Anti-money laundering detection method based on improved condition table generative adversarial network

Abstract

The invention belongs to the field of artificial intelligence, and relates to an anti-money laundering detection method based on an improved condition table generative adversarial network, which comprises the following steps: filtering a majority class sample set according to a minority class sample set to obtain a filtered majority class sample set; constructing a CTGAN model, and training the CTGAN model according to the minority class sample set; generating a candidate synthesis sample set by using the current CTGAN model, judging whether the candidate synthesis sample set meets requirements or not, and if yes, obtaining a trained CTGAN model; otherwise, adjusting the current CTGAN model and training according to the minority class sample set; generating a minority class synthesis sample set by using the trained CTGAN model, and combining the filtered majority class sample set, minority class sample set and minority class synthesis sample set to obtain a balanced training set; training an anti-money laundering detection model according to the balance training set; according to the method, the neighborhood cleaning rule is combined with the CTGAN, so that the high fidelity of the generated sample is ensured.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and relates to an anti-money laundering detection method based on an improved conditional table generative adversarial network. Background Technology

[0002] With the rapid development of mobile payment and e-commerce, mobile payment has become one of the mainstream payment tools globally. However, the annual economic losses caused by various illegal transactions are enormous, and the methods are showing a trend of "intelligence, concealment, and temporalization." Illegal activities are difficult to identify based on the characteristics of a single transaction and require analysis of long-term transaction sequences.

[0003] Existing methods for detecting illegal transactions mainly suffer from the following technical bottlenecks:

[0004] (1) The problem of extreme imbalance in data categories: In general transactions, normal transactions usually account for more than 99.5%, while illegal transactions account for less than 0.5%. In traditional processing methods, random oversampling ( Imbalance is addressed by replicating minority class samples, but this can lead to model overfitting; random undersampling ( Deleting majority class samples can easily lead to the loss of key characteristics of normal transactions (such as users’ special consumption patterns). Synthetic sampling methods generate minority class samples through interpolation, but the generated samples tend to fall into the majority class feature space, exacerbating class overlap; traditional Although it can generate synthetic samples, it has poor adaptability to tabular data (transaction data is mostly in tabular format, including continuous amounts and discrete merchant types), and does not consider the class overlap problem, resulting in low quality of synthetic samples.

[0005] (2) Overlapping Samples Interfere with Model Learning: There are many overlapping areas in the feature spaces of normal transactions and money laundering transactions—for example, normal travel spending by users and illegal operations by money launderers are highly similar in features such as "transaction location" and "amount." Such overlapping samples make it difficult for the model to learn clear classification boundaries, resulting in low recall (missed detection) for money laundering transactions or high false detection (false blocking) for normal transactions. Existing methods lack targeted overlapping sample cleaning mechanisms and only alleviate the problem through subsequent model optimization (such as adjusting the weights of the loss function), which has limited effect.

[0006] (3) Insufficient capture of long-term dependencies in time-series transaction data: The essence of violations is "time-series anomalies"—money laundering transactions usually violate users' long-established consumption habits. Traditional machine learning methods ( Random forests and logistic regression only analyze the isolated features of single transactions and cannot capture the temporal correlation of transaction sequences; in deep learning methods, , Although it can process time series data, it suffers from the gradient vanishing problem (when the sequence length exceeds 20, the gradient propagation decays severely), making it difficult to capture long-term dependencies of more than 10 transactions. Although it can extract local temporal features, its receptive field is fixed and cannot flexibly adapt to the differences in transaction frequency among different users (e.g., some users trade 3 times a day, while some users trade 1 time a week).

[0007] Therefore, there is an urgent need for a method to generate illegal transaction data that can simultaneously address the issues of "data imbalance, overlapping categories, and insufficient data volume" in order to meet the risk control requirements of high accuracy, low false positives, and real-time performance. Summary of the Invention

[0008] To address the aforementioned problems in the prior art, this invention employs an anti-money laundering detection method based on an improved conditional table generative adversarial network, comprising: acquiring user transaction data, inputting the user transaction data into a trained anti-money laundering detection model, and obtaining detection results; the training process of the anti-money laundering detection model includes:

[0009] S1. Obtain the transaction dataset, preprocess each transaction sample in the transaction dataset to obtain the training set;

[0010] S2. Divide the training set into majority class sample sets. With minority class sample set Based on the minority class sample set For the majority class sample set Filtering is performed to obtain the filtered majority class sample set. ;

[0011] S3. Construct the CTGAN model According to the minority class sample set For CTGAN model The CTGAN model is obtained through training. CTGAN is a conditional table-based generative adversarial network.

[0012] S4. Utilize the current CTGAN model Generate and minority class sample sets Given an equal number of candidate synthetic sample sets, determine whether the quality of the candidate synthetic sample sets meets the requirements; where t is the number of iterations for adjusting the CTGAN model.

[0013] S5. If the quality of the candidate synthetic sample set meets the requirements, then the current CTGAN model will be... This serves as the final trained CTGAN model; otherwise, the current CTGAN model is adjusted. The structure was used to obtain the adjusted CTGAN model. And return to step S3 to adjust the CTGAN model. Conduct training;

[0014] S6. Generate a synthetic minority class sample set using the finally trained CTGAN model. Merge and filter the majority class sample set Minority class sample set and minority class synthetic sample sets To obtain a balanced training set ;

[0015] S7. Based on the balanced training set Train the anti-money laundering detection model to obtain a well-trained anti-money laundering detection model.

[0016] Beneficial effects:

[0017] 1. This invention filters and removes nearest-neighbor overlapping samples of the minority class within the majority class, reducing feature space overlap, eliminating singularities that cause the WGAN-GP gradient penalty term to fail, ensuring the discriminator learns a continuous and smooth classification boundary, avoiding mode collapse, and enabling CTGAN to learn the real money laundering feature distribution to generate high-fidelity synthetic samples; 2. This invention constructs a model structure adaptive mechanism based on KL divergence feedback, calculating the KL divergence between the synthetic sample set and the real minority class sample set in real time during the generation process, automatically triggering topological reconstruction of the generator network structure until the generated distribution converges. This solves the problem that traditional static CTGAN networks cannot adapt to different data complexity scenarios; 3. The discriminator of this invention adopts a multi-sample batch input structure. The discriminator needs to determine whether the entire batch is a synthetic sample, reducing the discriminator's sensitivity to noise in a single sample, improving training stability, and solving the problem of synthetic sample distribution offset. Attached Figure Description

[0018] Figure 1 A flowchart of an anti-money laundering detection method based on an improved condition table generative adversarial network provided in an embodiment of the present invention;

[0019] Figure 2 The training flowchart of the CTGAN model provided in the embodiment of the present invention is shown. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, this invention employs an anti-money laundering detection method based on an improved conditional table generative adversarial network, comprising: acquiring transaction data, inputting the transaction data into a trained anti-money laundering detection model, and obtaining detection results; the training process of the anti-money laundering detection model includes:

[0022] S1. Obtain the transaction dataset, and preprocess each transaction sample in the transaction dataset to obtain the training set and test set;

[0023] The transaction dataset of this invention uses existing transaction datasets, and each transaction sample includes the following fields: transaction serial number, card number, transaction timestamp, transaction amount, merchant code, merchant type, transaction location (latitude and longitude), and transaction channel (…). The system includes 10 fields: account balance, time interval from the last transaction, etc.

[0024] Preprocessing for each transaction sample in the transaction dataset includes:

[0025] S11. Normalize the continuous features of each transaction sample in the transaction dataset, and perform one-hot encoding on the discrete features of each transaction sample in the transaction dataset to obtain the pre-processed transaction dataset.

[0026] The magnitudes of continuous features in transaction data vary greatly (e.g., the range of "transaction amount" is 0.01). 100,000 yuan, with a "time interval" range of 0. If the time interval is 86400 seconds, directly inputting it into the model will cause the model to overemphasize large-scale features (such as monetary amounts) and ignore small but important features (such as time intervals). This invention uses Z-Score normalization to map each continuous feature to the interval [-1, 1], as shown in the formula: ;in, This is the mean of the feature in the dataset. The standard deviation is used. One-hot encoding is performed on discrete features to convert them into binary features.

[0027] For example, the average transaction amount of a user Yuan, standard deviation A transaction of 1100 yuan, after normalization, is... Values ​​exceeding the range [-1, 1] need to be truncated (values ​​greater than 1 are set to 1, and values ​​less than -1 are set to -1) to avoid the impact of outliers on the model.

[0028] S12. Train a gradient boosting decision tree using the pre-processed transaction dataset. Calculate the split gain of each feature in the pre-processed transaction dataset based on the trained gradient boosting decision tree. Normalize the split gain of each feature to obtain the importance of each feature in the pre-processed transaction dataset.

[0029] Some features in transaction data are irrelevant to anti-money laundering detection, which increases model complexity and introduces noise. This invention uses Gradient Boosting Decision Tree (GBDT) to evaluate feature importance; the number of Gradient Boosting Decision Trees is 150 (to balance model accuracy and complexity), the maximum tree depth is 8 (to prevent noisy fitting caused by excessively deep decision trees), and the training learning rate is 0.05 (to avoid overfitting).

[0030] S13. Select features with importance greater than a preset threshold from the pre-processed transaction dataset to obtain the pre-processed transaction dataset.

[0031] The preset threshold is 0.03, which was determined through validation set testing. Features with scores below 0.03 improve model accuracy by less than 0.5%.

[0032] S14. Divide the preprocessed transaction dataset into a training set and a test set.

[0033] To ensure the model's generalization ability, a stratified random partitioning strategy is adopted (maintaining the ratio of money laundering samples in the training and test sets consistent with the original data), dividing the training and test sets in an 8:2 ratio. This avoids situations where the random partitioning results in too few money laundering samples in the test set, making it impossible to accurately evaluate the model's performance.

[0034] S2. Divide the training set into majority class sample sets. With minority class sample set Based on the minority class sample set For the majority class sample set Filtering is performed to obtain the filtered majority class sample set. ;

[0035] Based on minority class sample set For the majority class sample set Filtering includes:

[0036] S21. Initialize the empty set R;

[0037] S22. For each minority class sample Calculate the Euclidean distance between each sample and all other samples in the preprocessed transaction dataset, and select the first... The nearest samples form the minority class samples. nearest neighbor sample set ;

[0038] The Euclidean distance is:

[0039] ,

[0040] Where K is the number of key features after filtering. Samples , The value of the k-th feature.

[0041] The value of N is determined through 5-fold cross-validation: tests are performed on the validation set. Time model Select the optimal value (experiments show that...). It performs best on most datasets, covering sufficient neighborhood information while avoiding the introduction of too many irrelevant samples.

[0042] S23. For each minority class sample Iterate through its nearest neighbor sample set Each sample in ,like Then the sample Add to collection ;

[0043] Such samples Samples that are highly similar to money laundering samples in the feature space will interfere with the model's learning of money laundering patterns and need to be labeled as overlapping samples and added to the set R.

[0044] S24, From the majority class sample set Delete all items belonging to the set From the samples, we obtain the filtered majority class sample set. .

[0045] Experiments show that this step can be removed. 5% 10% overlap samples, making and The feature space overlap is reduced by 40%. 60%.

[0046] In this method, the majority class sample set filtering operation has a crucial indirect impact on the CTGAN model training. Although CTGAN is trained directly using minority class samples, the unfiltered majority class samples contain overlapping samples that are highly similar to the minority class, which pollutes the minority class feature space. This leads to CTGAN learning a pseudo-money laundering distribution mixed with normal transaction features. The filtering operation uses Euclidean distance to filter and remove the nearest neighbor overlapping samples of the minority class in the majority class, reducing the feature space overlap by 40%~60%, eliminating singularities that cause the WGAN-GP gradient penalty term to fail, ensuring that the discriminator learns a continuous and smooth classification boundary, avoiding mode collapse, and enabling CTGAN to learn the real money laundering feature distribution and generate high-fidelity synthetic samples. This lays the foundation for subsequent balanced training set construction and anti-money laundering detection model training.

[0047] S3. Construct the CTGAN model According to the minority class sample set For CTGAN model The CTGAN model is obtained through training. CTGAN is a conditional table-based generative adversarial network.

[0048] like Figure 2 As shown, the CTGAN model includes: an improved generator and a discriminator; the improved generator includes: a pattern-specific normalization layer, a conditional embedding layer, and a generator; the adjusted CTGAN model... The training includes:

[0049] S31, The minority class sample set The samples were divided into multiple batches, resulting in multiple subsets of the minority class. Where b is the index of the batch;

[0050] S32. Subset the minority class samples from each batch. Continuous feature input model The pattern-specific normalization layer yields a subset of minority class samples for each batch. Multimode parameters of continuous features;

[0051] The pattern-specific normalization layer includes a Gaussian Mixture Model (GMM), which is used to normalize the minority class subset of samples. The distribution is fitted to each continuous feature, thereby dividing each continuous feature into 3 to 5 patterns (e.g., amount can be divided into "small amount (0)"). 100 yuan) "Medium amount (100 yuan)" The system offers three patterns: “1000 yuan”, “large amount (above 1000 yuan)”, and “large amount”. Each pattern corresponds to independent Gaussian component parameters (mean and variance). This approach prevents the loss of patterns in long-tailed distributions (a very small number of ultra-large amount money laundering transactions) by traditional CTGAN.

[0052] S33. Subset the minority class samples from each batch. Discrete feature input model The conditional embedding layer obtains a subset of minority class samples for each batch. The conditional vector of discrete features;

[0053] S34. Obtain the batch random noise Z, and combine the batch random noise Z with the minority class sample subset of each batch. Conditional vectors of discrete features and multi-mode parameter input models of continuous features The generator produces a subset of minority class samples for each batch. The synthetic sample set;

[0054] The random noise here is batch random noise Z, not a single noise vector. Its quantity is consistent with the number of samples in the minority class subset of each batch, ensuring that each sample can incorporate unique noise during the generation process, improving the diversity of the synthesized samples and avoiding the generator outputting homogeneous samples.

[0055] The generator consists of a fully connected network; the generator handles random noise Z and a minority class subset of samples in batch b. The processing of conditional vectors for discrete features and multi-mode parameters for continuous features includes:

[0056] Combine the batch random noise Z with the minority class sample subset of batch b. The conditional vectors of discrete features are fused; the fused result is then combined with the minority class sample subset of batch b. The multi-mode parameters of continuous features are input into a fully connected network to obtain a subset of minority class samples in batch b. The synthetic sample set.

[0057] S35. Input the minority class sample subset and its composite sample set from each batch into the model. The discriminator obtains the discrimination results of the minority class sample subset and the synthetic sample set for each batch;

[0058] The discriminator uses The multi-sample batch input structure: the number of samples in each batch is m=4~8 (instead of 1 in traditional GAN). The discriminator needs to determine whether the entire batch is a synthetic sample, which reduces the discriminator's sensitivity to noise in a single sample and improves training stability.

[0059] S36. Calculate the loss function value based on the discrimination results of the minority class sample subsets and the composite sample set of all batches, and update the model based on the loss function value. The parameters are set such that the trained CTGAN model is obtained when the loss function value is minimized. .

[0060] use The loss function introduces a gradient penalty term to avoid the pattern collapse problem of traditional GANs (where the generator only generates a few types of samples). The loss function values ​​include:

[0061]

[0062] in, Let Z be the minority class subset of batch b, and Z be the batch random noise. For minority class sample set minority class sample subset The true distribution For the distribution of batch random noise, For the distribution of the interpolated sample set, As weight, For minority class sample set The gradient penalty term ensures the Lipschitz gradient of the discriminator is continuous between the interpolated sample set and its synthetic sample set.

[0063] S4. Utilize the current CTGAN model Generate and minority class sample sets Given an equal number of candidate synthetic sample sets, determine whether the quality of the candidate synthetic sample sets meets the requirements; where t is the number of iterations for adjusting the CTGAN model.

[0064] Determining whether the quality of the candidate synthetic sample set meets the requirements includes:

[0065] Calculate the candidate synthetic sample set and the minority class sample set The proportion of overlapping samples is calculated to determine the ratio between the candidate synthetic sample set and the minority class sample set. The KL divergence is used to determine the quality of the candidate synthetic sample set. If the proportion of overlapping samples is greater than the preset overlap threshold or the KL divergence is greater than the preset divergence threshold, the quality of the candidate synthetic sample set is determined to be unacceptable. Otherwise, it meets the requirements.

[0066] Preferably, the overlap threshold is 15%~20% and the divergence threshold is 0.1.

[0067] Calculating the overlap ratio between the candidate synthetic sample set and the minority class sample set includes:

[0068] For each synthetic sample in the candidate synthetic sample set Calculate its relationship with the minority class sample set. Each sample Euclidean distance Where u and v are the indices of the candidate synthetic sample set and the minority class sample set, respectively;

[0069]

[0070] Where K is the total number of features. (These are the k-th feature values ​​of the u-th synthetic sample in the candidate synthetic sample set and the v-th sample in the minority class sample set, respectively).

[0071] Preset distance threshold (After 5-fold cross-validation, the value is taken as 0.1~0.3), if the synthetic sample Each sample in the minority class sample set The Euclidean distances are all less than or equal to the distance threshold. If so, it is determined to be an overlapping sample;

[0072] Count the number of overlapping samples in the candidate synthetic sample set Calculate the number of overlapping samples Total number of samples in the candidate synthetic sample set The ratio of the two samples is used to obtain the proportion of overlapping samples. .

[0073] S5. If the quality of the candidate synthetic sample set meets the requirements, then the current CTGAN model will be... This serves as the final trained CTGAN model; otherwise, the current CTGAN model is adjusted. The structure was used to obtain the adjusted CTGAN model. And return to step S3 to adjust the CTGAN model. Conduct training;

[0074] Adjust the current CTGAN model The structure includes:

[0075] If the candidate synthetic sample set and the minority class sample set If the proportion of overlapping samples is greater than a preset threshold, then NCL cleaning is performed on the candidate synthetic sample set, and the NCL-cleaned candidate synthetic sample set and minority class sample set are calculated. The KL divergence is used to adjust the current CTGAN model. The structure; otherwise, based on the candidate synthetic sample set and the minority class sample set. Adjusting the KL divergence of the current CTGAN model The structure.

[0076] The core function of the NCL (Neighborhood Cleaning Rule) cleaning step here is to accurately remove overlapping samples in the candidate synthetic sample set that are highly similar in features to minority class samples. This avoids these "repeated / redundant samples" interfering with the accuracy of KL divergence calculation. Without cleaning, overlapping samples will cause the KL divergence to be too small, leading to a misjudgment that the distribution of synthetic samples has converged. Cleaning allows the calculated KL divergence to more accurately reflect the actual distribution difference between the synthetic sample set and the minority class sample set, providing an accurate basis for subsequent CTGAN model structure adjustments (such as adding or removing hidden layer neurons). This ensures that the model can be optimized in a targeted manner to generate samples that are more in line with real money laundering characteristics and are free of redundancy.

[0077] NCL cleaning of the candidate synthetic sample set includes:

[0078] Identifying candidate synthetic sample sets and minority class sample sets For overlapping samples, remove all overlapping samples from the candidate synthetic sample set to obtain the NCL-cleaned candidate synthetic sample set.

[0079] The specific steps include:

[0080] For each synthetic sample in the candidate synthetic sample set Calculate its relationship with the minority class sample set. Each sample The Euclidean distance;

[0081]

[0082] Where K is the total number of features. (These are the k-th feature values ​​of the u-th synthetic sample in the candidate synthetic sample set and the v-th sample in the minority class sample set, respectively).

[0083] Preset distance threshold (After 5-fold cross-validation, the value is taken as 0.1~0.3), if the synthetic sample Each sample in the minority class sample set The Euclidean distances are all less than or equal to the distance threshold. If the sample is an overlapping sample, then it is considered an overlapping sample.

[0084] Adjust the current CTGAN model based on KL divergence. The structure includes:

[0085] Determine whether the candidate synthetic sample set has feature homogenization; specifically including:

[0086] For each continuous feature, calculate its distribution concentration index, including the proportion of the largest / second largest pattern and the proportion of the standard deviation; if the proportion of the largest / second largest pattern is ≥3 and the proportion of the standard deviation is <50% of the real sample, then it is judged to meet the standard, that is, there is homogenization;

[0087] For each discrete feature, calculate its Shannon diversity index H. If H < 0.5 × lnk, it is considered to meet the standard, that is, there is homogenization; where k is the number of categories of the discrete feature.

[0088] If there is one qualifying feature in both the continuous and discrete features, then the candidate synthetic sample set is determined to have feature homogenization.

[0089] If the KL divergence is slightly higher than the preset divergence threshold (e.g., exceeding the threshold by 10%-30%), it indicates that the model is not capturing the distribution details well enough. The number of neurons in the generator hidden layer can be increased by 20% to improve the feature expression ability while avoiding excessive increase in complexity.

[0090] If the KL divergence is significantly higher than the divergence threshold (e.g., exceeding the threshold by more than 30%) and the candidate synthetic sample set does not exhibit feature homogenization, it indicates that the model is difficult to fit the complex patterns of minority class distributions. In this case, the number of hidden layer neurons in the generator needs to be increased by 50% to enhance the generator's ability to learn multi-dimensional feature associations.

[0091] If the KL divergence is significantly higher than the divergence threshold (e.g., exceeding the threshold by more than 30%) and the candidate synthetic sample set exhibits feature homogeneity, then the activation function of the generator's hidden layer should be replaced first, replacing ReLU with LeakyReLU (with a slope set to 0.2). This approach mitigates the "neuron death" problem by preserving the negative half-axis gradient information, enhances the generator's ability to learn from samples at the edge of the distribution, and ensures that the adjusted model can generate synthetic samples that more closely resemble the true distribution of the minority class.

[0092] S6. Generate a minority class sample set using the finally trained CTGAN model. Equal number of minority class synthetic sample sets Merge and filter the majority class sample set Minority class sample set and minority class synthetic sample sets To obtain a balanced training set ;

[0093] Generate a synthetic minority class sample set using the final trained CTGAN model. This includes: generating a candidate synthetic sample set using the final trained CTGAN model, performing NCL cleaning on the candidate synthetic sample set, calculating the number of samples in the NCL-cleaned candidate synthetic sample set, and determining whether the number is equal to the minority class sample set. If the number of samples is equal to the number of samples in the final synthetic money laundering sample set, then the final synthetic money laundering sample set is obtained. Otherwise, the final trained CTGAN model is used to regenerate a synthetic sample set with the missing number of samples. This is then combined with the candidate synthetic sample set cleaned by NCL and the regenerated synthetic sample set to obtain the combined minority class synthetic sample set. Continuous features in the combined minority class synthetic sample set are mapped back to the original data volume. For discrete features, the softmax function is used to determine the class with the highest probability to obtain the final discrete feature value, thus yielding the final minority class synthetic sample set. ;

[0094] The core function of the final step is to transform the "model-adaptive abstract features" generated by CTGAN into "business-usable data." This includes: denormalizing continuous features using multi-mode parameters to restore the original magnitude and business semantics of features such as transaction amount and time interval, avoiding feature failure caused by magnitude distortion; and using softmax to select the category with the highest probability for discrete features, which can transform the probability vector into specific business labels such as "transaction channel," eliminating feature ambiguity and ensuring that the feature format and distribution logic of the synthetic samples are consistent with those of the real samples, thus clearing format obstacles for subsequent processes.

[0095] majority class sample set and ∪ The sample size ratio is strictly controlled at 1:1±0.1 to avoid new class imbalances.

[0096] S7. Based on the balanced training set Train the anti-money laundering detection model to obtain a well-trained anti-money laundering detection model.

[0097] The anti-money laundering detection model is the TCN model (Temporal Convolutional Network). The TCN model captures the long-term dependencies of transaction sequences through its unique network structure, achieving high-precision anti-money laundering detection.

[0098] The TCN model consists of an input layer, a hidden layer, and an output layer;

[0099] Input layer: Receives a balanced training set A time-series transaction sequence, each sequence containing 5 consecutive transactions from a user. Feature data of 10 transactions.

[0100] Hidden layers: Multiple dilated causal convolutions are used, with the dilation factor d increasing in powers of 2, and the kernel size set to 3. 5; Each dilated causal convolutional layer is followed by a weight normalization layer (to accelerate training convergence). Activation function layer (introducing nonlinearity) Layer (drop probability 0.1) 0.3 (to prevent overfitting); adjacent hidden layers reuse features through residual connections to avoid gradient vanishing.

[0101] Output layer: using The activation function outputs the probability that a single transaction is money laundering. If the probability is greater than a preset threshold (such as 0.5), it is determined to be a money laundering transaction.

[0102] use Optimizer (learning rate 1e-4) 1e-3), binary cross-entropy loss function, with a batch size of 32 128. Training rounds: 50 200 training The model prevents overfitting through an early stopping mechanism (such as stopping training if the loss increases for 5 consecutive rounds of test set training);

[0103] Input the test set after training is complete. The model outputs the money laundering probability of each transaction and generates the final prediction result based on a threshold.

[0104] When the probability is greater than the preset threshold (0.4~0.6), it is judged as an illegal transaction; otherwise, it is a normal transaction.

[0105] use (Recall rate) (Accuracy) (F1 score) (Area under the curve) (Matthews correlation coefficient) (Geometric mean) is used as an evaluation metric, where:

[0106] , This represents the number of samples that were actually in violation and were predicted to be in violation. This represents the number of samples that actually violated regulations but were predicted to be normal.

[0107] , This represents the number of samples that were actually normal but were predicted to be non-compliant.

[0108] ; for Area under the curve The curve is composed of different thresholds and , (Drawn as the number of samples that are actually normal and predicted to be normal);

[0109] ;

[0110] .

[0111] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An anti-money laundering detection method based on an improved conditional table generative adversarial network, characterized in that, include: Obtain user transaction data, input the user transaction data into the trained anti-money laundering detection model, and obtain the detection results; The training process for an anti-money laundering detection model includes: S1. Obtain the transaction dataset, preprocess each transaction sample in the transaction dataset to obtain the training set; S2. Divide the training set into majority class sample sets. With minority class sample set Based on the minority class sample set For the majority class sample set Filtering is performed to obtain the filtered majority class sample set. ; S3. Construct the CTGAN model According to the minority class sample set For CTGAN model The CTGAN model is obtained through training. CTGAN is a conditional table-based generative adversarial network. S4. Utilize the current CTGAN model Generate and minority class sample sets Given an equal number of candidate synthetic sample sets, determine whether the quality of the candidate synthetic sample sets meets the requirements; where t is the number of iterations for adjusting the CTGAN model. S5. If the quality of the candidate synthetic sample set meets the requirements, then the current CTGAN model will be... This serves as the final trained CTGAN model; otherwise, the current CTGAN model is adjusted. The structure was used to obtain the adjusted CTGAN model. And return to step S3 to adjust the CTGAN model. Conduct training; S6. Generate a synthetic minority class sample set using the finally trained CTGAN model. Merged and filtered majority class sample sets Minority class sample set and minority class synthetic sample sets To obtain a balanced training set ; S7. Based on the balanced training set Train the anti-money laundering detection model to obtain a well-trained anti-money laundering detection model.

2. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 1, characterized in that, Preprocessing for each transaction sample in the transaction dataset includes: The continuous features of each transaction sample in the transaction dataset are normalized, and the discrete features of each transaction sample in the transaction dataset are one-hot encoded to obtain the preliminary processed transaction dataset. A gradient boosting decision tree is trained using the pre-processed transaction dataset. The split gain of each feature in the pre-processed transaction dataset is calculated based on the trained gradient boosting decision tree. The split gain of each feature is then normalized to obtain the importance of each feature in the pre-processed transaction dataset. Features with importance greater than a preset threshold are selected from the initially processed transaction dataset to obtain the preprocessed transaction dataset. The preprocessed transaction dataset is divided into a training set and a test set.

3. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 1, characterized in that, Based on minority sample sets For the majority class sample set Filtering includes: S21. Initialize the empty set R; S22. For each minority class sample Calculate the Euclidean distance between each sample and all other samples in the preprocessed transaction dataset, and select the first... The nearest samples form the minority class samples. nearest neighbor sample set ; S23. For each minority class sample Iterate through its nearest neighbor sample set Each sample in ,like Then the sample Add to collection ; S24, From the majority class sample set Delete all items belonging to the set From the samples, we obtain the filtered majority class sample set. .

4. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 1, characterized in that, The CTGAN model includes: an improved generator and discriminator; the improved generator includes: a pattern-specific normalization layer, a conditional embedding layer, and a generator; and the adjusted CTGAN model. The training includes: S31, The minority class sample set The samples were divided into multiple batches, resulting in multiple subsets of minority class samples. S32. Input the continuous features of the minority class sample subset of each batch into the model. The pattern-specific normalization layer yields multi-pattern parameters of continuous features for the minority class sample subset in each batch; S33. Input the discrete features of the minority class sample set in each batch into the model. The conditional embedding layer obtains the conditional vector of discrete features of the minority class sample set in each batch; S34. Obtain batch random noise, and input the batch random noise, the conditional vector of discrete features of minority class sample subsets in each batch, and the multi-mode parameters of continuous features into the model. The generator produces a composite sample set of minority class sample subsets for each batch; S35. Input the minority class sample subset and its composite sample set from each batch into the model. The discriminator obtains the discrimination results of the minority class sample subset and the synthetic sample set for each batch; S36. Calculate the loss function value based on the discrimination results of the minority class sample subsets and the composite sample set of all batches, and update the model based on the loss function value. The parameters are set such that the trained CTGAN model is obtained when the loss function value is minimized. .

5. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 1, characterized in that, Determining whether the quality of the candidate synthetic sample set meets the requirements includes: calculating the quality of the candidate synthetic sample set and the minority class sample set. The proportion of overlapping samples is calculated to determine the ratio between the candidate synthetic sample set and the minority class sample set. The KL divergence is used to determine the quality of the candidate synthetic sample set. If the proportion of overlapping samples is greater than the preset overlap threshold or the KL divergence is greater than the preset divergence threshold, the quality of the candidate synthetic sample set is determined to be unacceptable. Otherwise, it meets the requirements.

6. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 5, characterized in that, Calculating the proportion of overlapping samples between the candidate synthetic sample set and the minority class sample set includes: For each synthetic sample in the candidate synthetic sample set Calculate its relationship with the minority class sample set. Each sample The Euclidean distance; where u and v are the indices of the samples in the candidate synthetic sample set and the minority class sample set, respectively; Preset distance threshold If the synthetic sample Each sample in the minority class sample set The Euclidean distances are all less than or equal to the distance threshold. Then determine the synthetic sample These are overlapping samples; Count the number of overlapping samples in the candidate synthetic sample set Calculate the number of overlapping samples The ratio of overlapping samples to the total number of samples in the candidate synthetic sample set is used to obtain the proportion of overlapping samples.

7. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 5, characterized in that, Adjust the current CTGAN model The structure includes: If the candidate synthetic sample set and the minority class sample set If the proportion of overlapping samples is greater than a preset threshold, then NCL cleaning is performed on the candidate synthetic sample set, and the NCL-cleaned candidate synthetic sample set and minority class sample set are calculated. The KL divergence is used to adjust the current CTGAN model. The structure; otherwise, based on the candidate synthetic sample set and the minority class sample set. Adjusting the KL divergence of the current CTGAN model The structure is as follows: NCL is the neighborhood cleaning rule.

8. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 7, characterized in that, NCL cleaning of the candidate synthetic sample set includes: identifying the minority class sample set from the candidate synthetic sample set. For overlapping samples, remove all overlapping samples from the candidate synthetic sample set to obtain the NCL-cleaned candidate synthetic sample set.

9. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 7, characterized in that, Adjust the current CTGAN model based on KL divergence. The structure includes: Determine whether the candidate synthetic sample set exhibits feature homogenization; If the KL divergence exceeds the divergence threshold of 10%-30%, then the number of neurons in the generator hidden layer will be increased by 20%. If the KL divergence exceeds the divergence threshold of 30% and the candidate synthetic sample set does not exhibit feature homogenization, then the number of neurons in the generator hidden layer will be increased by 50%. If the KL divergence exceeds the divergence threshold of 30% and the candidate synthetic sample set exhibits feature homogenization, then the activation function of the generator hidden layer should be changed.

10. The anti-money laundering detection method based on an improved conditional table generative adversarial network according to claim 1, characterized in that, The anti-money laundering detection model is the TCN model; where the TCN model is a temporal convolutional network model.

Images