Continuous financial fraud detection method and system based on multi-modal data fusion

Abstract

The invention discloses a continuous financial fraud detection method and system based on multi-modal data fusion, and belongs to the technical field of financial fraud detection.The method comprises the steps that features of different modal financial transaction samples are extracted, and the features of the different modal samples are aligned; generating a reconstructed sample of the transaction sample based on the auto-encoder; and training the detection model by using the reconstructed sample and the new data, and then executing financial fraud detection. By aligning the features of different modal samples, the complementary information of different modal data can be effectively obtained, so that the fraud detection accuracy is improved. New data is adopted to train the detection model, so that the model can absorb unknown data types in real time; meanwhile, the reconstructed sample can supplement the feature coverage range of the current multi-modal financial transaction sample, and the reconstructed sample and the new data are adopted to train the model, so that the problem that the model forgets historical features due to only attention to the new data can be avoided, and the problem of privacy leakage due to reuse of old data can be prevented.

Description

Technical Field

[0001] The present invention relates to the technical field of financial fraud detection, and in particular to a continuous financial fraud detection method and system based on multimodal data fusion. Background Art

[0002] In the field of fraud detection in financial markets, existing methods face the dual challenges of diversified data sources and dynamic changes in the market environment. It is difficult to fully utilize the complementary information of different modal data and adapt to the rapid fluctuations of financial markets. Traditional prediction models usually statically model single modal data and ignore the potential correlation between multimodal data, such as historical transaction records, news texts, market sentiment and social media information. This not only leads to one-sided prediction results, but also fails to capture the key driving forces in the dynamic changes of the market. In recent years, although deep learning methods have been applied to fraud detection to automatically learn high-dimensional features in financial markets and adapt to nonlinear relationships, most of them rely on static data training and fail to fully consider the high dynamics and non-stationary characteristics of financial markets. As a result, the detection model is still difficult to adapt to the addition of new data. The model prediction performance is significantly reduced when processing new data. At the same time, there is also the problem of privacy leakage caused by reusing old data. Summary of the invention

[0003] The purpose of the present invention is to overcome the problems of the prior art and provide a continuous financial fraud detection method and system based on multimodal data fusion.

[0004] The purpose of the present invention is to achieve the following technical solution: a method for detecting persistent financial fraud based on multimodal data fusion, the method comprising the following steps: S1: Obtain multimodal financial transaction samples, including text data, time series data, and graph structure data; S2: Extract the features of financial transaction samples of different modalities and align the features of financial transaction samples of different modalities; S3: Reconstructed samples of financial transaction samples generated based on autoencoders; S4: Use reconstructed samples and newly added financial transaction samples to train the detection model; S5: Use the trained detection model to perform financial fraud detection and output the detection results.

[0005] In one example, after extracting the features of the financial transaction samples of different modalities, the multimodal feature space is optimized by sharing a comparative learning objective to achieve alignment processing of the features of the financial transaction samples of different modalities.

[0006] In one example, the objective loss function of contrastive learning is The expression is: ; in, represents the total number of samples; Represents positive samples from different modalities; Represents the similarity measurement function between features; represents the temperature coefficient; represents a negative sample pair; , All are serial number labels.

[0007] In one example, the autoencoder includes an encoder and a decoder, the encoder extracts features of an input financial transaction sample, and the decoder generates a reconstructed sample of the financial transaction sample based on the extracted features of the financial transaction sample.

[0008] In one example, after the step of generating the original financial transaction sample based on the autoencoder, the process further includes: Perform clustering processing on the reconstructed samples to obtain clustering results that appear as spherical regions in the feature space; The centroid of the spherical region is used as the representative sample of all reconstructed samples in the current spherical region; The detection model is trained using a dataset consisting of representative samples and newly added financial transaction samples.

[0009] In one example, the calculation expression of the centroid is: ; in, represents the cluster centroid; Represents the clustering result set; is the number of reconstructed samples in the clustering results; Indicates the clustering result A reconstruction sample.

[0010] In one example, the use of a data set consisting of representative samples and newly added financial transaction samples to train the detection model includes: Introducing a loss function based on representative samples of cluster centroids Train the detection model, loss function The expression is: ; in, represents the training loss of newly added financial transaction samples; is the regularization coefficient; Indicates the number of clustering results; is a serial number label; Represents the detection model, which is used to extract features and predict input data; It represents the representative sample after the cluster centroid processing; , represent the current and historical model parameters respectively.

[0011] It should be further explained that the technical features corresponding to the various examples of the above method can be combined or replaced with each other to form a new technical solution.

[0012] The present invention also includes a continuous financial fraud detection system based on multimodal data fusion, the system comprising interconnected terminals and servers, the server comprising interconnected storage servers and training servers; The terminal accesses the storage server and inputs the multimodal financial transaction sample into the storage server; The training server is used to execute steps S1-S4 of the method formed by any one of the above examples or a combination of multiple examples; The storage server is used to save the trained detection model; The terminal and / or the training server and / or the storage server performs financial fraud detection using the trained detection model and outputs the detection result.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. In one example, by extracting the features of financial transaction samples of different modalities and aligning the features of financial transaction samples of different modalities, it is possible to better adapt to changes in data of different modalities and improve the generalization ability on new data; at the same time, by aligning the features of samples of different modalities, multimodal data fusion processing is realized, which can effectively obtain complementary information of data of different modalities, enable the model to have a more comprehensive understanding of the characteristics of financial transaction samples, and then more accurately identify abnormal transaction behaviors, thereby improving the accuracy of fraud detection.

[0014] Using newly added financial transaction samples (new data) to train the detection model can enable the model to absorb unknown data types in real time, thereby improving the prediction performance when processing new data; at the same time, the reconstructed samples can supplement the feature coverage of the current multimodal financial transaction samples. Using reconstructed samples and new data to train the model can not only avoid the model forgetting historical features due to focusing only on new data, but also prevent privacy leakage caused by reusing old data (historical financial transaction samples).

[0015] 2. In one example, contrastive learning can bring the feature representations of related samples (positive samples) in different modalities closer together, while at the same time distance the representations of unrelated samples. The resulting multimodal embedding space can not only effectively fuse heterogeneous data, but also be sensitive to the volatility characteristics of the financial market.

[0016] 3. In one example, through clustering centroid processing, without modifying the characteristics of the real sample (original multimodal financial transaction sample), the originally scattered and changeable single reconstructed sample is converted into a more stable and representative centroid point, so that the reconstructed sample as a whole can more easily form a clear boundary with the real sample in the feature space, thereby significantly improving the model's ability to distinguish between real and reconstructed samples, and maximally protecting the real sample characteristics from identification and leakage risks, thereby effectively protecting the privacy security of the real sample.

[0017] 4. In one example, a loss function based on representative samples of cluster centroids is introduced to train the model, and a regularization constraint is added to the loss function. The model's attention to past knowledge can be adjusted by adjusting the regularization coefficient, that is, increasing the regularization coefficient can make the model pay more attention to past knowledge, and vice versa, the model will pay more attention to the learning of new knowledge. Through this mechanism, the model can effectively retain the memory of historical data patterns while adapting to changes in the dynamic financial market, thereby improving the stability and accuracy of predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The specific implementation methods of the present invention are further described in detail below in conjunction with the accompanying drawings. The accompanying drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The same reference numerals are used in these drawings to represent the same or similar parts. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0019] Figure 1 A method flow chart provided for an example of the present invention; Figure 2 An autoencoder architecture diagram provided for an example of the present invention; Figure 3 A comparative learning framework diagram provided for an example of the present invention; Figure 4 A system framework diagram provided for an example of the present invention.

[0020] In the figure: 101 - first terminal; 102 - second terminal; 103 - server. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] In the description of the present invention, ordinal numbers (e.g., "first and second", etc.) are used to distinguish objects, but are not limited to the order, and cannot be understood as indicating or implying relative importance. Unless otherwise clearly specified and limited, "connected" should be understood in a broad sense, for example, it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0023] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0024] In one example, if Figure 1 As shown, a method for detecting persistent financial fraud based on multimodal data fusion comprises the following steps:

[0025] S1: Obtain multimodal financial transaction samples.

[0026] Among them, multimodal financial transaction samples include transaction-related text data, (transaction) time series data, and graph structure data, such as consumption records, bank loan records, securities trading records, etc., which can reflect transaction time, transaction location, transaction frequency, transaction account, etc.

[0027] S2: Extract features of financial transaction samples of different modalities and align the features of financial transaction samples of different modalities, that is, use inter-modal self-supervised distillation to jointly train the extracted embedded features and learn cross-modal deep feature representation.

[0028] In order to fully explore the potential correlation between features between modalities, inter-modal self-supervised distillation is used to construct a multimodal embedding space, thereby achieving effective data fusion. Specifically, the present invention performs separate feature extraction on text data, time series data (digital information), and graph structure data. Among them, for text data, existing text analysis models, such as BERT model, Transformer model, etc., are used to convert sentences or articles into a series of features that can reflect their meaning; for digital information, a simple neural network, such as a fully connected neural network, can be used to pre-process the numerical data (such as data range adjustment), and then extract features for subsequent data fusion processing through linear calculation and appropriate nonlinear transformation. For graph structure data, an image recognition network is used to convert the image into a set of digital features that can reflect the content of the picture, such as a convolutional neural network, a graph convolutional neural network, etc. Through the above-mentioned feature extraction method, different types of data can be processed into similar expression formats, which is convenient for mapping the features of different modalities into a unified feature space during subsequent feature alignment processing. For example, the features of different modalities can be transformed through neural network layers (such as fully connected layers) to match the dimensions and scales of features of different modalities. The loss function can be further used to optimize the feature alignment processing effect to achieve effective fusion of multimodal data, thereby making full use of the complementary information of different modal data, enhancing the model's ability to understand market behavior, and thus improving the accuracy of the prediction model.

[0029] Preferably, before extracting the features of the financial transaction samples, data preprocessing may be performed, such as data cleaning, data normalization, data enhancement, etc.

[0030] S3: Reconstructed samples of financial transaction samples generated based on autoencoders.

[0031] like Figure 2 As shown in the figure, the autoencoder is an unsupervised learning neural network model, including an encoder and a decoder. The encoder is used to compress the input data (financial transaction samples) into a low-dimensional latent space, and the decoder is used to reconstruct the original data and generate reconstructed data (reconstructed samples) similar to the input data.

[0032] S4: Use reconstructed samples and newly added financial transaction samples to train the detection model for this round.

[0033] The method of the present invention is a continuous dynamic financial fraud detection method based on continuous learning. Specifically, continuous learning refers to training a model on a data stream of a task sequence, and the training goal is to train the model to achieve good performance on all learned tasks. Each task has its own separate training set, validation set and test set. The model can only access the data of the current round of training tasks during training. Figure 3 As shown, samples corresponding to different tasks , and tags , Different, so it is necessary to retrain the generator and solver for each task, so that reconstructed samples can be generated in the next task.

[0034] Preferably, when the model is trained, the model parameters are recorded to guide subsequent correction tasks. Further, after the model training is completed, the trained model is corrected, including:

[0035] While the model is predicting, combined with real-time dynamic data input, the loss function changes of the model input and the predicted value are calculated, and the input feature weights are adjusted according to real-time market dynamics to ensure the adaptability of the prediction to the dynamic environment, and the target model is recalibrated using the latest data and the adjusted loss function to output the corrected financial market fraud detection value. Then, the model is calibrated online using the latest data and the adjusted loss function.

[0036] In step S4, the detection model of this round is trained using the reconstructed samples and the newly added financial transaction samples, wherein the reconstructed samples can supplement the feature coverage of the current multimodal financial transaction samples. At the same time, the use of reconstructed samples to train the model can also prevent the privacy leakage caused by reusing old data (historical financial transaction samples). This is because the reconstructed samples are reconstructed based on the deep features of the extracted original financial transaction samples. These features are abstract representations after model processing, rather than the specific content of the original data. Therefore, the reconstructed samples reflect more of the statistical characteristics of the data rather than the specific information of the individual, which greatly reduces the risk of privacy leakage. Furthermore, the newly added financial transaction samples (new data) can provide the detection model with the latest data features and patterns, enabling the model to learn and adapt to unknown data types in real time. By using these new data to train the model, the prediction performance of the model when processing new data is significantly improved. For example, in bank fraud detection, new data may contain new fraud behavior patterns, and the trained model can better identify these newly emerging fraud behaviors. In addition, during the training process, using reconstructed samples and new data at the same time can prevent the model from only focusing on new data and forgetting historical features, thus solving the problem of catastrophic forgetting in continuous learning.

[0037] S5: Use the detection model that has completed this round of training to perform financial fraud detection and output the detection results.

[0038] Preferably, after completing step S5, it is determined whether the model has reached an update cycle. If so, it returns to step S4 and enters the next round of model training to optimize the performance of the detection model to adapt to dynamically changing financial fraud detection scenarios.

[0039] In this example, cross-modal features are extracted through inter-modal self-supervised distillation technology, and reconstructed samples of financial transaction samples are generated using autoencoders to achieve playback enhancement of market historical states. In the dynamic market observation process, knowledge distillation technology is used to align the features of financial transaction samples of different modalities, effectively alleviating the problem of catastrophic forgetting and ensuring the continuity and stability of the detection model during data updates. This method can significantly improve the real-time and accuracy of financial market fraud detection, is suitable for dynamic financial transaction detection scenarios, and provides an efficient and intelligent solution for financial fraud detection.

[0040] In one example, after extracting features from financial transaction samples of different modalities, the multimodal feature space is optimized by sharing a unified contrastive learning objective, thereby achieving alignment of features of financial transaction samples of different modalities. Contrastive learning is an unsupervised learning method that aims to learn effective feature representations of data by comparing similarities and differences between samples. The core idea of ​​contrastive learning is to make similar samples closer in the feature space and dissimilar samples farther apart. In contrastive learning, the model is trained by constructing positive sample pairs (similar samples, such as different enhanced versions of the same image) and negative sample pairs (dissimilar samples, such as different images), thereby achieving high-quality feature extraction on unlabeled data and providing a good basic representation for downstream tasks (such as classification or clustering). Optionally, the objective loss function of contrastive learning is used to measure the consistency of representations between different modalities, using It means that the definition is as follows: ; in, represents the total number of samples; Represents positive samples from different modalities; Represents the similarity measurement function between features; represents the temperature coefficient; represents a negative sample pair; , All are serial number labels.

[0041] In one example, in the dynamic data processing part, in order to cope with the dynamic changes in data distribution in the financial market, a feature-level joint generation adjustment mechanism is introduced, hereinafter referred to as the adjustment mechanism. The adjustment mechanism is mainly implemented based on the autoencoder, in which the encoder extracts the features of the input original financial transaction samples, and the decoder generates the reconstructed samples of the original financial transaction samples based on the features of the extracted financial transaction samples. The adjustment mechanism can not only effectively extract the deep features in the financial transaction samples, but also generate reconstructed samples similar to the input financial transaction samples.

[0042] Furthermore, the training objective of the autoencoder is to minimize the reconstruction error, using Indicates that the loss function is expressed as follows: ; in, is the input financial transaction sample, are the reconstructed samples (pseudo samples) of the decoder's reconstructed output.

[0043] Through the above-mentioned adjustment mechanism based on the autoencoder, not only can the feature distribution of the current market be learned, but also in subsequent training, when new data needs to be updated, the autoencoder generates reconstructed samples based on the stored features. These reconstructed samples are played back into the model training to supplement the feature coverage of the current data. This not only avoids the model forgetting historical features due to focusing only on new data, but also prevents privacy leakage caused by reusing old data.

[0044] Optionally, regularization methods and modularization methods can be used to replace the sample replay method, so that the model can remember important sample information of previous tasks and avoid catastrophic forgetting of the model during continuous learning.

[0045] In one example, in order to further improve the distinguishability between the reconstructed samples and the original samples, a fuzzy processing mechanism based on clustering of reconstructed samples is introduced. Specifically, after the reconstructed sample step of generating the original financial transaction sample based on the autoencoder, it also includes: a. Perform clustering processing on the reconstructed samples to obtain clustering results that appear as spherical regions in the feature space.

[0046] Specifically, the generated reconstructed samples are regarded as a set of data points in the feature space, and these data points are clustered, and each cluster can be regarded as a spherical region in the feature space.

[0047] b. Use the centroid of the spherical region as the representative sample of all reconstructed samples in the current spherical region.

[0048] The centroid of the sphere (i.e., the center of the sphere) is used as the representative point of the reconstructed sample, that is, the centroid represents all the reconstructed samples under the cluster category. This is because the centroid can more stably reflect the central characteristics of the cluster area, thereby reducing the impact of the characteristic fluctuation of a single reconstructed sample on model training.

[0049] Preferably, the calculation expression of the centroid is: ; in, represents the cluster centroid; Represents the clustering result set; is the number of reconstructed samples in the clustering results; Indicates the clustering result A reconstruction sample.

[0050] c. Use a dataset consisting of representative samples and newly added financial transaction samples to train the detection model.

[0051] In one example, to further enhance the effect of joint generation, regularization constraints were added when training new data to ensure that the model retains the ability to adapt to fraud detection in the old market while learning new market trend features. Specifically, a loss function based on representative samples of cluster centroids was introduced. Train the detection model, loss function The expression is: ; in, represents the training loss of newly added financial transaction samples; Indicates the number of clustering results; Represents the detection model, which is used to extract features and predict input data; It represents the representative sample after the cluster centroid processing; , denote the current and historical model parameters, respectively; is the regularization coefficient, by adjusting The size of the adjustment model pays attention to past knowledge, that is: increase This enables the model to pay more attention to the knowledge learned in previous tasks, and vice versa, the model will pay more attention to the learning of new knowledge. Through this mechanism, the model can effectively retain the memory of historical data patterns while adapting to changes in the dynamic financial market, thereby improving the stability and accuracy of predictions.

[0052] Combining the above examples, a preferred example of the present invention is obtained, in which the method comprises the following steps: S10: Obtain multimodal financial transaction samples; S20: extracting features of financial transaction samples of different modalities, and aligning the features of financial transaction samples of different modalities; S30: Generate a reconstruction sample of a financial transaction sample based on the autoencoder; S40: clustering the reconstructed samples to obtain clustering results presented as spherical regions in the feature space; using the centroid of the spherical region as a representative sample of all reconstructed samples in the current spherical region; using the data set consisting of the representative samples and the newly added financial transaction samples to train the detection model, and introducing a loss function of the representative samples based on the cluster centroid during the model training process; S50: Execute financial fraud detection using the trained detection model and output the detection result.

[0053] This method realizes efficient fusion of financial market fraud-related data through dynamic multimodal learning, and dynamically adapts to changes in the market environment by using a feature-level joint generation adjustment mechanism, thereby constructing a fraud detection model with market volatility adaptability. This method can accurately identify abnormal transaction patterns in a multimodal environment, improve the sensitivity and robustness of fraud detection, and provide effective support for financial risk control. After completing the two-stage training, the constructed model is the target fraud detection model of this embodiment. The model can detect fraudulent behavior in the market based on the multimodal financial data provided by the user, and return the corresponding prediction results. When new data arrives, multimodal information will be integrated in real time, and model parameters will be dynamically adjusted to ensure that the detection results always meet the latest market environment, thereby maintaining continuous and efficient fraud detection capabilities.

[0054] The present invention also includes a continuous financial fraud detection system based on multimodal data fusion, such as Figure 4 As shown, the system includes terminals and a server 103, the terminals include a first terminal 101 and a second terminal 102, the server 103 includes a storage server and a training server, and the first terminal 101 and the second terminal 102 are both connected to the storage server and the training server.

[0055] In this example, the first terminal may be a personal user terminal, an enterprise terminal, or other third-party data source terminal, and the first terminal may be a desktop computer, a smart phone, a tablet computer, a laptop computer, etc., but is not limited thereto. The first terminal uploads the modal financial transaction sample to the training server and / or storage server through a network interface, providing a basis for subsequent data processing and model training.

[0056] The storage server and the training server can be independent physical servers or server clusters. Preferably, the storage server is used to receive and save the multimodal financial transaction samples uploaded by the first terminal, and pre-process and classify and store them, while providing efficient data reading support for the training server. The training server trains the dynamic financial market fraud detection model based on the multimodal financial transaction samples (including newly added financial transaction samples) in the storage server to generate a high-precision market fraud detection model, that is, the training server is used to execute steps S1-S4 and steps S10-S40 of the method of the present invention. Optionally, the training server can be equipped with a Linux operating system and high-performance GPU computing resources to meet the needs of large-scale data processing and complex model training.

[0057] Furthermore, the storage server saves the trained detection model for access by the first terminal and the second terminal. At this time, the terminal and / or the training server and / or the storage server use the detection model to perform financial fraud detection and output the detection results.

[0058] The second terminal includes a device for receiving and displaying the detection results. The user can access the detection model stored in the server through the second terminal and input the financial transaction sample into the detection model to obtain the detection result. The second terminal can be a desktop computer, a smart phone, a tablet computer, a laptop computer, etc., but is not limited thereto.

[0059] Through the synergy of the various devices in the above system, multimodal financial data can be efficiently processed and financial market fraud detection results can be dynamically generated. It is worth noting that the system supports simultaneous access by users of multiple second terminals, which can meet the diverse needs of different users.

[0060] In summary, the system of the present invention first obtains financial transaction data of different modalities, and then fuses the data of different modalities. Specifically, through the inter-modal self-supervised distillation technology, the fusion modeling of multi-modal data features is carried out to improve the comprehensive understanding of market status; at the same time, the feature-level joint generation adjustment mechanism is introduced, so that the model can not only absorb unknown data types in real time, but also dynamically adjust the impact of outdated information on predictions, ensuring the stability and accuracy of the prediction results. This method not only improves the prediction accuracy, but also reasonably optimizes the allocation of computing resources, providing an efficient and flexible solution for coping with complex and changing financial markets.

[0061] The above specific implementation methods are detailed descriptions of the present invention. It cannot be determined that the specific implementation methods of the present invention are limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions and substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for detecting persistent financial fraud based on multimodal data fusion, characterized in that: The following steps are involved: S1: Obtain multimodal financial transaction samples, including text data, time series data, and graph structure data; S2: Extract the features of financial transaction samples of different modalities and align the features of financial transaction samples of different modalities; S3: Reconstructed samples of financial transaction samples generated based on autoencoders; S4: Use reconstructed samples and newly added financial transaction samples to train the detection model; S5: Use the trained detection model to perform financial fraud detection and output the detection results.

2. The method for detecting persistent financial fraud based on multimodal data fusion according to claim 1, characterized in that: After extracting the features of the financial transaction samples of different modalities, the multimodal feature space is optimized by sharing a comparative learning objective to achieve alignment processing of the features of the financial transaction samples of different modalities.

3. The method for detecting persistent financial fraud based on multimodal data fusion according to claim 2 is characterized in that: Objective loss function for contrastive learning The expression is: ； in, represents the total number of samples; Represents positive samples from different modalities; Represents the similarity measurement function between features; represents the temperature coefficient; represents a negative sample pair; , All are serial number labels.

4. The method for detecting persistent financial fraud based on multimodal data fusion according to claim 1, characterized in that: The autoencoder includes an encoder and a decoder. The encoder extracts features of an input financial transaction sample, and the decoder generates a reconstructed sample of the financial transaction sample based on the extracted features of the financial transaction sample.

5. The method for detecting persistent financial fraud based on multimodal data fusion according to claim 1, characterized in that: After the step of reconstructing samples based on the autoencoder to generate the original financial transaction samples, the method further includes: Perform clustering processing on the reconstructed samples to obtain clustering results that appear as spherical regions in the feature space; The centroid of the spherical region is used as the representative sample of all reconstructed samples in the current spherical region; The detection model is trained using a dataset consisting of representative samples and newly added financial transaction samples.

6. The method for detecting persistent financial fraud based on multimodal data fusion according to claim 5, characterized in that: The calculation expression of the centroid is: ； in, represents the cluster centroid; Represents the clustering result set; is the number of reconstructed samples in the clustering results; Indicates the clustering result A reconstruction sample.

7. The method for detecting persistent financial fraud based on multimodal data fusion according to claim 5, characterized in that: When the detection model is trained using a data set consisting of representative samples and newly added financial transaction samples, it includes: Introducing a loss function based on representative samples of cluster centroids Train the detection model, loss function The expression is: ； in, represents the training loss of newly added financial transaction samples; is the regularization coefficient; Indicates the number of clustering results; is a serial number label; Represents the detection model, which is used to extract features and predict input data; It represents the representative sample after the cluster centroid processing; , represent the current and historical model parameters respectively.

8. A continuous financial fraud detection system based on multimodal data fusion, characterized in that: The system includes interconnected terminals and servers, and the servers include interconnected storage servers and training servers; The terminal accesses the storage server and inputs the multimodal financial transaction sample into the storage server; The training server is used to perform steps S1-S4 in the method according to any one of claims 1-7; The storage server is used to save the trained detection model; The terminal and / or the training server and / or the storage server performs financial fraud detection using the trained detection model and outputs the detection result.

Images