A one-class adversarial fraud detection method based on explicit class representation

Abstract

The present invention discloses a one-class adversarial fraud detection method based on explicit class representation, using an improved autoencoder and an improved generative adversarial network, including four stages: the first stage: extracting preliminary features based on normal behavior data, and extracting its data features using the improved autoencoder; the second stage: modifying the objective function of the original generative adversarial network to obtain an improved generative adversarial network, and using the improved generative adversarial network to generate pseudo-abnormal behavior data; the third stage: inputting normal user behavior data and pseudo-abnormal behavior data into the improved autoencoder for training together, and extracting the final features of the normal behavior data using the encoder after training; the fourth stage: training the improved generative adversarial network with the final extracted normal behavior data features, and using the discriminator obtained after training to detect fraud as a fraud detector. The method of the present invention has significant improvements in the accuracy and stability of fraud detection.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a classification adversarial fraud detection method based on explicit class representation. Background Art

[0002] Fraud is widely present in our lives. Fraud is everywhere in the Internet, telecommunications, insurance (health, automobile, etc.) claims, banks (tax return claims, credit card transactions, etc.), etc. Fraud detection is a major problem that needs to be solved urgently.

[0003] Technically, fraud detection mainly has the following technical problems:

[0004] (1) Fraudulent behavior is generally adaptive and has a variety of types. Some traditional methods are based on robust features manually extracted by experts for detection. Once the fraud type changes, the previous methods are no longer applicable.

[0005] (2) The amount and dimension of fraudulent data are large. As the amount and dimension of data increase, the amount of computation required by traditional methods increases exponentially, and the computational cost is too high.

[0006] (3) Data on fraudulent behavior is not easy to obtain. Many emerging machine learning methods are very advantageous in processing large amounts of data, but machine learning methods require a large amount of data support to achieve good results. First of all, fraudulent behavior is a minority in the entire social behavior, and a large part of fraudulent behavior has no data records. Therefore, the diversity and imbalance of behavioral data also bring challenges to fraud detection.

[0007] Many traditional machine learning methods, such as the one-class nearest neighbor method (OCNN), one-class Gaussian process (OCGP) and one-class support vector machine (OCSVM), can successfully detect fraud, but these traditional methods are generally based on robust features manually extracted by experts. Once the type of fraud changes, the previous methods are no longer applicable. In addition, the fraudster behavior data that can be obtained is much less than the behavior data of normal users. The diversity and imbalance of behavioral data make traditional machine learning methods no longer applicable. Unsupervised or semi-supervised deep learning methods are more widely used in fraud detection because they can better overcome the above problems.

[0008] One-class adversarial network (OCAN) is an adversarial fraud detection network based on one-class training. Its training mainly includes two stages. The first stage is feature learning. LSTM-AE is a variant of autoencoder. It is different from conventional autoencoder in that both input and output are sequences. OCAN uses LSTM-AE to train the user's behavior data. After the training is completed, the fixed-length feature representation of the user data can be calculated through the encoder. The second stage is to train complementary GAN. Complementary GAN is obtained by modifying the generator and detector objective functions in the original generative adversarial network. It defines an ideal pseudo-abnormal data distribution, and the generator loss function is changed to the KL divergence between the generated data distribution and the ideal distribution. Through training, the generated distribution can gradually fit the ideal distribution of pseudo-abnormal data. The discriminator is slightly modified, and one more item is added on the basis of the original discriminator loss function, so that the discriminator is more inclined to detect normal users more accurately. After the training is completed, the discriminator can be used as a detector to distinguish normal users from malicious users.

[0009] OCAN uses an autoencoder to extract features from the behavior record data of normal users in Wikipedia, making it easier to process. Fence GAN also uses the discriminator as an anomaly detector, but it directly modifies the objective function of the generator so that the generated samples are distributed in the edge area of ​​the normal data, that is, the area where the discriminator judgment value is 0.5, and adds a loss function to prevent the generation of pseudo-abnormal data from aggregating, which also achieves good results. However, as the data dimension increases, it can be seen that the pseudo-abnormal data in Fence GAN is increasingly difficult to "surround" the normal data well.

[0010] OCAN adds an autoencoder to extract robust features of data, which enables it to better handle high-dimensional and variable-length data. We can see that although OCAN uses a generative network to solve the problem of lack of negative samples and extracts features from samples to improve the ability to process data, a good feature extractor should aim to maximize the distance between classes and minimize the distance within a class. Therefore, a feature extractor trained only on normal behavior data will ignore the goal of maximizing the distance between classes, causing deviations. At this time, feature extraction only focuses on "restoration" and ignores the "difference" between normal behavior data and abnormal behavior data that is concerned during detection, which often determines the upper limit of the detection effect that can be achieved on this basis. Summary of the invention

[0011] In view of the above problems, the present invention proposes a one-class adversarial fraud detection method based on explicit class representation, improves the autoencoder and the generative adversarial network, adds a loss function that reduces the intra-class spacing and increases the inter-class spacing in the middle layer, so that feature extraction can better serve the goal of fraud detection. At the same time, through two feature extractions, the autoencoder and the generative adversarial network are combined to complement each other, progress layer by layer, self-generate, and self-learn, thereby improving the detection accuracy and stability of subsequent detectors.

[0012] In order to achieve the above object, the present invention provides the following technical solutions:

[0013] A one-class adversarial fraud detection method based on class-explicit representation adopts improved autoencoder and improved generative adversarial network and consists of four stages:

[0014] Phase 1: Perform preliminary feature extraction based on normal behavior data and use the improved autoencoder to extract data robust features;

[0015] The second stage: modify the objective functions of the generator and discriminator in the original generative adversarial network to obtain an improved generative adversarial network, and use the improved generative adversarial network to generate pseudo abnormal behavior data;

[0016] Phase 3: Input normal user behavior data and pseudo-abnormal behavior data into the improved autoencoder for training together, and use the trained encoder to perform final feature extraction on the normal behavior data;

[0017] The fourth stage: Use the finally extracted normal behavior data features to train the improved generative adversarial network. The discriminator obtained after training is used as a fraud detector to detect fraud.

[0018] The improved autoencoder adds a central loss function to the middle layer of the autoencoder, so that the autoencoder satisfies the minimization of intra-class distance and maximization of inter-class distance during encoding; in the improved generative adversarial network, the loss function of the generator is the KL divergence of the generated distribution and the ideal distribution of the defined fraudulent data.

[0019] Furthermore, the specific process of the first stage is: train the autoencoder with normal data, and add central loss to its middle layer to minimize the intra-class distance; after the training is completed, encode the original normal data through encoder En1 to obtain its features, and retain the decoder De1 model weights.

[0020] Furthermore, the specific process of the second stage is: modify the loss function of the generator in the original generative adversarial network to the KL divergence of the generated distribution and the ideal distribution of the defined fraudulent data, and train the discriminator; after the training is completed, the pseudo-fraudulent data features generated by the generator are restored from the latent feature space to the original space through the decoder De1.

[0021] Furthermore, the specific process of the third stage is: adding a loss function that minimizes the intra-class distance and maximizes the inter-class distance in the middle layer of the autoencoder, using the pseudo-fraud behavior data as reference data and normal behavior data to train the autoencoder together, and obtaining the feature representation of the normal behavior data through the trained encoder En2.

[0022] Furthermore, the specific process of the fourth stage is: use the features of the normal behavior data extracted in the third stage to train the improved generative adversarial network. The pseudo-abnormal data generated by the generator will gradually gather towards the normal data until it is distributed on its edge. The discriminator is continuously trained to improve its ability to distinguish between normal data and generated pseudo-abnormal data. After the training is completed, the discriminator obtained is used as a fraud detector to detect fraud.

[0023] Furthermore, the loss function of the improved autoencoder is shown in formula (8);

[0024] L AE = l MSE +l latent (8)

[0025] Among them, l MSE is the loss function of the middle layer of the original encoder, as shown in formula (3); latent A loss function of a new objective function is added to the middle layer of the improved autoencoder, as shown in formula (5);

[0026]

[0027]

[0028] Among them, λ1, λ2 and λ3 are hyperparameters. When the model is in the first stage, λ2 and λ3 are 0, and λ1 is greater than 0; when the model is in the third stage, λ1, λ2 and λ3 are all greater than zero; d rf =||c r -c f || represents the center distance between the normal class and the fake fraud class, c r and c f are the normal class X in the latent feature space respectively. r and fake fraud X f The class center, c r and c f As shown in formula (6);

[0029]

[0030] Among them, n and m are the normal class X r and fake fraud X f The number of samples; r and δ f They are normal class X r and fake fraud X f The class variance of samples in feature space, δ r and δ f As shown in formula (7);

[0031]

[0032] Furthermore, the generation method of the improved generative adversarial network is:

[0033] First, we define the ideal distribution p of pseudo-abnormal data features. * As shown in formula (9):

[0034]

[0035] The pseudo-abnormal fraud data features generated Its dimension is the same as that of the middle layer of the autoencoder and belongs to the user feature space B v ; is the probability that the generated data in the feature space belongs to the normal class, ε is a hyperparameter that defines the edge of the normal class in the feature space; C is a constant, and τ is a normalization term;

[0036] The improved generator has a loss function of generating distribution p G and the ideal distribution p * The KL divergence of is defined as shown in formula (10):

[0037]

[0038] Where H(·) is the entropy function, |[·] is the indicator function, which is 1 when the condition is met and 0 otherwise;

[0039] The formula of the generator is shown in formula (11):

[0040]

[0041] Among them, f(·) represents the output of the feature in the middle layer of the discriminator;

[0042] The overall objective function of the improved generator is shown in formula (12):

[0043]

[0044] The pull-away method is used to measure H(p G ) value, the formula is shown in formula (13):

[0045]

[0046] The discriminator takes the normal class features and the generated pseudo-fraud class features as input, and its objective function is shown in formula (14):

[0047]

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The one-class adversarial fraud detection method based on explicit class representation proposed by the present invention is a new one-class model CS-OCAN, which includes an improved autoencoder and a generative adversarial network. In the first stage, we only perform preliminary feature extraction based on normal behavior data, use the autoencoder to extract data robust features, achieve dimensionality reduction for the data, and improve the model's ability to process big data and complex data dimensions; in the second stage, we use the improved generative adversarial network complementary GAN to generate pseudo-abnormal behavior data, which solves the problem that fraud data is difficult to obtain in fraud detection and overcomes the problem of data imbalance; in the third stage, we input normal user behavior data and pseudo-abnormal behavior data into the improved autoencoder for training together, so that the encoding features focus on minimizing the intra-class spacing and maximizing the inter-class distance, improving the effect of feature extraction, and making it more useful for the goal of fraud detection; in the fourth stage, we use the finally extracted normal behavior data features to train the fraud detector: the generator generates samples distributed in the low-density area of ​​the normal class, and the discriminator distinguishes the generated samples from the real samples through continuous training, thereby obtaining the ability to detect fraudulent behavior. In the feature extraction stage of the present invention, we conducted experiments on the UMD Wikipedia dataset and the credit card fraud detection dataset. The experiments showed that the model performance exceeded the existing technical level of a classification method, and the accuracy and stability of fraud detection were significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0051] Figure 1A flow chart of a one-class adversarial fraud detection method based on explicit class representation provided by an embodiment of the present invention. Figure (a) is the preliminary feature extraction in the first stage; Figures (b) and (c) are the generation of pseudo-abnormal behavior data in the second stage; Figure (d) is the final feature extraction; Figures (e) and (f) are the training of the fraud detector in the fourth stage.

[0052] Figure 2 The first stage flow chart provided for the embodiment of the present invention.

[0053] Figure 3 The second stage flow chart provided for the embodiment of the present invention.

[0054] Figure 4 The third stage flow chart provided for the embodiment of the present invention.

[0055] Figure 5 This is a flowchart of the fourth stage provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to better understand the technical solution, the method of the present invention is described in detail below with reference to the accompanying drawings.

[0057] A one-classification deep learning method refers to using a deep learning method or model to learn with only one class (in this method, only normal behavior data) to ultimately achieve the purpose of classification.

[0058] The following is an introduction to the models commonly used in classification deep learning methods:

[0059] (1) Autoencoder

[0060] In 1986, Rumelhart proposed the concept of autoencoder and used it to process high-dimensional complex data, which promoted the development of neural networks. Autoencoder neural network is an unsupervised learning algorithm that is often used to reduce the dimension of data. One of its most common applications is anomaly detection. After data dimension reduction, key features are extracted, making the classifier easier to process and more robust. The autoencoder consists of an encoder En and a decoder De. Suppose the input data X benign ={x1,…,x N}, output data The calculation formula of the intermediate layer features and output data is as follows:

[0061] v=En(x) (1)

[0062]

[0063] The characteristic of the autoencoder is that the dimension of the intermediate layer feature representation is generally smaller than the input and output, and the training goal is to make the output and input as similar as possible. Therefore, the intermediate layer feature representation retains most of the information of the input data. Its loss function is shown in formula (3).

[0064]

[0065] (2) Generative Adversarial Networks

[0066] Generative adversarial network (GAN) is a generative network, which consists of two parts: generator and discriminator. It has a strong data generation capability and can be used to make up for the lack of data. The main function of the generator G is to generate a priori p of the input noise z. z Generate fake samples and learn a distribution p that is consistent with the real sample distribution through training data Matching generated sample distribution p G On the other hand, the function of the discriminator D as a binary classifier is to determine whether an input sample is a real sample x or a generated sample G(z), and output the probability D(·) that this sample belongs to the real sample. The objective function of GAN is formalized as a maximum minimization strategy As shown in formula (4).

[0067]

[0068] The generator and the discriminator are trained alternately according to the objective function, so that the ability of the generator to generate real samples and the ability of the discriminator to judge real and fake samples are getting stronger and stronger. Finally, the entire generative adversarial network reaches a balance. In this state, the distribution of generated samples completely matches the distribution of real samples, that is, p G =p data , the discriminator can no longer distinguish between generated samples and real samples, and its output D(·) is 0.5.

[0069] The detection method (CS-OCAN) proposed in the present invention has a structure diagram as shown in FIG. Figure 1As shown in , there are four stages in total. The first stage is the preliminary feature extraction, which trains the improved autoencoder to learn its feature representation when there is only normal class training data; the second stage is the generation of pseudo-abnormal behavior data. We modify the objective functions of the generator and discriminator in the original generative adversarial network to obtain complementary GAN, and use the features obtained in the first stage to train it to generate data distributed on the edge of the normal sample features, and decode it through the decoder saved in the first stage to obtain pseudo-abnormal behavior data; the third stage is the final feature extraction. After obtaining the pseudo-abnormal behavior data, we will add it to the training data to train the improved autoencoder again, so that the generated features meet the characteristics of compactness within the class and large class spacing. Finally, the encoder after training is used to extract features from the normal behavior data; the fourth stage is the training of the fraud detector. We use the finally extracted normal class features to train the complementary GAN, and the discriminator continuously improves the ability to distinguish generated samples from normal classes, and finally has the ability to detect fraud.

[0070] In this invention, inspired by the Mahalanobis distance, which measures the inter-class separability, we add a new objective function to the middle layer of the autoencoder, so that the autoencoder can meet the requirements of minimizing the intra-class distance and maximizing the inter-class distance during encoding. Its loss function is shown in formula (5), which meets the first characteristic, compact intra-class distance.

[0071]

[0072] Among them, λ1, λ2 and λ3 are hyperparameters. When the model is in the first stage, λ2 and λ3 are 0, and λ1 is greater than 0, that is, only the feature mapping of the normal class is required, and they are expected to be as aggregated as possible; when the model is in the third stage, λ1, λ2 and λ3 are all numbers greater than zero. The feature extraction at this stage takes into account the difference between the normal class and the pseudo-fraud class, increases the spacing between the extracted features of different classes, and reduces the overlap, so as to improve the accuracy and stability of subsequent detection. rf =||c r -c f || represents the center distance between the normal class and the fake fraud class, c r and c f are the normal class X in the latent feature space respectively. r and fake fraud X f Class center.

[0073]

[0074] Among them, n and m are the normal class X r and fake fraud X f The number of samples. r and δf They are normal class X r and fake fraud X f The class variance of samples in feature space.

[0075]

[0076] Then the loss function of the improved autoencoder is shown in formula (8).

[0077] L AE = l MSE +l latent (8)

[0078] About the improved generative adversarial network:

[0079] We use the Complementary GAN model in OCAN to train the fraud detector. We first define the ideal distribution p of pseudo-anomaly data features. * As shown in formula (9).

[0080]

[0081] The pseudo-abnormal fraud data features generated Its dimension is the same as that of the middle layer of the autoencoder and belongs to the user feature space B v ; is the probability that the generated data in the feature space belongs to the normal class, ε is a hyperparameter that defines the edge of the normal class in the feature space; C is a very small constant, and τ is a normalization term. Formula (9) defines the ideal distribution of pseudo-fraud data features: when the point in the feature space enters the normal class range, its value decreases rapidly and is the largest at the edge; when the point in the feature space is distributed outside the normal class range, its value is also very small. Then the loss function of the improved generator is the generated distribution p G and the ideal distribution p * The KL divergence of is defined as follows.

[0082]

[0083] Where H(·) is the entropy function, and |[·] is the indicator function, which is 1 when the condition is met, otherwise it is 0. Since the gradients of the last two terms of formula (10) are almost zero, they can be ignored.

[0084] At the same time, the generator also introduces feature matching loss to ensure that the generated data is in the user feature space B v Here, the formula is as follows.

[0085]

[0086] Among them, f(·) represents the output of the feature in the middle layer of the discriminator. Then the overall objective function of the improved generator is shown in formula (12).

[0087]

[0088] Since the discriminator in a generative adversarial network can determine the possibility that a sample comes from the real distribution, we will train a regular GAN with normal class data as an alternative to estimate before formal model training The following is called the pre-trained network. G )We use the pull-away (PT) method to measure the size of its value, and the formula is as follows.

[0089]

[0090] The discriminator takes the normal class features and the generated fake fraud class features as input and tries to separate the two. Its objective function is as follows.

[0091]

[0092] The first half of formula (14) is the same as the objective function of the discriminator in the standard adversarial generative network. The last term of the formula is a conditional entropy, which is equivalent to adding an additional weight, making the discriminator more inclined to accurately judge the normal class.

[0093] In the improved generative adversarial network, the pseudo data generated by the generator will gradually change from being disorganized to being close to normal data, until all of them are scattered on the edge of the normal data. In this process, the discriminator will define the boundary between normal data and abnormal data more and more clearly and accurately. Ultimately, the discriminator can detect fraud very well.

[0094] The detailed four stages of the one-class adversarial fraud detection method (CS-OCAN) based on class explicit representation of the present invention are as follows.

[0095] The first stage is preliminary feature extraction.

[0096] Specifically, the autoencoder is trained with normal data, and the center loss is added to its middle layer to minimize the intra-class distance. After the training is completed, the original normal data is encoded by the encoder En1 to obtain its features, and the decoder De1 model weights are retained. The first stage training flow chart is as follows Figure 2 shown.

[0097] The second stage is the training of the pseudo-fraud data generator.

[0098] Specifically, we modify the loss function of the generator in the original generative adversarial network to the KL divergence between the generated distribution and the ideal distribution of the defined fraudulent data. The discriminator can distinguish between generated data and real data through training, and also has a certain fraud detection ability, but there is no requirement for its accuracy at this time. After the training is completed, the pseudo-fraud data features generated by the generator are restored from the latent feature space to the original space through the decoder De1. The second stage training flow chart is as follows Figure 3 shown.

[0099] The third stage is the final feature extraction.

[0100] Specifically, the third stage is to extract features again, add a loss function that minimizes the intra-class distance and maximizes the inter-class distance in the middle layer of the autoencoder, use the pseudo-fraud behavior data as reference data and normal behavior data to train the autoencoder together, and obtain the feature representation of the normal behavior data through the trained encoder En2. The third stage training flow chart is shown in Figure 4 shown.

[0101] The fourth stage is the final detector, which is the training of the discriminator in the improved generative adversarial network.

[0102] Specifically, the features of normal behavior data extracted in the third stage are used to train complementary GAN. The pseudo-abnormal data generated by the generator will gradually gather towards the normal data until they are distributed on its edge, while the discriminator is continuously trained to improve its ability to distinguish normal data from the generated pseudo-abnormal data. After the training is completed, the discriminator obtained is used as a fraud detector to detect fraud. The fourth stage training flow chart is shown in the figure below: Figure 5 shown.

[0103] In the present invention, we propose a new classification model CS-OCAN for fraud detection, which includes an improved autoencoder and a generative adversarial network. In the first stage, we only perform preliminary feature extraction based on normal behavior data, use the autoencoder to extract data robust features, achieve dimensionality reduction for the data, and improve the model's ability to process big data and complex data dimensions; in the second stage, we use the improved generative adversarial network complementary GAN to generate pseudo-abnormal behavior data, which solves the problem of fraud data being difficult to obtain in fraud detection and overcomes the problem of data imbalance; in the third stage, we input normal user behavior data and pseudo-abnormal behavior data into the improved autoencoder for training together, so that the encoding features focus on minimizing the intra-class spacing and maximizing the inter-class distance, improving the effect of feature extraction, and making it more useful for the goal of fraud detection; in the fourth stage, we use the finally extracted normal behavior data features to train the fraud detector: the generator generates samples distributed in the low-density area of ​​the normal class, and the discriminator distinguishes the generated samples from the real samples through continuous training, thereby obtaining the ability to detect fraudulent behavior.

[0104] We conducted experiments on the UMD Wikipedia dataset and the credit card fraud detection dataset respectively: we selected an extracted feature and performed 10 different rounds of detection on the fraud detector to obtain the four indicators of fraud class: precision, recall, F1 and ACC. The mean and standard deviation of the calculated test results are shown in Tables 1 and 2. The experimental results of the comparison method are from the literature (Panpan Zheng, Shuhan Yuan, Xintao Wu, Jun Li, and Aidong Lu. 2019. One-class adversarial nets for fraud detection. In AAAI. 1286–1293).

[0105] Table 2 Wikipedia fraud detection results

[0106] Algorithm Name Precision Recall F1 ACC OCNN 0.8314±0.0351 0.8028±0.0476 0.8150±0.0163 0.8181±0.0153 OCGP 0.8381±0.0225 0.8289±0.0374 0.8326±0.0158 0.8337±0.0139 OCSVM 0.6558±0.0058 0.9590±0.0096 0.7789±0.0064 0.7278±0.0080 OCAN 0.9067±0.0615 0.9292±0.0348 0.9010±0.0228 0.8973±0.0244 CS-OCAN 0.9263±0.0139 0.8904±0.0232 0.9076±0.0073 0.9095±0.0059

[0107] Table 2 Credit card fraud detection results

[0108] Algorithm Name Precision Recall F1 ACC OCNN 0.7058±0.1396 0.9390±0.0786 0.7910±0.0608 0.7401±0.1079 OCGP 0.8813±0.1177 0.8566±0.0822 0.8576±0.0417 0.8536±0.0623 OCSVM 0.6547±0.0151 0.9509±0.0101 0.7755±0.0127 0.7245±0.0185 OCAN 0.9067±0.0614 0.8320±0.0319 0.8656±0.0220 0.8701±0.0264 CS-OCAN 0.9368±0.0227 0.8561±0.0333 0.8938±0.0123 0.9086±0.0092

[0109] The above results show that the proposed method has the highest precision, F1 and ACC on the two data sets; in addition, the precision of the proposed method on the fraud class is significantly improved. The experiment shows that the model performance exceeds the existing technical level of the classification method, and the accuracy and stability of fraud detection are significantly improved.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A one-class adversarial fraud detection method based on explicit class representation, characterized in that: It adopts an improved autoencoder and an improved generative adversarial network and includes four stages: Phase 1: Perform preliminary feature extraction based on normal behavior data and use the improved autoencoder to extract data robust features; The second stage: modify the objective functions of the generator and discriminator in the original generative adversarial network to obtain an improved generative adversarial network, and use the improved generative adversarial network to generate pseudo abnormal behavior data; Phase 3: Input normal user behavior data and pseudo-abnormal behavior data into the improved autoencoder for training together, and use the trained encoder to perform final feature extraction on the normal behavior data; The fourth stage: Use the finally extracted normal behavior data features to train the improved generative adversarial network. After the training, the discriminator obtained is used as a fraud detector to detect fraud; The improved autoencoder adds a central loss function to the middle layer of the autoencoder, so that the autoencoder satisfies the minimization of intra-class distance and maximization of inter-class distance during encoding; in the improved generative adversarial network, the loss function of the generator is the KL divergence of the generated distribution and the ideal distribution of the defined fraudulent data.

2. The one-class adversarial fraud detection method based on class explicit representation according to claim 1, characterized in that: The specific process of the first stage is: train the autoencoder with normal data, and add central loss to its middle layer to minimize the intra-class distance; after the training is completed, encode the original normal data through encoder En1 to obtain its features, and retain the decoder De1 model weights.

3. The one-class adversarial fraud detection method based on class explicit representation according to claim 1, characterized in that: The specific process of the second stage is: modify the loss function of the generator in the original generative adversarial network to the KL divergence between the generated distribution and the ideal distribution of the defined fraud data, and train the discriminator; After training is completed, the pseudo-fraud data features generated by the generator are restored from the latent feature space to the original space through decoder De1.

4. The one-class adversarial fraud detection method based on class explicit representation according to claim 1, characterized in that: The specific process of the third stage is: adding a loss function that minimizes the intra-class distance and maximizes the inter-class distance in the middle layer of the autoencoder, using the pseudo-fraud behavior data as reference data and normal behavior data to train the autoencoder together, and obtaining the feature representation of the normal behavior data through the trained encoder En2.

5. The one-class adversarial fraud detection method based on class explicit representation according to claim 1, characterized in that: The specific process of the fourth stage is: use the features of the normal behavior data extracted in the third stage to train the improved generative adversarial network. The pseudo-abnormal data generated by the generator will gradually gather towards the normal data until they are distributed on its edge. The discriminator is continuously trained to improve its ability to distinguish between normal data and generated pseudo-abnormal data. After the training is completed, the discriminator obtained is used as a fraud detector to detect fraud.

6. The one-class adversarial fraud detection method based on class explicit representation according to claim 1, characterized in that: The loss function of the improved autoencoder is shown in formula (8); L AE ＝l MSE +l latent (8) Among them, l MSE is the loss function of the middle layer of the original encoder, as shown in formula (3); latent A loss function of a new objective function is added to the middle layer of the improved autoencoder, as shown in formula (5); Among them, λ1, λ2 and λ3 are hyperparameters. When the model is in the first stage, λ2 and λ3 are 0, and λ1 is greater than 0; when the model is in the third stage, λ1, λ2 and λ3 are all greater than zero; d rf =||c r -c f || represents the center distance between the normal class and the fake fraud class, c r and c f are the normal class X in the latent feature space respectively. r and fake fraud X f The class center, c r and c f As shown in formula (6); Among them, n and m are the normal class X r and fake fraud X f The number of samples; r and δ f They are normal class X r and fake fraud X f The class variance of samples in feature space, δ r and δ f As shown in formula (7); 7. The one-class adversarial fraud detection method based on class explicit representation according to claim 1, characterized in that: The generation method of the improved generative adversarial network is: First, we define the ideal distribution p of pseudo-abnormal data features. * As shown in formula (9): The pseudo-abnormal fraud data features generated Its dimension is the same as that of the middle layer of the autoencoder and belongs to the user feature space B v ; is the probability that the generated data in the feature space belongs to the normal class, ε is a hyperparameter that defines the edge of the normal class in the feature space; C is a constant, and τ is a normalization term; The improved generator has a loss function of generating distribution p G and the ideal distribution p * The KL divergence of is defined as shown in formula (10): Where H(·) is the entropy function, |[·] is the indicator function, which is 1 when the condition is met and 0 otherwise; The formula of the generator is shown in formula (11): Among them, f(·) represents the output of the feature in the middle layer of the discriminator; The overall objective function of the improved generator is shown in formula (12): The pull-away method is used to measure H(p G ) value, the formula is shown in formula (13): The discriminator takes the normal class features and the generated pseudo-fraud class features as input, and its objective function is shown in formula (14):

Images