Telecommunication fraud detection model based on mixed attention mechanism data repair and integrated classification fusion and method thereof

Abstract

The application discloses a telecom fraud detection model based on mixed attention mechanism data repair and integrated classification fusion and a method thereof, and comprises the following steps: a generative adversarial interpolation network CSAM-GAIN based on a mixed attention mechanism is used for telecom fraud data repair, a serial connection channel and a spatial attention mechanism are used for capturing data space-time correlation, an adversarial loss, a reconstruction loss and a non-negative constraint loss are combined for optimization, and completed data close to a real distribution are generated; an integrated classification module CTCN-CSIAM-SwinTransformer is used for parallel fusion of CNN and TCN, local space features and time features are extracted respectively, a multi-scale residual structure and a mixed attention are used for enhancing feature discriminability in the CSIAM module, and the SwinTransformer module is used for capturing global space-time dependence and making up for the limitations of local feature extraction. The model performance is optimized by combining space-time feature extraction and an integrated learning method, so that the accuracy and reliability of telecom fraud detection are improved.

Description

Technical Field

[0001] This invention belongs to the field of telecommunications fraud detection technology, specifically involving a telecommunications fraud detection model and method based on data repair and integrated classification fusion using a hybrid attention mechanism. Background Technology

[0002] With the deep integration of communication technology and the internet, telecommunications fraud has become increasingly rampant, posing a serious threat to social security and public property due to its cross-regional, intelligent, and industrialized characteristics. The core of telecommunications fraud lies in exploiting the spatiotemporal characteristics of communication networks to carry out precise deception. Criminals obtain victims' personal information through various means and then use online platforms and communication tools to conduct fraudulent activities. These activities leave traces on operator networks, such as device IDs, geographical locations, and call records; this telecommunications data is stored in databases as metadata. Operators and public security departments can analyze this data to identify fraudsters and shut down involved phone numbers. However, actual telecommunications data often suffers from missing values, high-dimensional nonlinearity, and spatiotemporal dynamic correlations, which limit the accuracy of traditional detection methods and make it difficult to effectively combat the increasingly complex telecommunications fraud.

[0003] Traditional detection methods primarily rely on rule engines and simple statistical analysis, which prove inadequate when faced with the complexity and diversity of telecommunications fraud. Rule engines require manually formulated rules, making it difficult to adapt to constantly evolving fraud tactics; while simple statistical analysis cannot handle high-dimensional nonlinear data and is susceptible to noise and outliers. Therefore, how to efficiently utilize the spatiotemporal characteristics of telecommunications fraud data to construct deep learning models and improve the accuracy and reliability of detection has become a core research issue.

[0004] In recent years, deep learning technology has achieved remarkable results in fields such as image recognition, speech recognition, and natural language processing. Its powerful feature extraction and pattern recognition capabilities have provided new approaches to solving complex problems. Deep learning models can learn the spatiotemporal distribution patterns in data, identify hotspots and correlations, and provide strong support for police in combating and preventing telecommunications fraud. However, the unique characteristics of telecommunications fraud data also bring challenges to the construction of deep learning models. Missing data values ​​can affect the training effect of the model, high-dimensional nonlinear data requires complex feature extraction methods, and spatiotemporal dynamic correlations require the model to be able to handle the fusion of time series and spatial information. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a telecommunications fraud detection model and method based on a hybrid attention mechanism for data repair and ensemble classification. The model solves the problem of missing data values ​​through data repair technology, processes high-dimensional nonlinear data using deep learning algorithms, and optimizes model performance by combining spatiotemporal feature extraction and ensemble learning methods, thereby improving the accuracy and reliability of telecommunications fraud detection.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution: On the one hand, this invention provides a method for detecting telecommunications fraud based on data repair and integrated classification fusion using a hybrid attention mechanism, including: We use the Generative Adversarial Interpolation Network CSAM-GAIN based on a hybrid attention mechanism to repair telecom fraud data. By using a concatenated channel and spatial attention mechanism to capture the spatiotemporal correlation of telecom fraud data, and combining adversarial loss, reconstruction loss and non-negativity constraint loss for optimization, we can generate telecom fraud data that is close to the real distribution. An integrated classification module CTCN-CSIAM-SwinTransformer is constructed. The CTCN module is used to fuse CNN and TCN in parallel to extract local spatial and temporal features of telecom fraud completion data. The multi-scale residual structure and hybrid attention of the CSIAM module are used to enhance the feature discriminativeness of telecom fraud completion data. The SwinTransformer module is used to capture the global spatiotemporal dependence of telecom fraud completion data, thus making up for the limitations of local feature extraction.

[0007] On the other hand, this invention provides a telecommunications fraud detection model based on a hybrid attention mechanism for data repair and integrated classification fusion, comprising: A generative adversarial interpolation network CSAM-GAIN based on a hybrid attention mechanism is used for repairing telecom fraud data. It captures the spatiotemporal correlation of data through a concatenated channel and spatial attention mechanism, and optimizes by combining adversarial loss, reconstruction loss and non-negativity constraint loss to generate complete data that is close to the real distribution. The integrated classification module CTCN-CSIAM-SwinTransformer comprises three modules: CTCN, CSIAM, and SwinTransformer, forming a three-layer processing system of local-feature enhancement-global. The CTCN module fuses CNN and TCN in parallel to extract local spatial and temporal features respectively; the CSIAM module enhances feature discriminativeness through multi-scale residual structures and hybrid attention; and the SwinTransformer module captures global spatiotemporal dependencies, compensating for the limitations of local feature extraction.

[0008] Furthermore, the Generative Adversarial Interpolation Network (CSAM-GAIN) constructs a hybrid attention mechanism by serially integrating channel and spatial attention, accurately capturing and fusing the correlation between channel and spatial distribution at different time points; after adversarial training between the generator and discriminator, it generates multi-dimensional enhanced data based on the prompt matrix, thus completing data reconstruction and enhancement. The generative adversarial interpolation network CSAM-GAIN consists of CSAM, GAIN, loss function and activation function. It enhances the global dependency capture capability through multi-subspace feature interaction and adds a spatiotemporal attention mechanism to the discriminator, which can improve feature understanding, training efficiency and stability, and accurately repair missing data in telecommunications fraud. The CSAM connects the Channel Attention Mechanism (CAM) and the Spatial Attention Mechanism (SAM) in a series, comprehensively focusing on the channel and spatial dimensions of the input features. It optimizes feature representation by processing channels first and then spatially, reducing data volume and acquiring rich features. The CAM calculates the weights of each channel through global pooling and adjusts them to amplify or weaken the influence of channels, generating information-rich feature descriptions. The SAM accurately locates spatial positions and uses a convolution kernel replacement strategy to reduce the parameter overhead of CAM, keeping the mechanism lightweight. It calculates position weights and adjusts their influence through convolution.

[0009] Furthermore, the CAM first performs global max pooling Fcavg and average pooling Fcmax on the input features to obtain two sets of weights consistent with the number of channels, and inputs them into a shared multilayer perceptron (MLP) to generate channel attention. After further compression using 1×1 convolution and restoration of the number of channels, the two sets of matrices are added together and then processed by LeakyReLU to obtain the final weight matrix. The shared multilayer perceptron (MLP) ensures consistent channel attention weights across different scenarios through weight sharing. The output feature vectors are then merged using element-wise addition. The channel attention is as follows: ; Where σ represents the sigmoid function, This represents a convolution operation with a filter size of 7×7.

[0010] Furthermore, the GAIN is an improved GAN framework based on cue vectors, which uses the cue matrix of real observations to perform an unsupervised data generation model, consisting of a generator, a discriminator, and a cue vector matrix; The generator conditionally fills in missing parts based on partial components of the observed data and outputs a complete vector; the discriminator receives the complete vector to distinguish between the observed values ​​and the filled values; the cue vector reveals the missing state of some samples to the discriminator, making it focus on the filling quality of specific components and ensuring that the generator follows the true data distribution.

[0011] Furthermore, a Generative Adversarial Network (GAN) consists of a generator G and a discriminator D; the generator G learns a mapping G(z) and attempts to map a random noise vector z to real data; the discriminator D attempts to find a mapping D(・) to determine the true probability of the input data. The generator learns complex data patterns through competitive training, with the goal of mapping the noise distribution Pz(z) to the real data distribution Pdata(x). The discriminator is responsible for distinguishing between real data and generated data. The generator G and the discriminator D are two competitors and are trained on minimax based on the value function V(D, G). Let Pz(z) be the input prior distribution, and Pdata(x) be the training data distribution; the discriminator By parameterizing the generator network weights, synthetic samples that are difficult to distinguish from real data are generated. Discriminator network weights are parameterized to distinguish between real and generated samples; the optimization goal of Generative Adversarial Network (GAN) is to find the optimal parameters. and The objective function is minimized as follows: ; Where x represents the real data, Pdata(x) represents the distribution of the real data, z is the input of the generator network, usually taken from the assumed prior data distribution Pz(z); D(x) and D(z) represent the outputs of the real data x and the generated data G(z) input to the discriminator network, respectively. The discriminator is guided to focus on the missing region by a cue matrix. During training, a binary cue matrix is ​​generated by randomly sampling the mask matrix Mmb. ; Let the original data matrix be mask matrix ,in express Not missing, hint matrix The generation process is formalized as follows: ;

[0012] in, Let P be an independent and identically distributed Bernoulli sampling matrix with sampling probability P, and ⊙ be the element-wise multiplication of the Hadamard product; each element The generation rules are: ; The discriminator is guided by two mechanisms to determine known values ​​( The original value is retained with probability p, allowing the discriminator to focus on the valid data region and provide deterministic prompts; for missing regions ( Injecting 0.5 intermediate noise helps prevent the generator from getting trapped in local optima and generating random perturbations.

[0013] Furthermore, the combined loss function is trained using a weighted loss function, which consists of discriminator and generator loss functions, enabling the generator G to generate more realistic samples and enabling the discriminator D to better distinguish between real samples and generated samples. The generator loss function includes: Adversarial loss measures the generator's ability to deceive the discriminator, and is based on the cross-entropy loss of the discriminator's output, used to deceive the discriminator. Reconstruction loss, or mean squared error loss, is used to measure the difference between the generated data and the known parts of the original data. ; in, It is a mask matrix. G is the output of the discriminator, and G is the output of the generator, used to avoid numerical instability in logarithmic operations. It is a hyperparameter with a value between 0.1 and 1.0, used to balance the adversarial loss and reconstruction accuracy, and to balance the two parts of the loss. This parameter is set to prevent numerical instability during logarithmic operations. The hyperparameter β controls the strength of this penalty term, typically between 0.1 and 0.5. In addition, nonnegativity constraint loss, negative value penalty, prevents the generation of negative values ​​and ensures physical interpretability, is expressed as: ; in, The loss is a region-based loss function that uses certain weights to... Loss and Loss combination, for the weighted combination loss function, can be used to better segment regions and ensure a more stable training process; where α is an arbitrary constant, α∈(0,1); The objective function is optimized as follows: ; The discriminator loss function is based on cross-entropy loss and measures the discriminator's ability to distinguish between real and generated data. The optimization objective function is as follows: ; in, It is a mask matrix. This is the output of the discriminator, used to avoid numerical instability in logarithmic operations. This parameter is set to prevent numerical instability during logarithmic operations; E represents expectation, which is the expectation of the loss expression within the parentheses over the entire data distribution.

[0014] Furthermore, in the integrated classification module CTCN-CSIAM-SwinTransformer, local spatiotemporal dynamic features are extracted by CTCN, then the features are optimized by CSIAM, and finally the global dependencies are captured by SwingTransformer to comprehensively capture the spatiotemporal correlation of the data. The CTCN module serves as a low-level feature extractor, integrating CNN and TCN. CNN extracts local spatial features, while TCN extracts local temporal features through causal convolution and dilated convolution. The CSIAM module is embedded between the CTCN module and the SwinTransformer module. It integrates convolutional features from different levels through a multi-scale depth-separable residual module while preserving details and global semantics. It also optimizes channel weights with channel attention and focuses on key regions and suppresses redundancy with spatial attention. The SwinTransformer module, as a high-level feature fusion unit, models global dependencies through a self-attention mechanism, captures long-distance temporal correlations, and makes up for the shortcomings of CTCN in long-range dependencies and global feature extraction.

[0015] Furthermore, TCN consists of causal convolution, dilated convolution, and residual connections. Causal convolution ensures that at a given time t, the model depends only on historical data, forming a strict time constraint and maintaining model stability; dilated convolution expands the receptive field through an exponential expansion factor, resolving long-distance time dependencies without increasing the number of layers; residual connections alleviate gradient vanishing / exploding, maintaining performance stability. Assume the convolution filter is represented as F =( f 1, f 2,⋯, f K The input sequence is X=(x1,x2,⋯ ,x T If ), then the causal convolution at position t can be expressed as: ; Time series prediction models require that the output at the current time step completely cover all its previous inputs. When the kernel size is K... C The input time length is L time When, the required number of convolutional layers L C satisfy: ; TCN expands the receptive field by introducing a dilation factor. The receptive field w of a specific layer is based on the number of layers below the current layer i, the dilation parameter b, the input length Ltime, the number of topmost layers n required to cover the complete input history, and the number of zeros p required for each layer. Spacing is inserted between convolutional kernel elements, with a spacing of b−1. The actual receptive field calculation formula after dilation (b=1 in standard CNNs) is as follows: ; ; To ensure that the input and output sequences have the same length, zero padding must be applied to the end of the sequence. Dilated convolution is used to achieve an exponentially expanded receptive field to capture local feature information at different scales. The operation definition of dilated causal convolution is as follows: ; Where f(t) represents the dilated convolution operation at time step t, f:{0,…,k−1}→ℝ represents the filter, and d is the dilation factor. d f represents a convolution operation with stride d; before the training samples are used as input to the CTCN module, missing values ​​are padded with the annual average to complete the initialization; A four-layer TCN is used, with a kernel size of 3 and dilation factors d of 1, 2, 4 and 8 respectively. Holes are added to the causal convolution to skip a certain number of time steps in the input sequence. Dilated causal convolution achieves a larger receptive field through interval sampling, which makes up for the limitations of causal convolution. By processing sequence data through causal dilated convolution, the model can effectively capture short-term dependencies.

[0016] Furthermore, in TCN, the residual module is passed across layers, and a dropout layer is added to the residual module for regularization to prevent overfitting; The input to the residual module is summed with the output after two convolution operations using an identity mapping to ensure dimensionality consistency. Residual connections train the network by directly adding the input to the output of a convolutional layer. The residual connection expression is as follows: ; Where x represents the input data, and F(x) represents the result of the input nonlinear transformation.

[0017] Furthermore, the residual module Inception is combined with the hybrid attention module CSAM to construct the residual module CSIAM based on the hybrid attention mechanism; Among them, the Hybrid Attention Module (CSAM) calculates attention weights through channel and spatial dimensions, CAM generates channel weights through global pooling and MLP to strengthen key channels, and SAM generates spatial weights through convolution and pooling to focus on important positions, thus highlighting effective information and suppressing redundancy. The residual module Inception achieves multi-scale feature fusion through a pyramid structure, integrating convolutional features from different levels, preserving details and global semantics, and solving the problem of insufficient semantics at a single scale. In addition, the model extracts local features from multiple scales and dynamically enhances key regions with attention. The combination of these two methods not only accurately captures multi-scale features but also enhances the discriminative power of low-contrast features, thereby improving the model's ability to process complex spatiotemporal data. The flow of input feature data is divided into two branches: One component first passes through the channel attention module, multiplies the resulting weight matrix with the original input data, and then passes through the spatial attention module and is multiplied by the corresponding weights. Another shortcut similar to ResNet is direct mapping, where the two data streams are ultimately added together and output through an activation function; the attention features output by each module are combined with the original features as follows: ; in, It is the output of channel attention. It is the output after applying spatial attention to the channel attention result.

[0018] Furthermore, the Swin Transformer module includes layer normalization (LN), multi-head self-attention, multilayer perceptron (MLP), and residual connections. The MLP contains the GELU activation function, and LN and residual connections are used for stabilizing training and preserving features, respectively. The computational complexity is reduced through window self-attention, the shifted window strategy enhances cross-window information interaction, and hierarchical feature representation efficiently captures long-distance dependencies and context, optimizing computational resources while maintaining architectural scalability. MLP introduces nonlinearity to learn complex patterns, and the classification head outputs predictions, achieving efficient fusion of local and global features, represented as: ; Among them, z l For the input of the l-th layer, z l1 This represents the output of the W-MSA layer before the MLP layer in the (l+1)-th module; (l+1)-th is the input of the (l+1)-th module (i.e., the output of the l-th module), z l+1 ,z l1+1 This represents the intermediate output after the SW-MSA layer and before the MLP layer in the (l+1)-th module, z. l-1 This represents the output characteristics of the previous layer; before each MHSA module and MLP, an LN layer is used to normalize the input; and after each module, residual connections are used to facilitate information flow. W-MSA is built upon multi-head self-attention constrained to a non-overlapping window. Compared to global self-attention, it helps reduce computational complexity, as expressed below: ; Where, 𝑄, 𝐾, and 𝑉 are the query, key, and value matrices, respectively, and 𝑊 𝐾 and 𝑊 𝑉 A learnable weight matrix for keys and values; SW-MSA achieves cross-window information exchange through window sliding and overlapping, thus preserving the ability to model global context information, as shown below: ; Among them, W 𝐾,shifted and W 𝑉,shifted This is the learnable weight matrix for the keys and values ​​after sliding.

[0019] At each stage, the number of patches is reduced by merging, and a hierarchical feature representation is generated after each stage. During the operation, adjacent patches are first concatenated, and then mapped to a lower dimension through linear projection, as shown below: ; Among them, W m and b m These are the learnable weight matrix and the bias term, respectively. The size of the patch determines the data segmentation method, which affects both the details captured and the computational efficiency.

[0020] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The telecom fraud detection model and method based on data repair and integrated classification fusion based on the hybrid attention mechanism provided by the present invention focuses on building a more general and efficient anti-fraud model. Through data augmentation and model classification optimization, it improves data quality and reliability and achieves accurate detection.

[0021] The technical solution of this invention is based on the generative adversarial network model CSAM-GAIN with a hybrid attention mechanism, which overcomes the limitation of capturing long-term data dependencies, generates data that is closer to reality, and improves the integrity of the original data; and constructs an integrated deep learning model CTCN-CSIAM-SwinTransformer.

[0022] Among them, the CTCN module extracts spatiotemporal features; the multi-scale hybrid attention enhancement module CSIAM improves the feature representation capability of CTCN and efficiently extracts spatiotemporal multidimensional features; the SwinTransformer module captures global spatial attributes and long-term dependency attributes.

[0023] The various modules in the model work together to process complex data relationships. By combining and optimizing classification algorithms and improving the model structure to adapt to the characteristics of telecom fraud data, the model improves the efficiency of spatiotemporal data processing and optimizes the prediction accuracy and robustness. The main technical innovations are as follows: 1) A generative adversarial network interpolation model based on a hybrid attention mechanism (CSAM-GAIN) is proposed to improve the integrity and reliability of the original data based on a spatiotemporal data repair algorithm.

[0024] 2) Construct a representation module based on spatiotemporal feature extraction (CTCN) and a feature enhancement module based on residual hybrid attention mechanism (CSIAM) to strengthen the differential expression of different types of features, fuse and enhance spatiotemporal representations, and generate more discriminative features.

[0025] 3) Construct an integrated deep learning model CTCN-CSIAM-SwinTransformer, which captures short-term and long-term dependencies in data through model structure optimization, extracts local and global features, realizes multi-level feature modeling of spatiotemporal data, and improves detection accuracy and generalization ability. Attached Figure Description

[0026] Figure 1 The following is an architecture diagram of a telecom fraud detection model CGTCTR based on a hybrid attention mechanism for data repair and integrated classification fusion, provided for embodiments of the present invention.

[0027] Figure 2 This is a data repair model CSAM-GAIN architecture diagram provided in an embodiment of the present invention.

[0028] Figure 3 This is a CSAM model architecture diagram provided for an embodiment of the present invention.

[0029] Figure 4 This is a CAM model structure diagram provided in an embodiment of the present invention.

[0030] Figure 5 This is a structural diagram of a spatial attention mechanism (SAM) model provided in an embodiment of the present invention.

[0031] Figure 6 This is a diagram of a generative adversarial interpolation network structure provided in an embodiment of the present invention.

[0032] Figure 7 This is a diagram of a generative adversarial network structure provided in an embodiment of the present invention.

[0033] Figure 8 The following is a classification model architecture diagram provided for an embodiment of the present invention: CTCN-CSIAM-SwinTransformer.

[0034] Figure 9 This is a TCN causal expansion residual structure diagram provided for an embodiment of the present invention.

[0035] Figure 10 This is a CSIAM structure diagram provided for an embodiment of the present invention.

[0036] Figure 11 This is a structural diagram of a Swing Transformer provided in an embodiment of the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0038] Combination Figure 1 This invention provides a telecom fraud detection model based on a hybrid attention mechanism for data imputation and fusion of classification, including a generative adversarial imputation network CSAM-GAIN and an fusion classification module CTCN-CSIAM-SwinTransformer.

[0039] Among them, the Generative Adversarial Imputation Network CSAM-GAIN based on the hybrid attention mechanism is used for repairing telecom fraud data. It captures the spatiotemporal correlation of data through the concatenation channel and spatial attention mechanism, and combines adversarial loss, reconstruction loss and non-negativity constraint loss for optimization to generate complete data that is close to the real distribution.

[0040] To address the high-dimensional complexity, nonlinearity, and close correlation of missing values ​​in telecommunications fraud data, and considering the difficulty of existing methods in accurately capturing data features and relationships, a Generative Adversarial Imputation Network (CSAM-GAIN) model with a hybrid attention mechanism is proposed. This model sequentially constructs a hybrid attention mechanism using channel and spatial attention to accurately capture and fuse the correlations between channel and spatial distributions at different time points. Through adversarial training between the generator and discriminator, multi-dimensional enhanced data is generated based on a cue matrix, completing data reconstruction and enhancement, and providing high-quality data support for subsequent processing.

[0041] The integrated classification module CTCN-CSIAM-SwinTransformer comprises three modules: CTCN, CSIAM, and SwinTransformer, forming a three-layer processing system of local-feature enhancement-global. The CTCN module fuses CNN and TCN in parallel to extract local spatial and temporal features respectively; the CSIAM module enhances feature discriminativeness through multi-scale residual structures and hybrid attention; and the SwinTransformer module captures global spatiotemporal dependencies, compensating for the limitations of local feature extraction.

[0042] Meanwhile, this invention also provides a method for detecting telecommunications fraud based on data repair and integrated classification fusion using a hybrid attention mechanism, including: We use the Generative Adversarial Interpolation Network CSAM-GAIN based on a hybrid attention mechanism to repair telecom fraud data. By using a concatenated channel and spatial attention mechanism to capture the spatiotemporal correlation of telecom fraud data, and combining adversarial loss, reconstruction loss and non-negativity constraint loss for optimization, we can generate telecom fraud data that is close to the real distribution. An integrated classification module CTCN-CSIAM-SwinTransformer is constructed. The CTCN module is used to fuse CNN and TCN in parallel to extract local spatial and temporal features of telecom fraud completion data. The multi-scale residual structure and hybrid attention of the CSIAM module are used to enhance the feature discriminativeness of telecom fraud completion data. The SwinTransformer module is used to capture the global spatiotemporal dependence of telecom fraud completion data, thus making up for the limitations of local feature extraction.

[0043] To address the sparsity and spatiotemporal characteristics of fraud data and overcome the performance limitations of traditional models that neglect spatiotemporal correlations, a CTCN-CSIAM-SwinTransformer model is proposed. In this model, CTCN is fused in parallel with TCN and CNN to capture spatial relationships and extract global spatial and local temporal features, making it suitable for fraud data containing spatiotemporal information. A CSIAM module with concatenated channels and spatial attention is introduced, adaptively learning weights to enhance the feature representation capability of CTCN and efficiently extract and fuse multi-scale spatiotemporal features. Finally, the encoder of the SwinTransformer model extracts temporal features and long-term dependencies, achieving high-precision prediction through fully connected layers.

[0044] GAIN is an unsupervised data generation model based on GANs, utilizing a cue matrix of real observations for generation. Targeting the spatiotemporal characteristics of telecommunications data, it introduces a serial spatial and channel attention mechanism on top of GAIN to extract deep global features, eliminate irrelevant information, accurately capture spatiotemporal features, and improve accuracy. Simultaneously, it designs a combination of adversarial, generative, and centrality loss functions to optimize incomplete data completion performance, and employs an activation function combining LeakyReLU and Sigmoid to retain more features and compensate for insufficient extraction. The model consists of CSAM, GAIN, loss functions, and activation functions (see reference). Figure 2 By enhancing global dependency capture capabilities through multi-subspace feature interaction and incorporating a spatiotemporal attention mechanism into the discriminator, feature understanding, training efficiency, and stability can be improved, and missing data related to telecommunications fraud can be accurately repaired.

[0045] Channel attention (CAM) and spatial attention (SAM) process the channel and spatial relationships of features. Since their contributions to classification differ, a hybrid optimized multi-scale attention mechanism, CSAM (refer to Figure 3), is designed. CSAM connects CAM and SAM in a concatenated manner, comprehensively focusing on the channel and spatial dimensions of the input features. By processing channels first and then spaces, it optimizes feature representation, reducing data volume and acquiring rich features. CAM calculates the weights of each channel through global pooling and adjusts them to amplify or weaken the influence of channels, generating information-rich feature descriptions. SAM precisely locates spatial positions and uses a convolution kernel replacement strategy to reduce the parameter overhead of CAM, keeping the mechanism lightweight. It calculates position weights and adjusts their influence through convolution.

[0046] CAM is an augmented feature representation in convolutional neural networks that processes feature data by weighting it according to the channel dimension (see reference). Figure 4 It first performs global max pooling (F) on the input features. c avg ) and average pooling (F c max Two sets of weights, consistent with the number of channels, are obtained and input into a shared multilayer perceptron (MLP) to generate channel attention. After 1×1 convolution compression to restore the number of channels, the two sets of matrices are added together and then processed by LeakyReLU to obtain the final weight matrix. The MLP ensures consistent channel attention weights across different scenarios through weight sharing, and the output feature vectors are merged using element-wise addition. The channel attention is as follows: ; Where σ represents the LeakyReLU function, the weights W0 and W1 of the MLP are shared in both input scenarios, and the LeakyReLU activation function is applied after W0.

[0047] CAM dynamically adjusts the features of each channel by learning the importance weights between channels, thus enabling the model to focus more on key features and reduce interference from unimportant features. This improves the discriminative and expressive power of features. This mechanism not only enhances the model's performance but also strengthens its robustness, resistance to interference, and interpretability.

[0048] Spatial Attention (SAM) leverages spatial relationships between features to generate spatial labels, assisting convolutional neural networks in deeply mining the spatial features of data. By learning the weights at each location, SAM enables the network to focus on spatial locations crucial to the task, thereby improving network performance and representation capabilities. As shown in Figure 5, SAM superimposes a differentiable weight matrix onto the input data plane, which can receive backpropagation gradients from the model. Using the output of Channel Attention (CAM) as the input feature, SAM first performs max pooling and average pooling on it respectively, obtaining two preliminary weight matrices with the same plane size as the input and a channel count of 1. After stacking these two matrices along the channel direction, a 1×1 convolution is performed to compress the channel dimension to 1, forming a weight matrix with the same plane size as the input feature and a channel count of 1. Max pooling and average pooling are then performed on the feature layer map to reduce the channel dimension, and features are extracted through a 7×7 receptive field convolution. The resulting feature elements are summed and fused, and then a sigmoid activation function is used to generate spatial attention features. The LeakyReLU activation function is used when calculating channel attention, which can constrain the range of weights and perform nonlinear modeling of channel features, effectively enhancing the performance of channel attention.

[0049] Finally, the spatial attention features are multiplied by the CAM output to obtain the final spatiotemporal features, calculated as follows: ; Where σ represents the sigmoid function, This represents a convolution operation with a filter size of 7×7.

[0050] Traditional GANs are prone to pattern collapse when imputing missing data, while cue vectors can reveal some missing location information, guiding the discriminator to specifically evaluate the imputation quality, preventing the generator from ignoring specific missing patterns, and enhancing the model's adaptability to complex missing mechanisms. GAIN, an improved GAN framework based on cue vectors, is a model that uses the cue matrix of real observations for unsupervised data generation (e.g.,...). Figure 6 As shown, GAIN consists of a generator, a discriminator, and a hint vector matrix. The generator conditionally fills in missing parts based on partial components of the observed data and outputs a complete vector. The discriminator receives the complete vector and distinguishes between the observed values ​​and the filled values. The hint vector reveals the missing status of some samples to the discriminator, allowing it to focus on the quality of filling in specific components and ensuring that the generator follows the true data distribution. GAIN transforms the filling problem into an adversarial optimization of local realism verification through an adversarial framework and hint mechanism, providing a new paradigm for handling missing data in high-dimensional and complex data.

[0051] Generative Adversarial Networks (GANs) consist of a generator (G) and a discriminator (D). The generator G learns a mapping G(z), attempting to map a random noise vector z to real data; the discriminator D tries to find a mapping D(・) to determine the true probability of the input data. The two compete with each other through alternating training. It can learn complex data patterns, with the goal of enabling the generator to map the noise distribution Pz(z) to the real data distribution Pdata(x), creating data that is as indistinguishable from real data as possible; the discriminator is responsible for distinguishing between real and generated data. The generator G and the discriminator D, as two competitors, undergo minimax training based on the value function V(D, G).

[0052] Let Pz(z) be the input prior distribution, and Pdata(x) be the training data distribution. Discriminator : By parameterizing the weights of the generator network, synthetic samples that are difficult to distinguish from real data are generated. The discriminator network weights are parameterized to distinguish between real and generated samples. The optimization goal of GANs is to find the optimal parameters. and The objective function is minimized as follows: ; Where x represents the real data, Pdata(x) represents the distribution of the real data, and z is the input to the generator network (usually taken from the assumed prior data distribution Pz(z)). D(x) and D(z) represent the outputs of the discriminator network, respectively, when the real data x and the generated data G(z) are input to the discriminator network.

[0053] The discriminator is guided to focus on missing regions by a cue matrix. During training, the mask matrix M is adjusted accordingly. mb Perform random sampling to generate a binary cue matrix. .

[0054] Let the original data matrix be mask matrix ,in express Not missing, hint matrix The generation process is formalized as follows: ;

[0055] in, Let be an independent and identically distributed Bernoulli sampling matrix with sampling probability P, and ⊙ be the Hadamard product (element-by-element multiplication). Each element The generation rules are: ; The discriminator is guided by two mechanisms. For known values ​​( The original value is retained with probability p, allowing the discriminator to focus on the valid data region and provide deterministic prompts. For missing regions ( Injecting 0.5 intermediate noise helps prevent the generator from getting trapped in local optima and generating random perturbations.

[0056] The model is primarily trained using a weighted loss function, consisting of discriminator and generator loss functions. This aims to enable the generator G to produce more realistic samples and to allow the discriminator D to better distinguish between real and generated samples.

[0057] The generator loss function consists of three parts: one part is the adversarial loss, which measures the generator's ability to deceive the discriminator. It is based on the cross-entropy loss of the discriminator's output and is used to deceive the discriminator. The other part is the reconstruction loss, which is the mean squared error loss and is used to measure the difference between the generated data and the known parts of the original data. ; in, It is a mask matrix. G is the output of the discriminator, and G is the output of the generator. This is used to avoid numerical instability in logarithmic operations. It is a hyperparameter with a value between 0.1 and 1.0, used to balance the adversarial loss and reconstruction accuracy, and to balance the two parts of the loss. This parameter is set to prevent numerical instability during logarithmic operations. The hyperparameter β controls the strength of this penalty term, typically between 0.1 and 0.5.

[0058] The final part is the nonnegativity constraint loss, which is a negative value penalty to prevent the generation of negative values ​​and ensure physical interpretability.

[0059] ; Class imbalances frequently occur in telecom fraud data, which can severely impact the effectiveness of the analysis. The loss is a region-based loss function, so its training effect may be unstable. Therefore, we use certain weights to... Loss and Loss combination, specifically a weighted combined loss function, can be used to better segment regions while ensuring a more stable training process. Here, α is an arbitrary constant, α ∈ (0,1). The optimization objective function is as follows: ; Discriminator loss function: Based on cross-entropy loss, it measures the discriminator's ability to distinguish between real and generated data. The optimization objective function is as follows: ; in, It is a mask matrix. This is the output of the discriminator, used to avoid numerical instability in logarithmic operations. This parameter is set to prevent numerical instability during logarithmic operations. E: represents the expectation, which is the expectation of the loss expression within the parentheses over the entire data distribution. In deep learning implementations, it is usually approximated by the average loss of a mini-batch of samples.

[0060] refer to Figure 8 To address the spatiotemporal correlation of telecom fraud data, the CTCN-CSIAM-SwinTransformer model is composed of three modules: CTCN, CSIAM, and SwinTransformer. Combining the complementary advantages of these three modules in time series data processing, a three-layer processing system of local-feature enhancement-global is formed, which improves the efficiency of spatiotemporal sequence processing and the ability to analyze complex time series data, and optimizes prediction accuracy and model robustness.

[0061] CTCN, as the underlying feature extractor, integrates CNN and TCN. CNN extracts local spatial features, while TCN extracts local temporal features through causal convolution and dilated convolution, ensuring temporal causality and efficiently capturing short-term dependencies. CSIAM, embedded between CTCN and SwinTransformer, integrates convolutional features from different levels through multi-scale, deeply separable residual modules, preserving details and global semantics. Channel attention optimizes channel weights, and spatial attention focuses on key regions to suppress redundancy. The residual module extracts information at multiple scales, improving model stability and feature discriminative power, and enhancing CTCN's feature representation capabilities. SwinTransformer, as the high-level feature fusion unit, models global dependencies through a self-attention mechanism, capturing long-distance temporal correlations and compensating for CTCN's shortcomings in long-range dependencies and global feature extraction. In terms of model flow, CTCN first extracts local spatiotemporal dynamic features, then CSIAM optimizes the features, and finally SwinTransformer captures global dependencies, comprehensively capturing spatiotemporal correlations of data and improving model performance.

[0062] Telecommunications fraud data contains multi-dimensional features including time and space. CNNs excel at capturing local spatial features and can learn unknown local spatial dependencies; while possessing strong feature extraction capabilities, they struggle to remember temporal information. TCNs, on the other hand, excel at capturing temporal features but lack spatial domain capture capabilities, and neither effectively addresses the temporal-spatial dependency problem. This embodiment connects TCN and CNN in parallel. CNN captures local spatial features, while TCN enhances local temporal feature extraction through causal dilated convolutions. The combination forms the CNNTCN model, specifically designed for processing data containing spatiotemporal features, improving local modeling and prediction performance.

[0063] The TCN module will be introduced in detail below.

[0064] TCN consists of causal convolution, dilated convolution, and residual connections. Causal convolution ensures that at a given time t, the model depends only on historical data, forming a strict time constraint and maintaining model stability. Dilated convolution expands the receptive field through an exponential expansion factor, resolving long-distance time dependencies without increasing the number of layers. Residual connections alleviate gradient vanishing / exploding, maintaining performance stability.

[0065] TCN employs a one-dimensional fully convolutional structure, with zero padding on the left side of the input to maintain the output length, and causal dilated convolution (see reference). Figure 9 It can efficiently capture long-term temporal dependencies. Although causal convolution solves the information leakage problem, its receptive field is limited, and it is still insufficient for capturing long-term dependencies. TCN operates unidirectionally based on the principle of causality and has no bidirectional capability. It uses temporal convolution as its core to ensure that the network depth matches the length of the input sequence and achieves full coverage of historical information.

[0066] Assume the convolution filter is represented as F =( f 1, f 2,⋯, f K The input sequence is X=(x1,x2,⋯ ,x T ), then the position t The causal convolution at a given point can be represented as: ; Time series prediction models require that the output at the current time step completely cover all its previous inputs, which means that when the kernel size is K... C The input time length is L time When, the required number of convolutional layers L C satisfy: ; Causal convolutions only access historical information to prevent information leakage, but their receptive field is limited, requiring more layers or larger filters to expand the perceptual range. Dilated convolutions expand the receptive field by introducing a gap, enabling the capture of multi-scale temporal information and long-distance relationships with fewer layers, compensating for the limitations of causal convolutions and improving model performance and generalization ability. TCNs expand the receptive field by introducing a dilation coefficient. The receptive field w of a specific layer is based on the number of layers below the current layer i, the dilation parameter b, the input length Ltime, the number of top layers n required to cover the complete input history, and the number of zeros padded per layer. A gap (b−1) is inserted between the convolution kernel elements. The formula for calculating the actual receptive field after dilation in standard CNNs (b=1, no gap) is as follows: ; ; Causal convolution preserves only feedforward neuron connections in a time series, ensuring that the output at time step t is only related to data from previous layers at time step t and earlier. To ensure the input and output sequences are of the same length, zero-padding is applied to the end of the sequence. Furthermore, dilated convolution is used to achieve an exponentially expanded receptive field to capture local feature information at different scales. The operation of dilated causal convolution is defined as follows: ; Where f(t) represents the dilated convolution operation at time step t, f:{0,…,k−1}→ℝ represents the filter, and d is the dilation factor. d f represents a convolution operation with stride d. Before using training samples as input to the CTCN module, missing values ​​are padded with annual averages to complete the initialization.

[0067] In this embodiment, a four-layer TCN is used, with a kernel size of 3 and dilation factors d of 1, 2, 4, and 8, respectively. Therefore, adding holes to the causal convolution allows skipping a certain number of time steps in the input sequence. Dilated causal convolution can achieve a larger receptive field through interval sampling, compensating for the limitations of causal convolution. By processing sequence data through causal dilated convolution, the model can effectively capture short-term dependencies.

[0068] Residual modules in TCN are passed across layers, thereby improving generalization ability, such as Figure 9 As shown. A dropout layer is added to the residual module for regularization to prevent overfitting. The input to the residual module is summed to the outputs after two convolutional operations via an identity mapping to ensure dimensionality consistency. Residual connections, by directly adding the input to the output of the convolutional layer, allow for better network training. The residual connection expression is as follows: ; Where x represents the input data, and F(x) represents the result of the input nonlinear transformation.

[0069] To address the issue of redundant information in the input due to long-distance spatial correlation, this embodiment introduces a convolutional block attention module to focus on key regions and time intervals to capture the complex spatiotemporal dependencies of telecommunications data. To enhance the local spatiotemporal feature extraction capability of CTCN and improve model performance and generalization ability, the residual module Inception is combined with the hybrid attention module CSAM to construct a residual module CSIAM based on a hybrid attention mechanism. Figure 9 CSIAM, comprising the hybrid attention module CSAM and the residual linking module Inception, excels at fusing local representations of spatiotemporal features of data, generating strong features and suppressing secondary features that interfere with classification, thereby increasing the attention given to the input data.

[0070] The Hybrid Attention (CSAM) module calculates attention weights based on channel and spatial dimensions. CAM generates channel weights through global pooling and MLP, strengthening key channels. SAM generates spatial weights through convolution and pooling, focusing on important locations to highlight effective information and suppress redundancy. The Inception module achieves multi-scale feature fusion through a pyramid structure, integrating convolutional features from different levels, preserving details and global semantics, and addressing the problem of insufficient semantics at a single scale. Inception extracts local features from multiple scales, and attention dynamically strengthens key regions. The combination of these two approaches accurately captures multi-scale features while enhancing the discriminative power of low-contrast features, improving the model's ability to handle complex spatiotemporal data.

[0071] The input feature data flows in two branches: one branch first passes through the channel attention module, multiplying the resulting weight matrix by the original input data, and then passes through the spatial attention module and is multiplied by the corresponding weights; the other branch, similar to a shortcut in a ResNet, is a direct mapping. Finally, the two data streams are added together and output through an activation function. The attention features output by each module are combined with the original features as follows: ; in, Mc ( F ) is the output of channel attention. Ms ( Mc ( F The output is the result of applying spatial attention to the channel attention result.

[0072] To address the computational complexity and inefficiency of traditional Transformers when processing global features, SwinTransformer proposes a window-based Transformer architecture. It optimizes the self-attention mechanism using sliding window technology, thereby improving the efficiency of high-resolution data processing. Its block structure (see reference) Figure 11 This architecture incorporates Layer Normalization (LN), Multi-Head Self-Attention, Multi-Layer Perceptron (MLP), and Residual Connections. The MLP includes the GELU activation function, while LN and Residual Connections are used for stable training and feature preservation, respectively. Window self-attention reduces computational complexity, a shifted window strategy enhances cross-window information interaction, and hierarchical feature representation efficiently captures long-range dependencies and context, optimizing computational resources while maintaining architectural scalability. The MLP introduces non-linearity to learn complex patterns, and the classification head outputs predictions, achieving efficient fusion of local and global features. (See below.) ; Among them, z l For the input of the l-th layer, z l1This represents the output of the W-MSA layer before the MLP layer in the (l+1)-th module; (l+1)-th is the input of the (l+1)-th module (i.e., the output of the l-th module), z l+1 ,z l1+1 This represents the intermediate output after the SW-MSA layer and before the MLP layer in the (l+1)-th module, z. l-1 This represents the output characteristics of the previous layer. Before each MHSA module and MLP, an LN layer is used to normalize the input; after each module, residual connections are used to facilitate information flow.

[0073] W-MSA is built upon multi-head self-attention, which is restricted to non-overlapping windows. Compared to global self-attention, it helps reduce computational complexity.

[0074] ; Where, 𝑄, 𝐾, and 𝑉 are the query, key, and value matrices, respectively, and 𝑊 𝐾 and 𝑊 𝑉 Let be the learnable weight matrix for keys and values.

[0075] SW-MSA enables cross-window information exchange through window sliding and overlapping, thus preserving the ability to model global context information.

[0076] ; Among them, W 𝐾,shifted and W 𝑉,shifted This is the learnable weight matrix for the keys and values ​​after sliding.

[0077] Patch merging reduces the number of patches at each stage, improving computational efficiency, and generates hierarchical feature representations after each stage. During the operation, adjacent patches are first concatenated, and then mapped to a lower dimension through linear projection (using learnable weight matrices and biases). Among them, W m and b m These are the learnable weight matrix and the bias term, respectively.

[0078] Patch size determines the data segmentation method, affecting both the detail captured and computational efficiency. Embedding dimension controls the richness of feature representation, while the number of patches and attention mechanisms influence the model's ability to capture local and long-range dependencies in the data.

Claims

1. A method for detecting telecommunications fraud based on data repair and integrated classification fusion using a hybrid attention mechanism, characterized in that, include: We use the generative adversarial interpolation network CSAM-GAIN based on a hybrid attention mechanism to repair telecom fraud data. By using a concatenated channel and spatial attention mechanism to capture the spatiotemporal correlation of telecom fraud data, and combining adversarial loss, reconstruction loss and non-negativity constraint loss for optimization, we can generate telecom fraud data that is close to the real distribution. An integrated classification module CTCN-CSIAM-SwinTransformer was constructed, and the CTCN module was used to fuse CNN and TCN in parallel to extract the local spatial and temporal features of the telecom fraud completion data. The discriminative power of telecom fraud completion data features is enhanced by using the multi-scale residual structure and hybrid attention of the CSIAM module, and the global spatiotemporal dependence of telecom fraud completion data is captured by the SwinTransformer module to make up for the limitations of local feature extraction.

2. The method for detecting telecommunications fraud based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 1, characterized in that, The generative adversarial interpolation network CSAM-GAIN uses serial channel and spatial attention to construct a hybrid attention mechanism, which accurately captures and fuses the correlation between channel and spatial distribution at different time points. After adversarial training between the generator and the discriminator, multi-dimensional augmented data is generated based on the prompt matrix to complete data reconstruction and augmentation. The generative adversarial interpolation network CSAM-GAIN consists of CSAM, GAIN, loss function and activation function. It enhances the global dependency capture capability through multi-subspace feature interaction and adds a spatiotemporal attention mechanism to the discriminator, which can improve feature understanding, training efficiency and stability, and accurately repair missing data in telecommunications fraud. The CSAM connects the Channel Attention Mechanism (CAM) and the Spatial Attention Mechanism (SAM) in a series, comprehensively focusing on the channel and spatial dimensions of the input features. It optimizes feature representation by processing channels first and then spatially, reducing data volume and acquiring rich features. The CAM calculates the weights of each channel through global pooling and adjusts them to amplify or weaken the influence of each channel, generating information-rich feature descriptions. The SAM accurately locates the spatial position and uses a convolution kernel replacement strategy to reduce the parameter overhead of the CAM, keeping the mechanism lightweight. It calculates position weights and adjusts the influence through convolution.

3. The telecommunications fraud detection method based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 2, characterized in that, The CAM first performs global max pooling Fcavg and average pooling Fcmax on the input features to obtain two sets of weights consistent with the number of channels. The input is shared by a multilayer perceptron (MLP) to generate channel attention. After further compression using 1×1 convolution and restoration of the number of channels, the two sets of matrices are added together and then processed by LeakyReLU to obtain the final weight matrix. The shared multilayer perceptron (MLP) ensures consistent channel attention weights across different scenarios through weight sharing. The output feature vectors are then merged using element-wise addition. The channel attention is as follows: ； Where σ represents the LeakyReLU function, the weights W0 and W1 of the MLP are shared in the two input scenarios, and the LeakyReLU activation function is applied after W0; The SAM (Spatial Attention Mechanism) receives model gradients through backpropagation by stacking differentiable weight matrices on the input data plane. It takes the output of the channel attention mechanism (CAM) as the input feature and performs max pooling and average pooling on the input feature plane to obtain two initial weight matrices with the same plane size as the input and a channel number of 1. After stacking in the channel direction, the channel dimension is compressed to 1 by 1×1 convolution to form a weight matrix with the same plane size as the input feature and a channel number of 1. Max pooling and average pooling are performed on the feature layer map to reduce the channel dimension. Then, features are extracted by convolution with a 7×7 receptive field. The obtained feature elements are summed and fused, and spatial attention features are generated by the Sigmoid activation function. When calculating channel attention, the LeakyReLU activation function is used, which can constrain the weight range and perform non-linear modeling of channel features, effectively enhancing the channel attention performance. Finally, the outputs of the spatial attention mechanism (SAM) and the channel attention mechanism (CAM) are multiplied to obtain the final spatiotemporal features, expressed by the following formula: ； Where σ represents the sigmoid function, This represents a convolution operation with a filter size of 7×7.

4. The telecommunications fraud detection method based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 2, characterized in that, The GAIN is an improved GAN framework based on cue vectors. It uses the cue matrix of real observations to perform an unsupervised data generation model, which consists of a generator, a discriminator, and a cue vector matrix. The generator conditionally fills in missing parts based on partial components of the observed data and outputs a complete vector. The discriminator receives the complete vector to distinguish between the observed values ​​and the imputed values; The cue vector reveals the missing state of some samples to the discriminator, enabling it to focus on the quality of filling in specific components and ensuring that the generator follows the true data distribution; Generative Adversarial Networks (GANs) consist of a generator G and a discriminator D; the generator G learns a mapping G(z) and attempts to map a random noise vector z to real data; The discriminator D then attempts to find a mapping D(・) to determine the true probability of the input data; The generator learns complex data patterns through alternating training, with the goal of mapping the noise distribution Pz(z) to the real data distribution Pdata(x); while the discriminator is responsible for distinguishing between real data and generated data. The generator G and the discriminator D are two competitors, and are trained using minimax based on the value function V(D, G). Let Pz(z) be the input prior distribution, and Pdata(x) be the training data distribution; the discriminator By parameterizing the generator network weights, synthetic samples that are difficult to distinguish from real data are generated. Discriminator network weights are parameterized to distinguish between real and generated samples; the optimization goal of Generative Adversarial Network (GAN) is to find the optimal parameters. and The objective function is minimized as follows: ； Where x represents the real data, Pdata(x) represents the distribution of the real data, z is the input of the generator network, usually taken from the assumed prior data distribution Pz(z); D(x) and D(z) represent the outputs of the real data x and the generated data G(z) input to the discriminator network, respectively. The discriminator is guided to focus on the missing region by a cue matrix. During training, a binary cue matrix is ​​generated by randomly sampling the mask matrix Mmb. ; Let the original data matrix be mask matrix ,in express Not missing, hint matrix The generation process is formalized as follows: ； in, Let P be an independent and identically distributed Bernoulli sampling matrix with sampling probability P, and ⊙ be the element-wise multiplication of the Hadamard product; each element The generation rules are: ； The discriminator is guided by two mechanisms to determine known values ​​( The original value is retained with probability p, allowing the discriminator to focus on the valid data region and provide deterministic prompts; for missing regions ( Injecting 0.5 intermediate noise helps prevent the generator from getting trapped in local optima and generating random perturbations.

5. The telecommunications fraud detection method based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 4, characterized in that, The combined loss function is trained using a weighted loss function, which consists of discriminator and generator loss functions, enabling the generator G to generate more realistic samples and enabling the discriminator D to better distinguish between real samples and generated samples. The generator loss function includes: Adversarial loss measures the generator's ability to deceive the discriminator, and is based on the cross-entropy loss of the discriminator's output, used to deceive the discriminator. Reconstruction loss, or mean squared error loss, is used to measure the difference between the generated data and the known parts of the original data. ； in, It is a mask matrix. G is the output of the discriminator, and G is the output of the generator. This is used to avoid numerical instability in logarithmic operations. It is a hyperparameter with a value between 0.1 and 1.0, used to balance the adversarial loss and reconstruction accuracy, and to balance the two parts of the loss. This parameter is set to prevent numerical instability during logarithmic operations. The hyperparameter β controls the strength of this penalty term, typically between 0.1 and 0.

5. In addition, nonnegativity constraint loss, negative value penalty, prevents the generation of negative values ​​and ensures physical interpretability, is expressed as: ； in, The loss is a region-based loss function that uses certain weights to... Loss and Loss combination, for the weighted combination loss function, can be used to better segment regions and ensure a more stable training process; where α is an arbitrary constant, α∈(0,1); The objective function is optimized as follows: ； The discriminator loss function is based on cross-entropy loss and measures the discriminator's ability to distinguish between real and generated data. The optimization objective function is as follows: ； in, It is a mask matrix. This is the output of the discriminator, used to avoid numerical instability in logarithmic operations. This parameter is set to prevent numerical instability during logarithmic operations; E represents expectation, which is the expectation of the loss expression within the parentheses over the entire data distribution.

6. The method for detecting telecommunications fraud based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 1, characterized in that, In the integrated classification module CTCN-CSIAM-SwinTransformer, local spatiotemporal dynamic features are extracted by CTCN, then the features are optimized by CSIAM, and finally the global dependencies are captured by SwingTransformer to comprehensively capture the spatiotemporal correlation of the data. The CTCN module serves as a low-level feature extractor, integrating CNN and TCN. CNN extracts local spatial features, while TCN extracts local temporal features through causal convolution and dilated convolution. The CSIAM module is embedded between the CTCN module and the SwinTransformer module. It integrates convolutional features from different levels through a multi-scale depth-separable residual module while preserving details and global semantics. It also optimizes channel weights with channel attention and focuses on key regions and suppresses redundancy with spatial attention. The SwinTransformer module, as a high-level feature fusion unit, models global dependencies through a self-attention mechanism, captures long-distance temporal correlations, and makes up for the shortcomings of CTCN in long-range dependencies and global feature extraction.

7. The method for detecting telecommunications fraud based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 6, characterized in that, TCN consists of causal convolution, dilated convolution, and residual connections. Causal convolution ensures that at a given time t, the model depends only on historical data, forming a strict time constraint and maintaining model stability. Dilated convolution expands the receptive field through an exponential expansion factor, resolving long-distance time dependencies without increasing the number of layers. Residual connections alleviate gradient vanishing / exploding, maintaining performance stability. Assume the convolution filter is represented as F =( f 1, f 2,⋯, f K The input sequence is X=(x1,x2,⋯ ,x T If ), then the causal convolution at position t can be expressed as: ； Time series prediction models require that the output at the current time step completely cover all its previous inputs. When the kernel size is K... C The input time length is L time When, the required number of convolutional layers L C satisfy: ； TCN expands the receptive field by introducing a dilation factor. The receptive field w of a specific layer is based on the number of layers below the current layer i, the dilation parameter b, the input length Ltime, the number of topmost layers n required to cover the complete input history, and the number of zeros p required for each layer. Spacing is inserted between convolutional kernel elements, with a spacing of b−1. The actual receptive field calculation formula after dilation (b=1 in standard CNNs) is as follows: ； ； To ensure that the input and output sequences have the same length, zero padding must be applied to the end of the sequence. Dilated convolution is used to achieve an exponentially expanded receptive field to capture local feature information at different scales. The operation definition of dilated causal convolution is as follows: ； Where f(t) represents the dilated convolution operation at time step t, f:{0,…,k−1}→ℝ represents the filter, and d is the dilation factor. d f represents a convolution operation with stride d; before the training samples are used as input to the CTCN module, missing values ​​are padded with the annual average to complete the initialization; A four-layer TCN is used, with a kernel size of 3 and dilation factors d of 1, 2, 4 and 8 respectively. Dilation is added to the causal convolution to skip a certain number of time steps in the input sequence. Dilated causal convolution achieves a larger receptive field through interval sampling, which makes up for the limitations of causal convolution. By processing sequence data through causal dilated convolution, the model can effectively capture short-term dependencies. In TCN, the residual module is passed across layers, and a dropout layer is added to the residual module for regularization to prevent overfitting; The input to the residual module is summed with the output after two convolution operations using an identity mapping to ensure dimensionality consistency. Residual connections train the network by directly adding the input to the output of a convolutional layer. The residual connection expression is as follows: ； Where x represents the input data, and F(x) represents the result of the input nonlinear transformation.

8. The telecommunications fraud detection model based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 7, characterized in that, By combining the residual module Inception with the hybrid attention module CSAM, a residual module CSIAM based on the hybrid attention mechanism is constructed. Among them, the Hybrid Attention Module (CSAM) calculates attention weights through channel and spatial dimensions, CAM generates channel weights through global pooling and MLP to strengthen key channels, and SAM generates spatial weights through convolution and pooling to focus on important positions, thus highlighting effective information and suppressing redundancy. The residual module Inception achieves multi-scale feature fusion through a pyramid structure, integrating convolutional features from different levels, preserving details and global semantics, and solving the problem of insufficient semantics at a single scale. In addition, the model extracts local features from multiple scales and dynamically enhances key regions with attention. The combination of these two methods not only accurately captures multi-scale features but also enhances the discriminative power of low-contrast features, thereby improving the model's ability to process complex spatiotemporal data. The flow of input feature data is divided into two branches: One component first passes through the channel attention module, multiplies the resulting weight matrix with the original input data, and then passes through the spatial attention module and is multiplied by the corresponding weights. Another shortcut similar to ResNet is direct mapping, where the two data streams are ultimately added together and output through an activation function; the attention features output by each module are combined with the original features as follows: ； in, It is the output of channel attention. It is the output after applying spatial attention to the channel attention result.

9. The telecommunications fraud detection model based on data repair and integrated classification fusion using a hybrid attention mechanism as described in claim 8, characterized in that, The Swin Transformer module includes Layer Normalization (LN), Multi-head Self-Attention, Multi-layer Perceptron (MLP), and Residual Connections. The MLP contains the GELU activation function, while LN and Residual Connections are used for stable training and feature preservation, respectively. The computational complexity is reduced through window self-attention, the shifted window strategy enhances cross-window information interaction, and hierarchical feature representation efficiently captures long-distance dependencies and context, optimizing computational resources while maintaining architectural scalability. MLP introduces nonlinearity to learn complex patterns, and the classification head outputs predictions, achieving efficient fusion of local and global features, represented as: ； Among them, z l For the input of the l-th layer, z l1 This represents the output of the W-MSA layer before the MLP layer in the (l+1)-th module; (l+1)-th is the input of the (l+1)-th module (i.e., the output of the l-th module), z l+1 ,z l1+1 This represents the intermediate output after the SW-MSA layer and before the MLP layer in the (l+1)-th module, z. l-1 This represents the output characteristics of the previous layer; before each MHSA module and MLP, an LN layer is used to normalize the input; and after each module, residual connections are used to facilitate information flow. W-MSA is built upon multi-head self-attention constrained to a non-overlapping window. Compared to global self-attention, it helps reduce computational complexity, as expressed below: ； Where, 𝑄, 𝐾, and 𝑉 are the query, key, and value matrices, respectively, and 𝑊 𝐾 and 𝑊 𝑉 A learnable weight matrix for keys and values; SW-MSA achieves cross-window information exchange through window sliding and overlapping, thus preserving the ability to model global context information, as shown below: ； Among them, W 𝐾,shifted and W 𝑉,shifted The learnable weight matrix for keys and values ​​after sliding; At each stage, the number of patches is reduced by merging, and a hierarchical feature representation is generated after each stage. During the operation, adjacent patches are first concatenated, and then mapped to a lower dimension through linear projection, as shown below: ； Among them, W m and b m These are the learnable weight matrix and the bias term, respectively. The size of the patch determines the data segmentation method, which affects both the details captured and the computational efficiency.

10. A telecom fraud detection model based on data repair and ensemble classification fusion using a hybrid attention mechanism, characterized in that... include: A generative adversarial interpolation network CSAM-GAIN based on a hybrid attention mechanism is used for repairing telecom fraud data. It captures the spatiotemporal correlation of data through a concatenated channel and spatial attention mechanism, and optimizes the generation of complete data that closely approximates the real distribution by combining adversarial loss, reconstruction loss and non-negativity constraint loss. The integrated classification module CTCN-CSIAM-SwinTransformer includes three modules: CTCN, CSIAM, and SwinTransformer, forming a three-layer processing system of local-feature enhancement-global. The CTCN module fuses CNN and TCN in parallel to extract local spatial features and temporal features respectively. The CSIAM module enhances feature discriminativeness through multi-scale residual structures and hybrid attention; the SwinTransformer module captures global spatiotemporal dependencies, making up for the limitations of local feature extraction.

Images