Mimicry judgment method based on semantic feature consistency

Abstract

The invention provides a mimicry judgment method based on semantic feature consistency, which belongs to the technical field of network security, and comprises the following steps of: firstly, performing zero padding and uniform segmentation on response data of different executors, and then extracting multi-level deep semantic features through a dual coding generation network; then, designing a multi-class loss function to jointly optimize parameters of a generator and a discriminator, finally, constructing a normal response data model through unsupervised training, calculating a difference degree between to-be-tested data and the model, and judging whether the to-be-tested data is attacked or not in combination with a threshold value; according to the technical scheme, the false positive problem caused by the difference of executors in a traditional violent comparison method is solved, the sensitivity of shallow data comparison to noise or normal difference is overcome, the detection domain of normal response data is expanded, and the detection capacity of unknown attacks is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of network security, and in particular relates to a mimicry decision method based on semantic feature consistency. Background Art

[0002] Mimetic defense technology is an active defense technology proposed by the team of Wu Jiangxing, an academician of the Chinese Academy of Engineering. Its core principle is to use the system's dynamic heterogeneous redundant architecture to break the relatively static and deterministic environment that network attacks rely on. Specifically: First, construct a dynamic heterogeneous redundant architecture. The mimetic defense system contains multiple executors with the same functions but different internal structures, implementation methods and operating mechanisms. For example, in a network server scenario, these executors can be server instances built by different manufacturers based on different chip architectures, operating systems, and Web service software. The heterogeneity makes it difficult for attackers to use the same vulnerability or attack method to carry out effective attacks on all executors at the same time; then, the system will dynamically schedule these executors. Within a certain time interval, the input data will be distributed to different executors for processing in a specific order. After the processing is completed, it will be dispatched according to the rules. Switching the working order of the executors. This dynamic change increases the difficulty of the attack because it is difficult for attackers to predict and adapt to changes in the running status of the executors. Then, the executors perform synchronous processing and result arbitration. When an external data request arrives at the system, it will be sent to each heterogeneous executor at the same time. Each executor independently processes the input data and generates its own processing results. For example, for an HTTP request, different Web server executors will generate corresponding response content. Finally, because the system has an arbiter, it will collect the processing results of all executors and determine the final output result based on majority voting or other arbitration algorithms. If an executor produces an incorrect result due to an attack, as long as the number of incorrect results does not exceed half (taking majority voting as an example), the arbiter can select the correct result as the system output, thereby ensuring that the system provides normal services to the outside world.

[0003] Although the existing technology has met the current more common needs, it still has the following defects:

[0004] 1. False positive problem: Traditional mimicry defense arbiters use a brute-force comparison method, requiring strict consistency in response data from different executors. However, due to differences in executor software and hardware, response data is inherently inconsistent, leading to non-aggressive discrepancies being misjudged as security threats.

[0005] 2. Shallow data sensitivity: Existing methods rely on shallow data comparison (such as direct comparison of numerical values or plain text), which is highly sensitive to data noise or normal differences and cannot effectively distinguish between attack events and normal differences.

[0006] 3. Limited detection domain: Traditional methods cannot learn the statistical patterns of normal response data, resulting in a limited detection range and difficulty in responding to unknown attacks or complex semantic differences.

[0007] In view of the above, the present application provides a mimicry decision method based on semantic feature consistency to solve the above problems. Summary of the Invention

[0008] In response to the above situation, in order to overcome the defects of the existing technology, the present invention provides a mimicry judgment method based on semantic feature consistency. Through deep semantic feature learning, it reduces the sensitivity to shallow data differences. At the same time, it combines multi-level semantic feature optimization to improve the accuracy and stability of mimicry judgment, reduce false positive phenomena, expand the detection domain of normal response data, and improve the detection capability of unknown attacks. It is suitable for scenarios where security threat detection is performed on response data output by heterogeneous executors in a cyberspace mimicry defense architecture.

[0009] A mimicry judgment method based on semantic feature consistency, characterized by comprising the following steps:

[0010] S1. Preprocessing: Fill the response data of different actuators with zeros and evenly divide them to unify the input format;

[0011] S2. Semantic feature extraction: Extract multi-level deep semantic features through a dual encoding generation network (Encoder-Decoder-Encoder) to learn the statistical laws of normal response data;

[0012] S3. Joint Optimization: Design a multi-category loss function (reconstruction loss, adversarial loss, and semantic feature loss) to jointly optimize the generator and discriminator parameters in the dual encoding generative network to enhance semantic consistency representation.

[0013] S4. Judgment: Build a normal response data model through unsupervised training, calculate the difference between the test data and the model, and use the threshold to determine whether it has been attacked.

[0014] Furthermore, the unified input format in S1 specifically refers to unifying response data of different sizes into segmented data of a fixed size.

[0015] Furthermore, the dual encoding generation network in S2 includes a generator and a discriminator.

[0016] Furthermore, the generator is composed of two autoencoders (Encoder1, Encoder2) and a decoder (Decoder1) for extracting multi-level semantic features.

[0017] Furthermore, the discriminator uses a five-layer simplified convolutional neural network to distinguish the generated pseudo-normal data from the real data.

[0018] Furthermore, the multi-category loss function in S3 is specifically shown in the following formula:

[0019] Loss=λ1L1+λ2L2+λ3L3

[0020] =λ1||Z1-Z2||2+λ2||XX′||2+λ3L3

[0021] Among them, Z1 refers to the encoding output of Encoder1 for the input data X, Z2 refers to the encoding output of Encoder2 for the generated data X′, X refers to the original input data, and X′ refers to the reconstructed data generated by Decoder1 based on Z1. Since the semantic feature loss L1 is the core innovation, it is given the highest weight of 0.5. The data reconstruction capability L2 is the basic function, with the second highest weight of 0.3. The adversarial loss L3 is mainly used to assist the generator to approximate the true distribution, with the lowest weight of 0.2. Therefore, the weights λ1 = 0.5, λ2 = 0.3, and λ3 = 0.2 are set.

[0022] Furthermore, the threshold determination standard in S4 is to convert the output of the discriminator into a probability distribution through the Softmax activation function, that is, the output is [P 正常 , P 异常 ], satisfying P 正常 +P 异常 =1, if P 正常 >P 异常 , the current response data is judged to be normal, otherwise it is judged to be abnormal.

[0023] The beneficial effects of the above technical solution are:

[0024] 1. Reduce false positive rates: Through deep semantic feature learning, we can distinguish between aggressive and normal execution differences, reducing the false positive rate by more than 30%;

[0025] 2. Enhanced robustness: Joint optimization of multi-level semantic features improves tolerance to noise and normal differences, increasing detection accuracy by 25%;

[0026] 3. Efficient unsupervised training: No need to label attack samples; unknown threats can be detected through normal data modeling, improving training efficiency by 40%;

[0027] 4. Resource saving: Reduce the computational overhead of traditional brute force comparison and shorten system response time by 50%. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the system architecture of the present invention;

[0029] Figure 2 Schematic diagram of the pretreatment process of the present invention;

[0030] Figure 3 Schematic diagram of the network layer design of the generator and discriminator of the present invention;

[0031] Figure 4 Schematic diagram showing the comparison of false positive rates of different methods in a non-aggressive difference scenario in a specific implementation manner;

[0032] Figure 5 Schematic diagram of detection accuracy stability under different signal-to-noise ratios (SNRs) in a specific implementation manner. DETAILED DESCRIPTION

[0033] The aforementioned and other technical contents, features and effects of the present invention will be clearly presented in the following detailed description of the embodiments with reference to the drawings of this application. The contents mentioned in the following embodiments are all based on the drawings of the specification as a reference.

[0034] Example 1, overall working principle and process overview:

[0035] 1. Construct a learning method for semantic feature consistency based on generative adversarial networks (GANs). The workflow is as follows:

[0036] 1.1 Preprocessing: Fill the response data of different actuators with zeros and evenly divide them to unify the input format;

[0037] 1.2 Semantic Feature Extraction: Extract multi-level deep semantic features through a dual encoding generation network (Encoder-Decoder-Encoder) to learn the statistical laws of normal response data;

[0038] 1.3 Joint Optimization: Design multi-category loss functions (reconstruction loss, adversarial loss, and semantic feature loss), jointly optimize the generator and discriminator parameters, and enhance semantic consistency representation;

[0039] 1.4 Adjudication: Build a normal response data model through unsupervised training, calculate the difference between the test data and the model, and use the threshold to determine whether an attack has occurred.

[0040] It should be noted that the calculation of the difference is based on the generator's ability to model the semantic features of normal data, and is comprehensively evaluated through the following two dimensions:

[0041] Data reconstruction error: The data to be tested is input into the generator. The difference between the reconstructed data output by the generator and the original data to be tested at the pixel or numerical level (mean absolute error) is the difference. This error reflects the consistency of the data to be tested with the normal data model in the shallow format.

[0042] Semantic Feature Deviation: Extract the multi-level semantic features (feature vectors of Encoder1 and Encoder2) output by the generator encoder and calculate the statistical distance (Euclidean distance) between the test data features and the normal data feature library (constructed during the training phase). This deviation measures the degree of deep semantic match between the test data and the normal model.

[0043] The difference is ultimately a weighted comprehensive value of the above two types of indicators. The weight is dynamically adjusted according to the feature importance during normal data training to ensure sensitivity to aggressiveness differences while tolerating differences in normal execution bodies.

[0044] It is also important to note that the rules for threshold determination are specifically formulated as follows:

[0045] Statistical benchmarks during the training phase: After unsupervised training is completed, all normal response data are used to calculate the distribution of differences and record their maximum value, mean, and standard deviation;

[0046] Dynamic threshold selection: Based on the historical maximum value of normal data difference, a certain tolerance margin is added (such as the mean plus 3 times the standard deviation) to prevent misjudgment due to environmental noise or accidental fluctuations;

[0047] Adaptive adjustment: During the deployment phase, continuously monitor normal data differences and regularly update thresholds to avoid missed detections due to data distribution drift (such as system upgrades).

[0048] Therefore, the specific rules for threshold determination are:

[0049] If the comprehensive difference of the data to be tested is lower than the threshold, it is judged as a normal execution body difference and no alarm is triggered; if the comprehensive difference exceeds the threshold, it is judged as a potential attack and the mimicry defense mechanism is activated.

[0050] 2. Build a semantic feature consistency learning system based on generative adversarial network (GAN), such as Figure 1 As shown, its structure is as follows:

[0051] 2.1 Preprocessing module: It includes zero-padding unit and uniform segmentation unit, which unifies response data of different sizes into segmented data of fixed size;

[0052] 2.2 Generative Adversarial Network Design:

[0053] Generator (G): consists of two autoencoders (Encoder1, Encoder2) and a decoder (Decoder1), which is used to extract multi-level semantic features;

[0054] Discriminator (D): A five-layer simplified convolutional network is used to distinguish between the generated pseudo-normal data and the real data.

[0055] 2.3 Joint Optimization of Semantic Features:

[0056] Integrate multi-category loss functions (L1, L2, L3) and improve semantic consistency by alternately optimizing generator and discriminator parameters.

[0057] 3. Make key relationship connections:

[0058] 3.1 Connect the output of the preprocessing module to the input of the generator;

[0059] 3.2 The dual-encoded feature output of the generator interacts with the discriminator and adjusts the network parameters through feedback from the joint optimization module;

[0060] It is important to note that the interaction between the generator's dual-encoded feature output and the discriminator is as follows: the generator extracts primary semantic features of the input data through the first-layer encoder (Encoder1). After being reconstructed into pseudo-normal data by the decoder (Decoder1), the second-layer encoder (Encoder2) extracts deep semantic features from the reconstructed data, forming a multi-level feature representation. These features are transmitted to the discriminator through the parameter connection between the generator and the discriminator. The discriminator receives the pseudo-normal data output by the generator and the real data, outputs a judgment based on the difference in the semantic feature distribution between the two, and backpropagates the judgment result to the joint optimization module. The joint optimization part in 1.3 dynamically adjusts the network weights of the generator and discriminator based on the feedback from the discriminator (adversarial loss) and the generator's own reconstruction loss and semantic consistency loss. Specifically, the generator improves its semantic reconstruction ability by minimizing the multi-level loss function, while the discriminator enhances its ability to distinguish between real and fake data by maximizing the adversarial loss. The two achieve a dynamic balance in alternating training, ultimately making the semantic features output by the generator approach the statistical laws of real data, and the discriminator can accurately identify abnormal semantic deviations.

[0061] 3.3 The output of the discriminator is used to calculate the adversarial loss and drive the generator to approach the real data distribution.

[0062] It should be noted that in the attack prediction method, the generator simulates and generates pseudo data that is highly similar to real normal data, aiming to help the discriminator establish an accurate normal data distribution model. The core goal of the generator is not to directly generate attack samples, but to continuously optimize itself through unsupervised learning so that the pseudo data it outputs approximates the statistical laws of real normal data at the semantic feature level. This design enables the discriminator to deeply understand the inherent pattern of normal responses, so that it can keenly identify abnormal data (such as attack behavior) that deviates from this pattern during the detection phase. Through adversarial training between the generator and the discriminator, the system can dynamically enhance its ability to predict unknown threats without relying on prior attack knowledge, while avoiding misjudgments caused by inherent differences in the executors, ultimately achieving high-precision, low-false-positive attack detection.

[0063] Example 2 further describes the overall working process of this method:

[0064] 1. Data preprocessing:

[0065] like Figure 2 As shown, the response data of different execution bodies are input (size is M1 = 256 × 256, M2 = 128 × 128), expanded to 512 × 512 by zero padding, and evenly divided into 64 × 64 sub-blocks;

[0066] 2. Generate adversarial network training, such as Figure 3 As shown:

[0067] Generator (G): Encoder1 contains 5 layers of convolution (64×64→2×2), Decoder1 deconvolution restores the size, and Encoder2 encodes again to generate multi-level features. The specific network structure is shown in Table 1;

[0068] Discriminator (D): Input 64×64 data, output 1×2 discrimination results after 5 layers of convolution. The specific network structure is shown in Table 1;

[0069] Table 1 Generative Adversarial Network Structure for Semantic Feature Consistency

[0070]

[0071]

[0072] Loss function:

[0073] Loss=λ1L1+λ2L2+λ3L3

[0074] =λ1||Z1-Z2||2+λ2||XX′||2+λ3L3

[0075] Here, Z1 refers to the encoding output of Encoder1 for the input data X (i.e., the first-layer semantic features); Z2 refers to the encoding output of Encoder2 for the generated data X′ (i.e., the second-layer semantic features); X refers to the original input data (the preprocessed segmented response data); and X′ refers to the reconstructed data generated by Decoder1 based on Z1. The mathematical relationship between them is: Z1 = Encoder1(X); X′ = Decoder1(Z1); Z2 = Encoder2(X′).

[0076] Due to the physical meaning of the loss function:

[0077] L1=||Z1-Z2||2: Constrains the semantic feature consistency of the two encodings (Encoder1 and Encoder2) to ensure the stability of the features extracted by the generator;

[0078] L2 = ||XX′||2: constrains the similarity between the reconstructed data X′ and the original data X, ensuring the generator’s ability to restore normal data;

[0079] L3: Adversarial loss (such as cross entropy loss) drives the generator to generate outputs that conform to the real data distribution, and the discriminator distinguishes between real data and generated data.

[0080] Therefore, through experimental optimization methods, different weight combinations are experimented with through grid search, and the weight that makes the false positive rate (FPR) and detection accuracy of the verification set optimal is selected.

[0081] Finally determine the loss top importance priority:

[0082] Semantic feature consistency (L1) is the core innovation, and is given the highest weight λ1 = 0.5 to ensure that the model focuses on capturing deep semantic differences;

[0083] Data reconstruction capability (L2) is the basic function, with the second highest weight being λ2 = 0.3;

[0084] The adversarial loss (L3) is mainly used to assist the generator in approaching the true distribution, with the minimum weight λ3 = 0.2.

[0085] 3. Threshold setting and adjudication:

[0086] 3.1 Discriminator output processing:

[0087] The final output of the discriminator (D) is a 1×2 vector, which is converted into a probability distribution through the Softmax function, that is, the output is [P 正常 , P 异常 ], satisfying P 正常 +P 异常 =1, if P 正常 >P 异常 , the current response data is judged to be normal; otherwise, it is judged to be abnormal. The specific process is as follows:

[0088] Design input data:

[0089] The input of the discriminator (D) is two types of data: real normal data, which is the normal response data from the preprocessing module (such as 64×64 sub-blocks); generator output data, which is the pseudo normal data generated by the generator (G) (that is, the reconstruction result of the generator on the input data).

[0090] Design the discriminator network structure:

[0091] The discriminator consists of 5 convolutional layers (see Table 1 for specific parameters). Figure 1-3 ), and finally outputs a 1×2 vector, which represents the original score (logits) of whether the input data belongs to "normal" or "abnormal".

[0092] Probability conversion is performed through Softmax, and the formula is defined at the same time:

[0093] The original output of the discriminator is converted into a probability distribution through the Softmax function:

[0094]

[0095] Among them, s 正常 and s 异常 are the two original scores in the 1×2 vector output by the discriminator, satisfying P 正常 +P 异常 =1. P 正常 is the probability that the input data conforms to the normal response semantic features; P 异常 : The probability that the input data deviates from normal semantic features (possibly attack or abnormal data).

[0096] Finally, the judgment rules and thresholds are set. The basic rules are:

[0097] If P 正常 >P 异常 , the current data is judged to be normal; otherwise it is judged to be abnormal (equivalent to P 正常 <0.5) triggers abnormal alarm.

[0098] Example 3, training and verification of the feasibility of this method:

[0099] Normal data training: Use unsupervised training method, input only normal response data, optimize the parameters of generator (G) and discriminator (D), so that the discriminator can detect the normal data P 正常 maximize.

[0100] It is important to note that the unsupervised training process uses only normal response data as input, without relying on any attack samples or manually annotated labels, allowing the generator to learn to reconstruct the semantic feature distribution of normal data. At the same time, the discriminator gradually improves its ability to identify normal data features by comparing the pseudo-normal data reconstructed by the generator with the real normal data. Specifically, during the training phase, the generator continuously attempts to encode and decode the input normal data into pseudo-data that is close to the real distribution. The discriminator drives the generator to optimize its reconstruction capabilities by determining whether the input data is real normal data or data forged by the generator, ultimately enabling the generator to accurately capture the deep semantic patterns of normal data, while the discriminator establishes sensitivity to the features of normal data, providing a basis for subsequent anomaly detection.

[0101] It's also important to note that parameter optimization is achieved by alternately training the generator and discriminator: first, the generator's parameters are fixed, and normal data is fed into the generator to generate pseudo data. This pseudo data is then mixed with real data and fed into the discriminator. Backpropagation is used to update the discriminator's parameters to improve its ability to distinguish true from false. Then, the discriminator's parameters are fixed, and normal data is fed into the generator to generate pseudo data. The discriminator's judgment of the pseudo data is backpropagated to update the generator's parameters, making the generated pseudo data closer to the distribution of real data. During training, the generator optimizes its parameters by reducing reconstruction error and its ability to deceive the discriminator, while the discriminator adjusts its parameters by improving its accuracy in distinguishing true from false data. The two are iterated alternately until the model converges.

[0102] Anomaly detection verification: In the verification phase, a mixed data set containing normal and abnormal (simulated attacks) is input to observe the probability distribution of the discriminator output to ensure that P 异常 It increases significantly in attack scenarios.

[0103] It should be noted that the abnormal probability P of the discriminator is required in the verification stage. 异常 The core purpose of significantly increasing in attack scenarios is to ensure that the model can accurately identify security threats and trigger defense mechanisms by clearly distinguishing the probability distribution differences between normal data and attack data. Specifically, if P 异常 If it does not increase significantly, it means that the model cannot effectively learn the semantic feature boundaries of normal data, which may lead to missed attacks (false negatives) or mistakenly judge normal differences as attacks (false positives). 异常 The significant changes in , can prove that the model has the following capabilities: first, it can extract stable semantic rules from normal data to avoid being overly sensitive to the inherent differences of the execution body; second, it has a strong response to the semantic deviation introduced by the attack, and even if the attack method is unknown, it can be captured due to inconsistent semantic features; third, it provides a reliable basis for the subsequent setting of judgment thresholds, for example, when P 异常When the threshold is exceeded, defense actions are triggered, thereby reducing the false positive rate and improving the confidence of threat detection.

[0104] Example 4, an example to illustrate the overall process:

[0105] Input data: preprocessed segmented response data (size 64×64);

[0106] Generator processing: Generate pseudo-normal data and extract multi-level semantic features;

[0107] Discriminator decision: Input the semantic features output by the generator. If the discriminator outputs probabilities of 0.92 or 0.08, it is judged to be a safe state.

[0108] Attack scenario: If the output probability is 0.35 or 0.65, it is judged as abnormal and the mimicry defense mechanism is triggered.

[0109] The specific experimental process is:

[0110] 1. Experimental environment and dataset:

[0111] 1. Experimental platform:

[0112] Hardware: Intel Xeon E5-2698v4 CPU, NVIDIA RTX 4090D GPU

[0113] Software: PyTorch 1.8, CUDA 11.1

[0114] 2. Dataset:

[0115] Normal data: 5,000 HTTP response data items collected from three types of heterogeneous executables (ARM architecture, x86 architecture, and MIPS architecture), including differences generated by normal business interactions (such as timestamps, session IDs, and hardware identifiers);

[0116] Attack data: 1,000 entries total, including 5 known attack types (SQL injection, XSS, buffer overflow, ROP attack, and DDoS command) and 2 unknown attack types (zero-day vulnerability exploits).

[0117] 3. Comparison method:

[0118] Traditional brute force comparison (Baseline 1): A decision method based on exact match of hash values or strings;

[0119] Rule-based feature comparison (Baseline2): Use regular expressions to filter the difference fields before comparison.

[0120] 2. Experimental verification of specific technical issues:

[0121] Specific problem 1: Solve the false positive problem:

[0122] Ⅰ Experimental Design:

[0123] On a normal data set, the misjudgment rates of “non-offensive differences” by the traditional method and the present invention are statistically analyzed.

[0124] Non-offensive difference scenario:

[0125] Differences in the executable hardware identifiers (e.g., MAC address, CPU serial number);

[0126] Dynamically generate field differences (such as timestamps, random numbers);

[0127] Differences in encoding formats (such as response headers encoded in UTF-8 and GBK).

[0128] Ⅱ Experimental results, such as Figure 4 As shown in Table 2:

[0129] Table 2 Comparison of false positive rates of different methods in non-aggressive difference scenarios

[0130]

[0131] Conclusion: The false positive rate is reduced by 2.8% (8.3% → 5.5%) compared to traditional brute force matching, and the false positive rate is reduced by more than 30%. The efficiency calculation formula is as follows:

[0132]

[0133] Specific Problem 2: Overcoming Shallow Data Sensitivity:

[0134] Ⅰ Experimental Design:

[0135] Gaussian noise (SNR = 10dB ~ 20dB) was injected into the normal data to compare the detection accuracy of different methods in a noisy environment.

[0136] Ⅱ Experimental results are as follows Figure 5 As shown in Table 3:

[0137] Table 3 Detection accuracy stability under different signal-to-noise ratios (SNR)

[0138]

[0139] Conclusion: In a noisy environment, the detection accuracy is improved by 19.5% (74.3% → 93.8%) compared to the traditional method, and the detection accuracy is improved by 25%. The efficiency formula is as follows:

[0140]

[0141] Specific question 3, training efficiency evaluation:

[0142] Ⅰ Experimental Design:

[0143] Compare the training time of traditional supervised learning (which requires labeled attack samples) and the unsupervised learning of the present invention.

[0144] The results are shown in Table 4:

[0145] Table 4 Comparison of training time of different learning methods

[0146]

[0147] Conclusion: The training efficiency has increased by 40%. The efficiency calculation formula is as follows:

[0148]

[0149] Specific question 4, system response time evaluation:

[0150] Ⅰ Experimental Design:

[0151] The average response time of the traditional brute force comparison and the present invention was measured on 1000 pieces of data.

[0152] The results are shown in Table 5:

[0153] Table 5 Comparison of average response time of different decision-making methods

[0154]

[0155] Conclusion: The system response time is shortened by 50%. The efficiency formula is as follows:

[0156]

[0157] The above description is only for illustrating the present invention. It should be understood that the present invention is not limited to the above embodiments, and various variations that conform to the concept of the present invention are within the scope of protection of the present invention.

Claims

1. A mimicry decision method based on semantic feature consistency, characterized in that: The steps include: S1. Preprocessing: Fill the response data of different actuators with zeros and evenly divide them to unify the input format; S2. Semantic feature extraction: Extract multi-level deep semantic features through a dual encoding generation network (Encoder-Decoder-Encoder) to learn the statistical laws of normal response data; S3. Joint Optimization: Design a multi-category loss function (reconstruction loss, adversarial loss, and semantic feature loss) to jointly optimize the generator and discriminator parameters in the dual encoding generative network to enhance semantic consistency representation. S4. Judgment: Build a normal response data model through unsupervised training, calculate the difference between the test data and the model, and use the threshold to determine whether it has been attacked.

2. The mimicry decision method based on semantic feature consistency according to claim 1, characterized in that: The unified input format in S1 specifically refers to unifying response data of different sizes into segmented data of a fixed size.

3. The mimicry decision method based on semantic feature consistency according to claim 1, characterized in that: The dual encoding generation network in S2 includes a generator and a discriminator.

4. The mimicry decision method based on semantic feature consistency according to claim 3, characterized in that: The generator consists of two autoencoders (Encoder1, Encoder2) and a decoder (Decoder1) for extracting multi-level semantic features.

5. The mimicry decision method based on semantic feature consistency according to claim 3, characterized in that: The discriminator uses a five-layer simplified convolutional neural network to distinguish the generated pseudo-normal data from the real data.

6. The mimicry decision method based on semantic feature consistency according to claim 1, characterized in that: The multi-category loss function in S3 is specifically shown in the following formula: Loss=λ1L1+λ2L2+λ3L3 =λ1||Z1-Z2||2+λ2||XX′||2+λ3L3 Among them, Z1 refers to the encoding output of Encoder1 for the input data X, Z2 refers to the encoding output of Encoder2 for the generated data X′, X refers to the original input data, and X′ refers to the reconstructed data generated by Decoder1 based on Z1. Since the semantic feature loss L1 is the core innovation, it is given the highest weight of 0.

5. The data reconstruction capability L2 is the basic function, with the second highest weight of 0.

3. The adversarial loss L3 is mainly used to assist the generator to approximate the true distribution, with the lowest weight of 0.

2. Therefore, the weights λ1 = 0.5, λ2 = 0.3, and λ3 = 0.2 are set.

7. The mimicry decision method based on semantic feature consistency according to claim 1, characterized in that: The threshold judgment standard in S4 is to convert the output of the discriminator into a probability distribution through the Softmax activation function, that is, the output is [P 正常 , P 异常 ], satisfying P 正常 +P 异常 =1, if P 正常 >P 异常 , the current response data is judged to be normal, otherwise it is judged to be abnormal.

Images