Anti-fraud model training method based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an anti-fraud model training method based on deep learning, which comprises the following steps: step 1, preprocessing and constructing multi-modal data, carrying out abnormal value processing, missing value filling and feature standardization on original data, representing processed transaction data and geographic information through a feature extraction module, and carrying out feature extraction; the equipment information and the social relation graph are embedded through an embedding layer and a graph neural network; step 2, based on data enhancement of self-supervised learning, pre-training unlabeled data by designing a behavior prediction task and a mask filling task so as to initialize a main model by using generated model parameters, and generating a synthetic sample extension training data set; by designing a behavior prediction task and a mask filling task, automatic learning and enhancement of features are realized by using label-free data, the dependence on labeled data is reduced, the detection capability of the model on fraudulent behaviors is remarkably improved, and the diversity of training data is enriched.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and specifically to an anti-fraud model training method based on deep learning. Background Art

[0002] The rapid development of Internet finance, e-commerce and mobile payment technology has significantly increased the complexity and diversity of fraudulent behaviors, placing higher demands on real-time identification and effective prevention of fraudulent behaviors. Traditional anti-fraud methods mostly rely on rule-based detection systems or simple machine learning models, which are unable to cope with complex and changing fraud methods. In addition, data imbalance problems, the dynamic evolution of fraud patterns and the need for model interpretability and security further increase the challenges of anti-fraud technology.

[0003] In recent years, deep learning technology has demonstrated its feature extraction capabilities and modeling advantages in image processing, natural language processing, and recommendation systems, bringing new ideas to the anti-fraud field. However, existing anti-fraud methods based on deep learning still have shortcomings:

[0004] Fraudulent behavior data is usually extremely scarce. Compared with the massive amount of normal transaction data, the proportion of fraud samples is low, resulting in imbalanced training data. At the same time, it is costly to annotate fraudulent data and difficult to obtain on a large scale, which limits the performance of the model.

[0005] The data sources in anti-fraud scenarios are complex, including transaction records, device information, geographic locations, and social relationships between users. Existing methods cannot fully integrate multimodal data and are difficult to effectively capture the correlation information between different modalities, which reduces the ability to identify complex fraud patterns.

[0006] Fraudulent behavior has the characteristics of dynamic evolution, and new fraud methods emerge in an endless stream. Traditional static models lack a real-time update mechanism and are difficult to quickly adapt to new fraud patterns, which can easily lead to a decrease in recognition accuracy and affect the actual application effect;

[0007] Existing deep learning models are usually "black box models" with poor interpretability of their prediction results, making it difficult to meet the compliance requirements of high-demand fields such as finance. At the same time, the models are not robust enough to adversarial samples and can be easily deceived by malicious attackers by exploiting vulnerabilities, threatening the reliability of anti-fraud systems.

[0008] Therefore, those skilled in the art provide an anti-fraud model training method based on deep learning to solve the above-mentioned problems. Summary of the invention

[0009] In view of the deficiencies in the prior art, the present invention provides an anti-fraud model training method based on deep learning to solve the problems raised in the above background technology.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: a deep learning-based anti-fraud model training method, comprising the following steps:

[0011] Step 1: Preprocess and construct multimodal data. Perform outlier processing, missing value filling and feature standardization on the original data. The processed transaction data and geographic information are represented by the feature extraction module, and the device information and social relationship graph are embedded by the embedding layer and graph neural network.

[0012] Step 2: Data enhancement based on self-supervised learning: pre-training the unlabeled data by designing behavior prediction tasks and mask filling tasks to initialize the main model with the generated model parameters and generate synthetic samples to expand the training data set;

[0013] Step 3: Design a model architecture combining multi-layer deep learning and graph neural network. Use convolutional neural network, embedding layer and graph neural network to extract features of transaction data, device information and social relationship graph, so as to concatenate the output feature vectors of each feature extraction layer into a joint feature representation and reduce the dimension of the fully connected layer.

[0014] Step 4: Online learning and dynamic model updating: incremental training and regular fine-tuning of the model based on real-time data to adapt to changes in data distribution and new fraud patterns;

[0015] Step 5: Design model interpretability and security. Use the Shapley value method to quantify the contribution of input features to the model prediction results, and enhance the adversarial robustness of the model through adversarial training.

[0016] Preferably, in step 1, the original data is processed for outliers using the 3σ detection method, specifically: let the feature f j The mean value is μ j and the standard deviation is σ j , for each data point x ij Requirements:

[0017] |x ij -μ j |>3σ j When x ij are considered as outliers and removed.

[0018] Among them, x ij The feature f j The value in the i-th sample, μ j f j The mean value, σ j f j The standard deviation of .

[0019] Preferably, in step 2, the goal of the behavior prediction task is to achieve self-supervised learning by predicting the next behavior in the transaction data sequence, and its loss function is:

[0020]

[0021] Among them, x ti+1 represents the i+1th transaction behavior, is the i-th transaction behavior, f θ is a self-supervised model, parameter θ is the model parameter, n is the total number of samples, mse is the mean square error loss function.

[0022] Preferably, in step 2, the mask filling task randomly masks some features in the input feature matrix so that the model predicts the masked feature values, and its loss function is:

[0023]

[0024] Among them, x j is the masked feature, g θ (X) is the model's predicted value for the input feature matrix X, parameter θ is the model parameter, n is the total number of samples, mask is the mean square error loss function.

[0025] Preferably, in step 3, a convolutional neural network is used to extract time series features of transaction data, and a convolution operation is performed on the transaction sequence through a sliding window, and the features of each time step are expressed as:

[0026]

[0027] in, is the feature representation of the i-th time step, W c is the convolution kernel weight, * indicates the convolution operation, b c is the bias, is the input feature of the i-th time step, and ReLU is the rectified linear unit activation function.

[0028] Preferably, in step 3, the graph neural network extracts features from the social relationship graph, and uses a graph sampling and aggregation algorithm to generate an embedded feature representation of the node, and the calculation formula is:

[0029]

[0030] in, is the embedding representation of node v at layer k, is the set of neighbor nodes of node v, is the previous layer embedding representation of neighbor node u, σ is the activation function, W kand b k are weights and biases.

[0031] Preferably, in step 3, the joint feature representation is performed by concatenating the transaction data feature vector, the device information embedding vector and the social network embedding feature vector, specifically:

[0032] H = Concat(h CNN ,h Emb ,h GNN ),

[0033] Among them, h CNN is the convolution feature vector of transaction data, h Emb is the embedded feature vector of device information, h GNN is the graph embedding feature vector of the social network, and H is the joint feature representation after concatenation.

[0034] Preferably, in step 4, online learning updates model parameters through real-time data, and the parameters are incrementally updated using the gradient descent method, and the update formula is:

[0035]

[0036] Among them, θ t is the current model parameter, θ t+1 is the updated model parameter, α is the learning rate, y t is the true label, is the model prediction value, is the loss function.

[0037] Preferably, in step 5, the contribution of the input features to the model prediction results is quantified by the Shapley value method, and the calculation formula of the Shapley value is:

[0038]

[0039] Among them, φ i represents the contribution of feature i, f(S) represents the output of the model using feature set S, S is the feature set, N is the full set of features, f(S∪{i}) represents the model output after adding feature i to feature subset S, |S| is the number of elements in feature subset S, |N| is the number of elements in the full set of features, |S|! and (|N|-|S|-1)! represent the number of permutations of feature subset S and the number of permutations of the remaining features.

[0040] Preferably, in step 5, adversarial training is used to improve the adversarial robustness of the model and generate adversarial samples X adv , add it to the training set, and the loss function is:

[0041] ∫adv =∫(f θ (X adv ),y),

[0042] Among them, X adv is an adversarial sample, y is the corresponding label, and f θ is the model prediction function, parameter θ is the model parameter, and ∫ is the loss function.

[0043] The present invention provides an anti-fraud model training method based on deep learning. It has the following beneficial effects:

[0044] 1. The present invention designs behavior prediction tasks and mask filling tasks, uses unlabeled data to realize automatic learning and enhancement of features, reduces dependence on labeled data, and significantly improves the model's ability to detect fraudulent behavior. At the same time, the generated model parameters are used to initialize the main model, providing a good learning starting point for the model, enriching the diversity of training data, and enabling the model to enhance its ability to capture fraud patterns.

[0045] 2. The present invention combines convolutional neural networks, embedding layers and graph neural networks to extract features and model transaction data, device information and social relationship graphs, effectively integrating the advantages of time series features, discrete features and graph structure data, supporting unified modeling of multimodal data, fully exploring the correlation and complexity of fraudulent behaviors, retaining key information, and improving the performance of the model in complex fraud scenarios.

[0046] 3. Through incremental training and regular fine-tuning of real-time data, the model of the present invention can dynamically adjust parameters to adapt to changes in data distribution, and update the model online in combination with new labeled data, thereby improving the model's adaptability to new fraud patterns, ensuring the timeliness and accuracy of the model, and enabling it to cope with rapidly changing fraud methods.

[0047] 4. The present invention generates adversarial samples and conducts adversarial training to improve the robustness of the model against abnormal inputs and malicious data, enhance the stability of the model in the face of adversarial attacks, reduce the possibility of fraudsters attacking the system through forged or malicious data, and provide technical support for reliable operation in harsh and complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Figure of the present invention. DETAILED DESCRIPTION

[0049] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0050] The present invention is described in detail below in conjunction with the accompanying drawings:

[0051] Example:

[0052] Please refer to the attached Figure 1 The embodiment of the present invention provides an anti-fraud model training method based on deep learning, comprising the following steps:

[0053] Step 1: Preprocess and construct multimodal data. Perform outlier processing, missing value filling and feature standardization on the original data. The processed transaction data and geographic information are represented by the feature extraction module, and the device information and social relationship graph are embedded by the embedding layer and graph neural network.

[0054] Step 2: Data enhancement based on self-supervised learning: pre-training the unlabeled data by designing behavior prediction tasks and mask filling tasks to initialize the main model with the generated model parameters and generate synthetic samples to expand the training data set;

[0055] Step 3: Design a model architecture combining multi-layer deep learning and graph neural network. Use convolutional neural network, embedding layer and graph neural network to extract features of transaction data, device information and social relationship graph, so as to concatenate the output feature vectors of each feature extraction layer into a joint feature representation and reduce the dimension of the fully connected layer.

[0056] Step 4: Online learning and dynamic model updating: incremental training and regular fine-tuning of the model based on real-time data to adapt to changes in data distribution and new fraud patterns;

[0057] Step 5: Design model interpretability and security. Use the Shapley value method to quantify the contribution of input features to the model prediction results, and enhance the adversarial robustness of the model through adversarial training.

[0058] By processing outliers, filling missing values, and standardizing features, we can eliminate noise and defects in the data and ensure the integrity and consistency of the input data. The feature extraction modules of transaction data and geographic information can effectively capture the temporal and spatial characteristics. The embedding layer of device information and orthogonal relationship graphs and the graph neural network can extract discrete features and related information in the network structure, providing the model with rich and unified input features to support subsequent training.

[0059] By designing behavior prediction tasks and mask filling tasks, we can make full use of unlabeled data, improve the model's ability to learn transaction behavior patterns and feature associations while reducing annotation costs, and use the generated model parameters to initialize the main model, providing a good parameter starting point. The expanded synthetic samples alleviate the data imbalance problem, enhance the model's ability to identify scarce fraud samples, and provide more comprehensive support for subsequent model training.

[0060] Through the combination of convolutional neural networks, embedding layers, and graph neural networks, we design specialized feature extraction methods for different modal data to ensure full learning of time series features, discrete features, and network structure features. The joint feature representation combines the key information of multiple data modes through the concatenation of multimodal features and the dimensionality reduction of the fully connected layer, providing accurate input from multiple angles for fraud detection and adapting to complex fraud scenarios.

[0061] Through incremental training and regular fine-tuning of real-time data, the model can dynamically adjust parameters to adapt to changes in data distribution and the emergence of new fraud patterns. The online learning mechanism ensures that the model is continuously updated under real-time requirements, while regular fine-tuning adjusts feature distribution by integrating new and old data, so that the model can maintain efficient detection capabilities for a long time and adapt to the rapid changes in fraudulent behavior.

[0062] Through the Shapley value method, the contribution of input features to the model prediction results is quantified, making the model's prediction process more transparent and providing users with an intuitive explanation of the model's output results. Adversarial training generates and learns adversarial samples to enhance the model's robustness under malicious attacks and abnormal data, improve the system's stability and reliability in identifying fraudulent behavior, and ensure the security of the model in practical applications.

[0063] In step 1, the 3σ detection method is used to process the outliers of the original data. Specifically, let the feature f j The mean value is μ j and the standard deviation is σ j , for each data point x ij Requirements:

[0064] |x ij -μ j |>3σ j When x ij are considered as outliers and removed.

[0065] Among them, x ij The feature f j The value in the i-th sample, μ j f j The mean value, σ j f j The standard deviation of .

[0066] By using the 3σ detection method to screen and eliminate outliers in the original data, the noise data that deviates from the normal distribution range can be effectively removed, thereby improving the quality and reliability of the data. The 3σ detection method sets clear boundary conditions based on the statistical distribution characteristics of the data, and marks extreme values ​​that exceed 3 times the standard deviation of the mean as abnormal. It is particularly suitable for processing data with normal distribution or close to normal distribution. It can avoid the interference of outliers on parameter estimation during model training and reduce the deviation caused by outliers. At the same time, it retains the integrity of most normal samples and provides stable and reliable basic data input for subsequent feature extraction and model training.

[0067] In step 2, the goal of the behavior prediction task is to achieve self-supervised learning by predicting the next behavior in the transaction data sequence, and its loss function is:

[0068]

[0069] in, represents the i+1th transaction behavior, is the i-th transaction behavior, f θ is a self-supervised model, parameter θ is the model parameter, n is the total number of samples, mse is the mean square error loss function.

[0070] The behavior prediction task predicts the next behavior of the transaction data sequence, utilizes the inherent temporal correlation in the time series, constructs an effective self-supervised learning objective, learns the dynamic pattern of transaction behavior on unlabeled data, avoids dependence on labeled data, and at the same time, expands the training data range of the model so that the model can accurately quantify the difference between the predicted value and the true value, thereby optimizing the model parameters.

[0071] In step 2, the mask filling task randomly masks some features in the input feature matrix so that the model predicts the masked feature values. Its loss function is:

[0072]

[0073] Among them, x j is the masked feature, g θ (X) is the model's predicted value for the input feature matrix X, parameter θ is the model parameter, n is the total number of samples, mask is the mean square error loss function.

[0074] The mask filling task randomly masks some features in the input feature matrix and trains the model to restore the masked feature values, so that the model can learn the potential correlation between features, make full use of unlabeled data, extract information from different feature interaction patterns, and make up for the problem of insufficient data annotation. The mean square error loss function is used to measure the error between the predicted value and the true value, so that the model can gradually optimize its ability to restore masked features during the training process. By randomly masking features, the model can adapt to the diversity of input data and improve the prediction ability of missing values ​​or some abnormal data, providing a more comprehensive and stable feature representation for subsequent main model training.

[0075] In step 3, the convolutional neural network is used to extract the time series features of the transaction data. The transaction sequence is convolved through a sliding window, and the features of each time step are expressed as:

[0076]

[0077] in, is the feature representation of the i-th time step, W c is the convolution kernel weight, * indicates the convolution operation, b c is the bias, is the input feature of the i-th time step, and ReLU is the rectified linear unit activation function.

[0078] By extracting the time series features of transaction data through convolutional neural networks and using sliding windows to perform convolution operations on transaction sequences, local patterns and short-term dependencies in time series data can be efficiently captured. Convolution operations extract features on time series by sharing weights, reducing the number of parameters, improving computational efficiency and generalization capabilities, and introducing bias and modified linear unit activation functions to enhance the model's ability to express nonlinear features, allowing the model to adapt to complex transaction behavior patterns.

[0079] In step 3, the graph neural network extracts features from the social relationship graph and uses the graph sampling and aggregation algorithm to generate the embedded feature representation of the node. The calculation formula is:

[0080]

[0081] in, is the embedding representation of node v at layer k, is the set of neighbor nodes of node v, is the previous layer embedding representation of neighbor node u, σ is the activation function, W k and b k are weights and biases.

[0082] The social relationship graph is extracted through graph neural networks, and the graph sampling and aggregation algorithm is used to generate the embedded feature representation of the node, which can effectively capture the relationship between the node and its neighbor nodes. The feature information of the neighbor nodes is used to generate the embedded feature of the current node through aggregation operations and activation functions, and the graph structure information is integrated into the node representation. This allows the model to flexibly adapt to different graph structures and feature distributions, effectively modeling complex relationship patterns in social networks, and at the same time reducing the computational complexity of processing large-scale graph data.

[0083] In step 3, the joint feature representation is achieved by concatenating the transaction data feature vector, the device information embedding vector, and the social network embedding feature vector, specifically:

[0084] H = Concat(h CNN ,h Emb ,h GNN ),

[0085] Among them, h CNN is the convolution feature vector of transaction data, h Emb is the embedded feature vector of device information, h GNN is the graph embedding feature vector of the social network, and H is the joint feature representation after concatenation.

[0086] By splicing the convolutional feature vector of transaction data, the embedded feature vector of device information and the graph embedded feature vector of social network to generate a joint feature representation, the feature information of multimodal data can be fully integrated. Various feature vectors represent the data characteristics of the corresponding modality. The splicing operation retains the important information of each modal data. At the same time, it unifies them into the input representation of the same dimension, providing a complete multimodal feature perspective for subsequent models.

[0087] In step 4, online learning updates model parameters through real-time data, and uses the gradient descent method to incrementally update the parameters. The update formula is:

[0088]

[0089] Among them, θ t is the current model parameter, θ t+1 is the updated model parameter, α is the learning rate, y t is the true label, is the model prediction value, is the loss function.

[0090] Online learning updates model parameters through real-time data and uses gradient descent for incremental updates. This enables the model to gradually adjust its own parameters while continuously receiving new data. By calculating the loss function in real time, the model parameters are optimized according to the difference between the true label and the model prediction value, ensuring that the model can adapt to the latest fraud patterns and data distribution changes. The incremental update formula avoids the high computational cost of retraining the model each time new data is received. At the same time, the update amplitude is controlled by the learning rate to prevent overfitting or insufficient updating. The online learning mechanism provides dynamic adaptability for fraud detection, so that the model can respond to changes in the data stream in a timely manner in practical applications and effectively deal with evolving fraud behaviors.

[0091] In step 5, the contribution of the input features to the model prediction results is quantified by the Shapley value method. The calculation formula of the Shapley value is:

[0092]

[0093] Among them, φ i represents the contribution of feature i, f(S) represents the output of the model using feature set S, S is the feature set, N is the full set of features, f(S∪{i}) represents the model output after adding feature i to feature subset S, |S| is the number of elements in feature subset S, |N| is the number of elements in the full set of features, |S|! and (|N|-|S|-1)! represent the number of permutations of feature subset S and the number of permutations of the remaining features.

[0094] The Shap Ley value method is used to quantify the contribution of input features to the model prediction results and accurately evaluate the importance of each feature in the model decision. The Shap Ley value is based on cooperative game theory. By allocating contributions to feature combinations and calculating the marginal contribution of features in the entire feature set, the fairness and uniqueness of contribution allocation are ensured.

[0095] The advantage of this method is that it can provide a clear explanation for the model prediction results. Especially in high-risk scenarios such as finance and fraud detection, explainability is an important indicator of model compliance. By quantifying the contribution of features, it helps users understand the decision-making basis of the model and enhances the transparency of the model. In addition, the Shap l ey value can guide feature selection, identify the features that contribute most to the prediction results, and optimize the model input, further improving model performance and computational efficiency. It is suitable for complex multimodal data scenarios and can provide important support for subsequent decision-making and risk assessment of anti-fraud systems.

[0096] In step 5, adversarial training is used to improve the adversarial robustness of the model and generate adversarial samples X adv , add it to the training set, and the loss function is:

[0097] ∫adv =∫(f θ (X adv ),y),

[0098] Among them, X adv is an adversarial sample, y is the corresponding label, and f θ is the model prediction function, parameter θ is the model parameter, and ∫ is the loss function.

[0099] By generating adversarial samples and adding them to the training set, and using adversarial training to optimize the model's loss function, the model's robustness in the face of malicious inputs or abnormal data can be improved. Adversarial samples are generated by adding subtle perturbations to the original samples, which can simulate model vulnerabilities exploited by attackers. By introducing adversarial samples during the training process, the model can learn how to correctly respond to abnormal inputs, improve stability and reliability in complex environments, effectively reduce the interference of malicious attacks on the model's prediction results, and enhance the model's defense capabilities in anti-fraud scenarios. Especially in the case of frequent adversarial attacks, adversarial training can expand the model's learning scope, adapt it to a more diverse data distribution, and further improve the accuracy and consistency of predictions, providing important technical guarantees for the practical application of anti-fraud systems in harsh and complex environments.

[0100] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based anti-fraud model training method, characterized in that: The steps include: Step 1: Preprocess and construct multimodal data. Perform outlier processing, missing value filling and feature standardization on the original data. The processed transaction data and geographic information are represented by the feature extraction module, and the device information and social relationship graph are embedded by the embedding layer and graph neural network. Step 2: Data enhancement based on self-supervised learning: pre-training the unlabeled data by designing behavior prediction tasks and mask filling tasks to initialize the main model with the generated model parameters and generate synthetic samples to expand the training data set; Step 3: Design a model architecture combining multi-layer deep learning and graph neural network. Use convolutional neural network, embedding layer and graph neural network to extract features of transaction data, device information and social relationship graph, so as to concatenate the output feature vectors of each feature extraction layer into a joint feature representation and reduce the dimension of the fully connected layer. Step 4: Online learning and dynamic model updating: incremental training and regular fine-tuning of the model based on real-time data to adapt to changes in data distribution and new fraud patterns; Step 5: Design model interpretability and security. Use the Shapley value method to quantify the contribution of input features to the model prediction results, and enhance the adversarial robustness of the model through adversarial training.

2. According to the anti-fraud model training method based on deep learning according to claim 1, it is characterized in that: In step 1, the original data is processed for outliers using the 3σ detection method, specifically: let the feature f j The mean value is μ j and the standard deviation is σ j , for each data point x ij Requirements: |x ij -μ j |>3σ j When x ij are considered as outliers and removed. Among them, x ij The feature f j The value in the i-th sample, μ j f j The mean value, σ j f j The standard deviation of .

3. According to the anti-fraud model training method based on deep learning according to claim 1, it is characterized in that: In step 2, the goal of the behavior prediction task is to achieve self-supervised learning by predicting the next behavior in the transaction data sequence, and its loss function is: in, represents the i+1th transaction behavior, is the i-th transaction behavior, f θ is a self-supervised model, parameter θ is the model parameter, n is the total number of samples, is the mean square error loss function.

4. According to the anti-fraud model training method based on deep learning according to claim 1, it is characterized in that: In step 2, the mask filling task randomly masks some features in the input feature matrix so that the model predicts the masked feature values, and its loss function is: Among them, x j is the masked feature, g θ (X) is the model's predicted value for the input feature matrix X, parameter θ is the model parameter, n is the total number of samples, mask is the mean square error loss function.

5. According to the anti-fraud model training method based on deep learning in claim 1, it is characterized in that: In step 3, the convolutional neural network is used to extract the time series features of the transaction data, and the transaction sequence is convolved through a sliding window to express the features of each time step as follows: in, is the feature representation of the i-th time step, W c is the convolution kernel weight, * indicates the convolution operation, b c is the bias, is the input feature of the i-th time step, and ReLU is the rectified linear unit activation function.

6. According to the deep learning-based anti-fraud model training method of claim 1, it is characterized in that: In step 3, the graph neural network extracts features from the social relationship graph and uses a graph sampling and aggregation algorithm to generate an embedded feature representation of the node. The calculation formula is: in, is the embedding representation of node v at layer k, is the set of neighbor nodes of node v, is the previous layer embedding representation of neighbor node u, σ is the activation function, W k and b k are weights and biases.

7. The anti-fraud model training method based on deep learning according to claim 1 is characterized in that: In step 3, the joint feature representation is performed by concatenating the transaction data feature vector, the device information embedding vector and the social network embedding feature vector, specifically: H=Concat(h CNN ,h Emb ,h GNN ), Among them, h CNN is the convolution feature vector of transaction data, h Emb is the embedded feature vector of device information, h GNN is the graph embedding feature vector of the social network, and H is the joint feature representation after concatenation.

8. The anti-fraud model training method based on deep learning according to claim 1, characterized in that: In step 4, online learning updates model parameters through real-time data, and uses the gradient descent method to incrementally update the parameters. The update formula is: Among them, θ t is the current model parameter, θ t+1 is the updated model parameter, α is the learning rate, y t is the true label, is the model prediction value, is the loss function.

9. The anti-fraud model training method based on deep learning according to claim 1, characterized in that: In step 5, the contribution of the input features to the model prediction results is quantified by the Shapley value method. The calculation formula of the Shapley value is: Among them, φ i represents the contribution of feature i, f(S) represents the output of the model using feature set S, S is the feature set, N is the full set of features, f(S∪{i}) represents the model output after adding feature i to feature subset S, |S| is the number of elements in feature subset S, |N| is the number of elements in the full set of features, |S|! and (|N|-|S|-1)! represent the number of permutations of feature subset S and the number of permutations of the remaining features.

10. The method for predicting power grid load based on machine learning according to claim 1, characterized in that: In step 5, adversarial training is used to improve the adversarial robustness of the model and generate adversarial samples X adv , add it to the training set, and the loss function is: ∫ adv =∫(f θ (X adv ),y), Among them, X adv is an adversarial sample, y is the corresponding label, and f θ is the model prediction function, parameter θ is the model parameter, and ∫ is the loss function.

Images