number of iterations is reached, resulting in a fine-tuned neural network model to be detected, including: The sample set is randomly divided into a training sample set and a test sample set, wherein the sample set includes a dataset and a label set that corresponds one-to-one with the dataset; The dataset from the training sample set is input into the original neural network model to be detected to obtain the label set output by the model; The label set in the training sample set and the label set output by the model are processed using the stochastic gradient descent method. Based on the processing results, the parameters of the original neural network model to be detected are updated using the cross-entropy loss function without steganalysis constraints to obtain new model parameters. The initial learning rate in the fine-tuning is 10 times the last learning rate during the training of the original neural network model to be detected. The new neural network model is constructed using the new model parameters, and the parameters of the new neural network model are fine-tuned using the test sample set to obtain the fine-tuned neural network model to be detected. The process involves iteratively performing label set acquisition, stochastic gradient descent processing, parameter update, model construction, and fine-tuning operations until the preset number of iterations is met, resulting in the fine-tuned neural network model to be detected.

7. An electronic device, comprising: One or more processors; Storage device for storing one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors perform the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 5.

Images