Number multi-dimensional risk assessment method and system based on dynamic feature extraction

Abstract

The invention relates to a number multi-dimensional risk assessment method and system based on dynamic feature extraction, and the method comprises the steps: obtaining multi-source heterogeneous telecommunication abnormal data of a client, carrying out the calculation power assessment, and carrying out the dynamic sampling according to a calculation power score, and forming a local training set; feature extraction is carried out on the text data and the image data, and multi-modal feature alignment and fusion are realized through a generative adversarial network and an attention mechanism; in a federated learning framework, dynamically adjusting a regularization coefficient according to the difference between a local model and a global model by utilizing a reinforcement learning model, and guiding a client to carry out adaptive local training; the server performs intelligent weighted aggregation on client parameters based on data quality and model similarity, and updates a global model; and finally, a user heterogeneous graph is constructed based on the global model and the multi-modal features, neighbor selection is optimized through reinforcement learning, and abnormal users are identified by using graph neural network aggregation information. According to the method and the device, on the premise of protecting data privacy, the risk assessment accuracy of the number is effectively improved.

Description

Technical Field

[0001] This application relates to the field of information security technology, and in particular to a method and system for multi-dimensional risk assessment of phone numbers based on dynamic feature extraction. Background Technology

[0002] With the rapid development of information and communication technologies, telecommunications network violations are increasingly characterized by cross-platform, multimodal, and covert nature. Illegal methods have evolved from early simple text messages to complex attacks combining fake images, forged voice messages, and even synthetic videos, posing a serious challenge to user property security and social stability.

[0003] Currently, the mainstream methods for identifying telecommunications violations or number risks are based on feature analysis and rule matching of single-modal data. For example, statistical analysis is performed only on the time sequence of call behavior, or text classification models are used to identify keyword patterns in illegal text messages. Another approach attempts to use a federated learning framework to address data privacy constraints. The basic idea is that each data holder trains a model locally and only uploads the model parameters to a central server for aggregation, thereby obtaining a general global risk identification model.

[0004] The aforementioned technologies rely solely on single-modal information such as call behavior or text content, making it difficult to effectively identify complex illegal methods that combine text, images, audio, and video. They also lack the ability to respond to new forms of spoofing and are not convenient for identifying abnormal numbers.

[0005] Based on this, this application provides a method and system for multi-dimensional risk assessment of phone numbers based on dynamic feature extraction. Summary of the Invention

[0006] To address the shortcomings of relying solely on single-modal information such as call behavior or text content, which makes it difficult to effectively identify complex illegal methods combining text, images, audio, and video, and lacks the ability to respond to new forms of spoofing, and is not convenient for identifying abnormal numbers, this application provides a multi-dimensional risk assessment method and system for numbers based on dynamic feature extraction.

[0007] Firstly, this application provides a multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction, employing the following technical solution: including: Acquire multi-source heterogeneous telecommunications anomaly data from clients, evaluate the computing power of each client, and generate a corresponding computing power score. The computing power score is calculated based on the client's CPU performance, GPU memory capacity, and remaining memory percentage. Based on each client's computing power score, the corresponding data sampling ratio is dynamically determined. For clients with computing power scores below a preset threshold, their local datasets are stratified and sampled according to the data sampling ratio to form a sampled dataset for training. For clients with computing power scores above the preset threshold, their complete local datasets are used. For text data in each client's sampled dataset, a pre-trained language model is used to extract text feature vectors; for image data in each client's sampled dataset, a convolutional neural network is used to extract global and local image features; a generative adversarial network training framework is adopted to align the text feature vectors, global image features, and local image features into a unified semantic space, and to perform weighted fusion through a dynamic attention mechanism to generate multimodal features for each client. A global recognition model is distributed to each client. Each client uses its local sampled dataset and the corresponding multimodal features to train the received global recognition model locally. The loss function of the local training includes a regularization term. The parameter value of the regularization term is dynamically selected by the reinforcement learning model based on the difference between the current client's model classification head weight and the global recognition model's classification head weight. The system receives the local model parameters trained by each client, calculates adaptive aggregation weights based on the data quality scores of each client and the similarity between their local models and the global recognition model, and then performs weighted aggregation on the local model parameters uploaded by all clients accordingly to update the global recognition model. Based on the updated global recognition model and the multimodal features, a user heterogeneous graph is constructed, and reinforcement learning is used to dynamically optimize neighbor selection. Information is aggregated through graph neural networks to identify abnormal users.

[0008] Preferably, the step of acquiring multi-source heterogeneous telecommunications anomaly data from clients, evaluating the computing power of each client, and generating a corresponding computing power score includes: The system retrieves telecommunications anomaly data stored locally by multiple clients. This telecommunications anomaly data includes text data, image data, and communication behavior data. For each client, collect the hardware performance metrics of its computing node. The hardware performance metrics include at least CPU performance, GPU memory capacity, and current remaining memory capacity. Based on hardware performance indicators, a corresponding computing power score for the client is generated through weighted calculation and normalization. Among them, CPU performance, GPU memory capacity, and the proportion of the remaining memory capacity in the total memory are used as calculation items, and each is configured with a corresponding weight coefficient.

[0009] Preferably, for text data in each client's sampled dataset, a pre-trained language model is used to extract text feature vectors; for image data in each client's sampled dataset, a convolutional neural network is used to extract global and local image features; a generative adversarial network training framework is employed to align the text feature vectors, global image features, and local image features into a unified semantic space, and a dynamic attention mechanism is used for weighted fusion to generate multimodal features for each client, including: For the text data in the sampled dataset of each client, a pre-trained BERT model is used to encode the text feature vector. For the image data in the sampled dataset of each client, the ResNet model is used to extract features to obtain the global image features of the image; at the same time, features are extracted from multiple preset local regions in the image to obtain multiple local image features. A generative adversarial network (GAN) is constructed, which includes a text encoder, an image encoder, and a discriminator. The text feature vector, the global image features, and multiple local image features are respectively input into the corresponding encoders. The discriminator distinguishes the modality categories to which the features output by different encoders belong. Through adversarial training, the features of different modalities are forced to be mapped to a unified semantic space, thereby generating unified text features and unified image features after semantic space alignment. Based on the unified text features and unified image features after semantic space alignment, the attention weights of the unified text features on the global image features and each of the local image features are calculated through a dynamic attention mechanism. The global image features and local image features are then weighted and summed according to the attention weights to obtain the fused image feature representation. The unified text features are concatenated with the image feature representation to generate the multimodal features corresponding to the client.

[0010] Preferably, for the image data in the sampled dataset of each client, a ResNet model is used for feature extraction to obtain the global image features of the image; simultaneously, feature extraction is performed on multiple preset local regions in the image to obtain multiple local image features, including: For each input image in the sampled dataset, it is fed into the pre-trained ResNet model; The high-level semantic representation of the input image is extracted through the forward propagation of the ResNet model, which serves as the global image feature of the corresponding image. A preset masking strategy is applied to the input image to generate multiple local image regions. The masking strategy is defined by a binary matrix of the same size as the image. Pixels with a value of 1 in the matrix constitute the local region to be extracted, and pixels with a value of 0 are non-masked regions. Each local image region defined by the mask is input into the ResNet model to extract the high-level semantic representation corresponding to each region, forming a set of local image features.

[0011] Preferably, the step of distributing the global recognition model to each client, and each client using its local sampled dataset and the corresponding multimodal features to locally train the received global recognition model, includes: The global recognition model for the current round is distributed to all clients participating in the training, and each client uses the received global recognition model as the initial state of its local model. For each client, observe the current state based on the deployed reinforcement learning model. The current state is the local model classification head weight and the classification head weight of the global recognition model after the previous training round of the corresponding client. The control reinforcement learning model outputs coefficients for the regularization term used in the loss function during the current local training process, based on the current state. Based on the coefficient values ​​selected by the reinforcement learning model, local training is performed, and its loss function is defined as the sum of cross-entropy loss and regularization loss. The regularization loss is determined by the coefficient values ​​and the difference between the global recognition model parameters and the local model parameters. During local training, the control client constructs training experience into experience tuples and stores them in the local experience pool. At the same time, based on the model classification head weights uploaded by each client, the similarity between clients is calculated, and each client is matched with a collaborative client with similar data distribution. This allows the client to sample experience tuples from the collaborative client's experience pool and update its own reinforcement learning model. Receive updated local model parameters uploaded by the client after local training is completed.

[0012] Preferably, the control reinforcement learning model outputs coefficient values ​​for the regularization term used in the loss function during the current local training process, based on the current state, including: Define the input state space of the reinforcement learning model, which is a combined vector consisting of the local model classification head weights of the current client and the classification head weights of the global recognition model; Define the output action space of the reinforcement learning model, which is a set of continuous or discrete actions used to adjust the coefficient values ​​of the regularization term; The reinforcement learning model receives a state space vector, calculates and outputs specific actions through its policy network, and maps the specific actions to the coefficient values ​​of the regularization term used in the loss function during this local training process. After the client completes local training based on the selected coefficient values, an immediate reward is calculated based on the performance changes of the local model and the global recognition model on the validation set. The immediate reward is determined by at least the improvement in the accuracy of the local model, the improvement in the accuracy of the global recognition model, and the training round factor. Using the immediate rewards and state transition information, the policy network parameters of the reinforcement learning model are updated through the policy gradient algorithm, thereby optimizing its decision-making ability to select the optimal coefficient values ​​for different states.

[0013] Preferably, the step of constructing a user heterogeneous graph based on the updated global recognition model and the multimodal features, dynamically optimizing neighbor selection using reinforcement learning, and aggregating information through a graph neural network to identify abnormal users includes: Using users in the system as nodes and the interaction relationships between users as edges, a user heterogeneous graph is constructed. The feature vector of each node is the multimodal feature of the corresponding user extracted or encoded by the updated global recognition model, and each edge is attached with the relationship feature of its corresponding interaction relationship type. For each node in the user heterogeneous graph, sort it according to its similarity with all neighbor nodes in the feature space to generate an initial candidate list of neighbors; Using a reinforcement learning model, the neighbor selection threshold is dynamically optimized for each relationship type. The state of the reinforcement learning model includes the current threshold, the average neighbor feature distance of nodes in the graph, and the node degree. Its actions are to increase or decrease the threshold. Its reward is based on the change in the neighbor aggregation effect before and after the threshold adjustment and the performance index of the auxiliary classification task. The optimized neighbor selection threshold is applied to filter the final set of neighbors for information aggregation from the initial neighbor candidate list; Based on the filtered neighbor set, message passing and feature aggregation are performed through a relation-specific graph convolutional network layer to update the embedding representation of each node in the user heterogeneous graph; Based on the updated node embedding representation, a classifier is used to predict whether the corresponding node is a normal user or an abnormal user, and the identification result is output.

[0014] Secondly, this application discloses a multi-dimensional risk assessment device for phone numbers based on dynamic feature extraction, which adopts the following technical solution, including: The data acquisition module is used to acquire multi-source heterogeneous telecommunications anomaly data from clients, evaluate the computing power of each client, and generate a corresponding computing power score. The computing power score is calculated based on the client's CPU performance, GPU memory capacity, and remaining memory percentage. The data sampling module is used to dynamically determine the corresponding data sampling ratio based on the computing power score of each client. For clients with computing power scores below a preset threshold, their local datasets are stratified and sampled according to the data sampling ratio to form a sampled dataset for training. For clients with computing power scores above the preset threshold, their complete local datasets are used. The feature generation module is used to extract text feature vectors from text data in each client's sampled dataset using a pre-trained language model; and to extract global and local image features from image data in each client's sampled dataset using a convolutional neural network. A generative adversarial network training framework is adopted to align the text feature vectors, global image features, and local image features into a unified semantic space, and to perform weighted fusion through a dynamic attention mechanism to generate multimodal features for each client. The local training module is used to distribute the global recognition model to each client. Each client uses its local sampled dataset and the corresponding multimodal features to train the received global recognition model locally. The loss function of the local training includes a regularization term. The parameter value of the regularization term is dynamically selected by the reinforcement learning model based on the difference between the current client's model classification head weight and the global recognition model classification head weight. The global update module is used to receive the local model parameters trained by each client, calculate the adaptive aggregation weight based on the data quality score of each client and the similarity between its local model and the global recognition model, and perform weighted aggregation on the local model parameters uploaded by all clients accordingly to update the global recognition model. An anomaly detection module is used to construct a user heterogeneous graph based on the updated global detection model and the multimodal features, dynamically optimize neighbor selection using reinforcement learning, and aggregate information through a graph neural network to identify abnormal users.

[0015] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed, such as the multi-dimensional risk assessment method for numbers based on dynamic feature extraction described above.

[0016] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding the multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction.

[0017] In summary, this application's system acquires multi-source heterogeneous data (text, images, and communication behaviors) from multiple clients and dynamically evaluates the computing power of each client's computing nodes based on their CPU, GPU, and memory status. Data is then adaptively hierarchically sampled based on the computing power score to balance computational load and training efficiency. Subsequently, pre-trained models are used to extract features from the text and image data, and a generative adversarial network (GAN) is used to align features from different modalities to a unified semantic space. A dynamic attention mechanism is then used to fuse these features, forming a unified multimodal feature representation for each client. Under the federated learning framework, the server distributes the global model to each client. Each client trains using local data and multimodal features. The key innovation lies in the introduction of a reinforcement learning model. This model dynamically determines the optimal regularization coefficient based on the difference between the classification head weights of the local and global models, thereby controlling the deviation of local training from the global model and adapting to different data distributions. After training, each client uploads its parameters, and the server intelligently weights and aggregates them based on data quality and model similarity to generate a more robust updated global model. Finally, a heterogeneous user graph (nodes representing users and edges representing interaction relationships) is constructed based on the global identification model and the multimodal features of all users. Reinforcement learning is then used to dynamically optimize the neighbor selection threshold for each node in the graph, filtering out interference. Next, a relationship-specific graph neural network aggregates neighbor information and updates node representations, ultimately accurately identifying abnormal users. This approach solves the training efficiency problem in heterogeneous environments through adaptive sampling, enhances the identification capability for complex illegal activities through multimodal feature alignment and fusion, improves the model's generalization and robustness under non-independent and identically distributed data through reinforcement learning-driven federated aggregation, and uncovers the association patterns of abnormal groups through graph structure analysis and dynamic neighbor selection. Thus, accurate identification of telecommunications illegal risks is achieved while strictly protecting data privacy. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction.

[0019] Figure 2 This is a structural block diagram of a multi-dimensional risk assessment device for phone numbers based on dynamic feature extraction. Detailed Implementation

[0020] The following combination Figure 1 - Figure 2 This application will be described in further detail.

[0021] This application aims to construct an adaptive intelligent system that can both strictly protect data privacy and efficiently coordinate multi-party data for accurate risk identification. It is not an isolated improvement of a single algorithm, but rather a systematic solution to four major challenges—data silos, heterogeneous computing power, modal differences, and hidden relationships—through the design of a series of interconnected and progressively advancing collaborative mechanisms. Specifically, the solution starts with the most basic data acquisition and resource adaptation, ensuring the training feasibility of heterogeneous participants through computing power-aware dynamic sampling. Then, it utilizes generative adversarial and attention mechanisms to break down barriers between multimodal information such as text and images, forming a unified feature representation. Building upon this, it embeds the decision-making capabilities of reinforcement learning into the local training process of federated learning, achieving adaptive modeling and robust aggregation for complex data distributions. Finally, it combines a global model with graph structure analysis, using reinforcement-guided neighbor selection to mine risk propagation patterns in graph neural networks, achieving a three-dimensional risk assessment from individual to group, and from features to associations, to detect the health of phone numbers and determine whether a number is abnormal or normal.

[0022] Reference Figure 1 The embodiments of this application include at least steps S10 to S60.

[0023] S10: Obtain multi-source heterogeneous telecommunications anomaly data from clients, evaluate the computing power of each client, and generate a corresponding computing power score. The computing power score is calculated based on the client's CPU performance, GPU memory capacity, and remaining memory percentage.

[0024] S20: Based on the computing power score of each client, dynamically determine the corresponding data sampling ratio. For clients with computing power scores below a preset threshold, perform stratified sampling of their local datasets according to the data sampling ratio to form a sampled dataset for training. For clients with computing power scores above a preset threshold, use their complete local datasets.

[0025] S30: For text data in each client's sampled dataset, a pre-trained language model is used to extract text feature vectors; for image data in each client's sampled dataset, a convolutional neural network is used to extract global and local image features; a generative adversarial network training framework is adopted to align text feature vectors, global image features, and local image features into a unified semantic space, and to perform weighted fusion through a dynamic attention mechanism to generate multimodal features for each client.

[0026] S40: Distribute the global recognition model to each client. Each client uses its local sampled dataset and corresponding multimodal features to train the received global recognition model locally.

[0027] S50 receives the local model parameters trained by each client, calculates adaptive aggregation weights based on the data quality scores of each client and the similarity between their local models and the global recognition model, and then performs weighted aggregation on the local model parameters uploaded by all clients to update the global recognition model.

[0028] S60 constructs a heterogeneous user graph based on the updated global recognition model and multimodal features, dynamically optimizes neighbor selection using reinforcement learning, and aggregates information through graph neural networks to identify abnormal users.

[0029] The loss function for local training includes a regularization term, the parameter value of which is dynamically selected by the reinforcement learning model based on the difference between the current client's model classification head weights and the global recognition model's classification head weights.

[0030] Specifically, the system first collects multi-source heterogeneous telecommunications anomaly data from various clients and dynamically evaluates computing power based on their hardware configurations. This allows for differentiated data sampling, ensuring that even nodes with weaker computing power can participate efficiently. For the sampled data, text and image features are extracted separately, and a generative adversarial network (GAN) is used to align different modal features to a unified semantic space. These features are then fused using an attention mechanism to form a unified multimodal feature for each client. In the federated learning phase, the server distributes a global model, and each client trains using local features. The core innovation lies in using a reinforcement learning model to dynamically determine the coefficient of the regularization term in the training loss based on the weight difference between the local and global model classification heads, thus adaptively adjusting the local training direction. After training, the server intelligently aggregates the parameters of each client based on data quality and model similarity, updating the global model. Finally, a heterogeneous graph is constructed based on this model and all user features. Reinforcement learning is then used to optimize the neighbor selection strategy for each node in the graph, and a graph neural network is used to aggregate related information, accurately identifying abnormal users and potential abnormal networks. Thus, through a multi-layered adaptive mechanism, the system significantly improves the accuracy and robustness of risk identification in complex heterogeneous environments while protecting data privacy.

[0031] In some embodiments, step S10 specifically includes the following steps: obtaining telecommunications anomaly data stored locally by multiple clients, the telecommunications anomaly data including text data, image data and communication behavior data; for each client, collecting the hardware performance indicators of its computing node, the hardware performance indicators including at least CPU performance, GPU memory capacity and current remaining memory capacity; based on the hardware performance indicators, generating the corresponding client's computing power score through weighted calculation and normalization processing, wherein the proportion of CPU performance, GPU memory capacity and remaining memory capacity in the total memory is used as calculation items, and each is configured with a corresponding weight coefficient.

[0032] Client i's computing power score Calculated using the following formula: ; in, This represents the normalized CPU performance metrics. GPU memory capacity (unit: GB); and These represent the remaining memory and total memory capacity, respectively; α, β, and γ are preset weighting coefficients, and satisfy α+β+γ=1.

[0033] Furthermore, sampling ratio Based on computing power score Determined dynamically using the following Sigmoid function: ; in, This represents the data sampling ratio for client i; This is a scaling factor used to adjust the rating scale. The offset parameter controls the center of the function. This function smoothly maps the computing power score to the (0,1) interval. The higher the score, the closer the sampling ratio is to 1, thus enabling the use of more complete data for high computing power nodes.

[0034] Specifically, the system first collects multi-source heterogeneous data from the local storage of various clients (such as operators and enterprises), including text content, associated images, and communication records related to risky behaviors. To adapt to the heterogeneous computing environment of subsequent distributed training, the system simultaneously collects the core hardware performance indicators of each client's computing node—CPU processing power, GPU memory capacity, and current available memory percentage. Based on these indicators, a quantified computing power score is calculated using a preset weighted summation and normalization formula. This score will serve as a key basis to drive differentiated data processing and training strategies for different computing power nodes in subsequent steps, thereby optimizing global training efficiency and resource utilization.

[0035] In some embodiments, step S30 specifically includes the following steps: encoding text data in the sampled dataset of each client using a pre-trained BERT model to obtain text feature vectors; extracting features from image data in the sampled dataset of each client using a ResNet model to obtain global image features; simultaneously, extracting features from multiple preset local regions in the image to obtain multiple local image features; constructing a generative adversarial network containing a text encoder, an image encoder, and a discriminator, inputting the text feature vectors, global image features, and multiple local image features into the corresponding encoders, and using the discriminator to distinguish the modality categories to which the features output by different encoders belong, forcing the features of different modalities to be mapped to a unified semantic space through adversarial training, generating unified text features and unified image features after semantic space alignment; based on the unified text features and unified image features after semantic space alignment, calculating the attention weights of the unified text features on the global image features and each local image feature through a dynamic attention mechanism, and performing a weighted summation of the global image features and local image features according to the attention weights to obtain a fused image feature representation; concatenating the unified text features and image feature representation to generate multimodal features for the corresponding client.

[0036] Text feature T and the kth local image feature Attention weights between The calculation is as follows: ; in, The attention weights for the k-th local image feature are represented by T; T is the unified text feature vector. Let be the feature vector of the j-th local image; N is the total number of local regions. Scoring function. , , Let be a learnable weight matrix, and b be a bias term. The global and local features are weighted and summed according to the weights to obtain the fused image feature representation.

[0037] Specifically, firstly, deep feature extraction is performed on text and image data using BERT and ResNet models respectively, obtaining text vectors and global and local features of the images. Then, a generative adversarial network framework is used to train text and image encoders, forcing features from different modalities to map to a unified semantic space, eliminating modal differences. Building on this, a dynamic attention mechanism is introduced to calculate the attention weights of text features to various local regions of the image, thereby weighted and fused global and local image features to form an image feature representation highly correlated with the text semantics. Finally, the unified text features and the fused image features are concatenated to generate a standardized, complementary, and semantically consistent multimodal feature representation for each client.

[0038] Furthermore, considering the problem of extracting image features using a combination of global and local methods, the corresponding processing steps are as follows: For each input image in the sampled dataset, it is input into a pre-trained ResNet model; through the forward propagation of the ResNet model, the high-level semantic representation of the input image is extracted as the global image feature of the corresponding image; a preset masking strategy is applied to the input image to generate multiple local image regions; the masking strategy is defined by a binary matrix of the same size as the image, where pixels with a value of 1 constitute the local region to be extracted, and pixels with a value of 0 are non-masked regions; each local image region defined by the mask is input into the ResNet model to extract the high-level semantic representation corresponding to each region, forming a set of local image features.

[0039] Specifically, for each input image, a pre-trained ResNet model is first used to extract its overall high-level semantic representation as global image features to capture the overall content and contextual information of the image. To deeply identify potential local tampering or key details in the image, a masking strategy based on a binary mask matrix is ​​adopted. This strategy predefines multiple local regions in the image that require special attention. By inputting the image pixel blocks corresponding to the masked regions into the same ResNet model, the high-level semantic representation of each local region is extracted, thereby obtaining a set of local image features. This method, by combining a grasp of global image information with fine feature extraction from multiple local regions, provides a more comprehensive image feature foundation for subsequent multimodal feature fusion and deep risk analysis.

[0040] In some embodiments, step S40 specifically includes the following steps: distributing the global recognition model of the current round to all clients participating in training, with each client using the received global recognition model as the initial state of its local model; for each client, observing the current state based on the deployed reinforcement learning model, the current state being the classifier head weights of the local model after the previous round of training and the classifier head weights of the distributed global recognition model; controlling the reinforcement learning model to output the coefficient values ​​of the regularization term used in the loss function during the current local training process based on the current state; and performing local training based on the coefficient values ​​selected by the reinforcement learning model, with the loss function defined as the sum of cross-entropy loss and regularization loss, the regularization loss being jointly determined by the coefficient values ​​and the difference between the parameters of the global recognition model and the parameters of the local model. During local training, the control client constructs training experience into experience tuples and stores them in the local experience pool. At the same time, based on the model classification head weights uploaded by each client, the similarity between clients is calculated, and each client is matched with a collaborative client with similar data distribution. This allows the client to sample experience tuples from the collaborative client's experience pool and update its own reinforcement learning model. The system also receives updated local model parameters uploaded by the client after completing local training.

[0041] Specifically, after the server distributes the global model, each client uses the reinforcement learning model as a decision-maker. By analyzing the differences in the classification head weights between the local and global models, the strength of regularization constraints during training is dynamically determined, thereby balancing model individuality and global consistency. After performing local training based on this coefficient, the client stores the training process as experience data and shares it with collaborating clients in an experience pool based on model similarity. This improves the training efficiency and effectiveness of the reinforcement learning strategy. Finally, the updated local model parameters are uploaded, completing this round of collaborative learning.

[0042] Furthermore, considering the specific operational mechanism of the reinforcement learning decision-maker in federated learning, the corresponding processing steps are as follows: Define the input state space of the reinforcement learning model, which is a combined vector consisting of the local model classification head weights and the global recognition model classification head weights of the current client; define the output action space of the reinforcement learning model, which is a set of continuous or discrete actions used to adjust the coefficient values ​​of the regularization term; receive the state space vector through the reinforcement learning model, calculate and output specific actions through its policy network, and map the specific actions to the coefficient values ​​of the regularization term used in the loss function during this local training; after the client completes local training based on the selected coefficient values, calculate the immediate reward based on the performance changes of the local model and the global recognition model on the validation set. The immediate reward is determined by at least the improvement in the accuracy of the local model, the improvement in the accuracy of the global recognition model, and the training epoch factor; use the immediate reward and state transition information to update the policy network parameters of the reinforcement learning model through the policy gradient algorithm to optimize its decision-making ability to select the optimal coefficient values ​​for different states.

[0043] Specifically, the difference in classification head weights between the local and global models is used as the state input. The policy network outputs actions (coefficients) that determine the regularization strength, guiding local training to achieve a balance between personalization and global consistency. After training, rewards are calculated based on the performance improvement of the local and global models on the validation set, and the policy network is updated using a policy gradient algorithm based on the rewards and state transition information. Its core function is to enable the system to automatically adapt to the data distribution characteristics of different clients, dynamically optimize the local training process, and thus significantly improve the convergence speed, final performance, and overall robustness of the federated model in heterogeneous data environments.

[0044] Furthermore, in step S50, adaptive aggregation weights are calculated based on the data quality scores of each client and the similarity between their local and global models. The adaptive aggregation weight for client i in round t is... Determined by the following formula: ; in, This represents the aggregate weight of client i in round t; ∈[0,1] represents the data quality score based on historical reward normalization; a larger value indicates higher data quality. Classification head weights for client-side local models Compared with the previous round of global model classification head weights The cosine similarity is denoted by λ, where λ is a hyperparameter balancing data quality and model similarity, and K is the total number of clients participating in the aggregation. This weighting method ensures that clients with high data quality and consistent with the global direction contribute more to the global model update.

[0045] In some embodiments, step S60 specifically includes the following steps: constructing a user heterogeneous graph with users in the system as nodes and interaction relationships between users as edges, wherein the feature vector of each node is the multimodal feature of the corresponding user extracted or encoded by the updated global recognition model, and each edge is attached with the relationship feature of its corresponding interaction relationship type; for each node in the user heterogeneous graph, sorting it according to its similarity with all neighbor nodes in the feature space to generate an initial neighbor candidate list; using a reinforcement learning model to dynamically optimize the neighbor selection threshold for each relationship type; wherein the state of the reinforcement learning model includes the current threshold, the average neighbor feature distance of the nodes in the graph, and the node degree, and its action is to increase or decrease the threshold, and its reward is based on the change in the neighbor aggregation effect before and after adjusting the threshold and the performance index of the auxiliary classification task; applying the optimized neighbor selection threshold to filter out the final set of neighbors for information aggregation from the initial neighbor candidate list; based on the filtered neighbor set, performing message passing and feature aggregation through a relationship-specific graph convolutional network layer to update the embedding representation of each node in the user heterogeneous graph; based on the updated node embedding representation, performing prediction through a classifier to identify the corresponding node as a normal user or an abnormal user, and outputting the identification result.

[0046] For relation type r, node In the Layer Embedded Representation Update as follows: ; in, Represents a node In the Layer embedding; Let 'R' be the GELU activation function; and R be the set of all relation types. For nodes under relation r The set of neighbors; For the neighbor of relation r For nodes Attention coefficient; For relation r in the th... Layer-specific learnable weight matrices; This indicates embedding the node one level above the neighbor node. With edge features Perform the splicing; the final This is a residual connection. The process updates the final representation of all nodes through multiple iterations for classification.

[0047] Specifically, firstly, a heterogeneous graph is constructed using users as nodes and interaction relationships as edges. Node features are multimodal features, and edges carry relationship attributes. For each node, an initial neighbor list is generated based on feature similarity. Subsequently, a reinforcement learning model, for each relationship type, is optimized by increasing or decreasing the threshold, using the current threshold, average node distance, and degree as states. The reward signal comes from changes in aggregation effect and classification performance. After applying the optimized threshold to filter out a high-quality neighbor set, neighbor information is aggregated through a relationship-specific graph convolutional network to update node embeddings. Finally, the abnormal user judgment result is output by a classifier. This method significantly improves the accuracy and robustness of identifying abnormal behavior of individuals and groups in complex social topologies by adaptively filtering spoofed nodes and noisy relationships in the network using reinforcement learning.

[0048] The implementation principle of a multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction in this application is as follows: The system acquires multi-source heterogeneous data (text, images, communication behavior) from multiple clients and dynamically evaluates the computing power of each client's computing node based on its CPU, GPU, and memory status. Based on the computing power score, the data is adaptively hierarchically sampled to balance computational load and training efficiency. Subsequently, pre-trained models are used to extract features from the text and image data respectively, and a generative adversarial network is used to align the features of different modalities to a unified semantic space. Then, a dynamic attention mechanism is used for fusion to form a unified multi-modal feature representation for each client. Under the federated learning framework, the server distributes the global model to each client. Each client trains using local data and multi-modal features. The key innovation lies in the introduction of a reinforcement learning model. This model dynamically determines the optimal regularization coefficient based on the difference between the classification head weights of the local and global models, thereby controlling the degree of deviation of local training from the global model and adapting to different data distributions. After training, each client uploads parameters, and the server performs intelligent weighted aggregation based on data quality and model similarity to generate a more robust updated global model. Finally, a heterogeneous user graph (nodes representing users and edges representing interaction relationships) is constructed based on the global identification model and the multimodal features of all users. Reinforcement learning is then used to dynamically optimize the neighbor selection threshold for each node in the graph, filtering out interference. Next, a relationship-specific graph neural network aggregates neighbor information and updates node representations, ultimately accurately identifying abnormal users. This approach solves the training efficiency problem in heterogeneous environments through adaptive sampling, enhances the identification capability for complex illegal activities through multimodal feature alignment and fusion, improves the model's generalization and robustness under non-independent and identically distributed data through reinforcement learning-driven federated aggregation, and uncovers the association patterns of abnormal groups through graph structure analysis and dynamic neighbor selection. Thus, accurate identification of telecommunications illegal risks is achieved while strictly protecting data privacy.

[0049] Figure 1 This is a flowchart illustrating a multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0050] Based on the same technical concept, referring to Figure 2 This application also provides a multi-dimensional risk assessment device for phone numbers based on dynamic feature extraction, which adopts the following technical solution: The device includes: The data acquisition module is used to acquire multi-source heterogeneous telecommunications anomaly data from clients, evaluate the computing power of each client, and generate a corresponding computing power score. The computing power score is calculated based on the client's CPU performance, GPU memory capacity, and remaining memory percentage. The data sampling module is used to dynamically determine the corresponding data sampling ratio based on the computing power score of each client. For clients with computing power scores below a preset threshold, their local datasets are stratified and sampled according to the data sampling ratio to form a sampled dataset for training. For clients with computing power scores above the preset threshold, their complete local datasets are used. The feature generation module is used to extract text feature vectors from text data in each client's sampled dataset using a pre-trained language model; and to extract global and local image features from image data in each client's sampled dataset using a convolutional neural network. A generative adversarial network training framework is adopted to align text feature vectors, global image features, and local image features into a unified semantic space, and to perform weighted fusion through a dynamic attention mechanism to generate multimodal features for each client. The local training module is used to distribute the global recognition model to each client. Each client uses its local sampled dataset and corresponding multimodal features to train the received global recognition model locally. The loss function of the local training includes a regularization term. The parameter value of the regularization term is dynamically selected by the reinforcement learning model based on the difference between the classification head weight of the current client's model and the classification head weight of the global recognition model. The global update module receives the local model parameters trained by each client, calculates the adaptive aggregation weight based on the data quality score of each client and the similarity between its local model and the global recognition model, and then performs weighted aggregation on the local model parameters uploaded by all clients to update the global recognition model. The anomaly detection module is used to construct a user heterogeneous graph based on the updated global detection model and multimodal features, dynamically optimize neighbor selection using reinforcement learning, and aggregate information through graph neural networks to identify abnormal users.

[0051] In some embodiments, the data acquisition module is specifically used to acquire telecommunications anomaly data stored locally by multiple clients, including text data, image data, and communication behavior data. For each client, collect the hardware performance metrics of its computing node. The hardware performance metrics include at least CPU performance, GPU memory capacity, and current remaining memory capacity. Based on hardware performance metrics, a corresponding computing power score for the client is generated through weighted calculation and normalization. The proportions of CPU performance, GPU memory capacity, and remaining memory capacity in the total memory are used as calculation items, each with a corresponding weight coefficient.

[0052] In some embodiments, the feature generation module is specifically used to encode the text data in the sampled dataset of each client using a pre-trained BERT model to obtain text feature vectors. For the image data in the sampled dataset of each client, the ResNet model is used to extract features to obtain the global image features of the image; at the same time, features are extracted from multiple preset local regions in the image to obtain multiple local image features. A generative adversarial network (GAN) consisting of a text encoder, an image encoder, and a discriminator is constructed. Text feature vectors, global image features, and multiple local image features are input into the corresponding encoders. The discriminator distinguishes the modality categories to which the features output by different encoders belong. Through adversarial training, the features of different modalities are forced to be mapped to a unified semantic space, generating unified text features and unified image features after semantic space alignment. Based on the unified text features and unified image features after semantic space alignment, the attention weights of the unified text features on the global image features and each local image feature are calculated through a dynamic attention mechanism. The global image features and local image features are then weighted and summed according to the attention weights to obtain the fused image feature representation. By concatenating unified text features with image feature representations, multimodal features corresponding to the client are generated.

[0053] In some embodiments, the feature generation module is specifically used to input each input image in the sampled dataset into a pre-trained ResNet model; By using the forward propagation of the ResNet model, the high-level semantic representation of the input image is extracted as the global image feature of the corresponding image; A pre-defined masking strategy is applied to the input image to generate multiple local image regions. The masking strategy is defined by a binary matrix of the same size as the image. Pixels with a value of 1 in the matrix constitute the local region to be extracted, while pixels with a value of 0 are non-masked regions. Each local image region defined by the mask is input into the ResNet model to extract the high-level semantic representation corresponding to each region, forming a set of local image features.

[0054] In some embodiments, the local training module is specifically used to distribute the global recognition model of the current round to all clients participating in the training, and each client uses the received global recognition model as the initial state of its local model. For each client, observe the current state based on the deployed reinforcement learning model. The current state is the classification head weight of the local model after the previous round of training for the corresponding client and the classification head weight of the global recognition model. The control reinforcement learning model outputs coefficients for the regularization term used in the loss function during the current local training process, based on the current state. Based on the coefficient values ​​selected by the reinforcement learning model, local training is performed. Its loss function is defined as the sum of cross-entropy loss and regularization loss. The regularization loss is determined by the coefficient values ​​and the difference between the global recognition model parameters and the local model parameters. During local training, the control client constructs training experience into experience tuples and stores them in the local experience pool. At the same time, based on the model classification head weights uploaded by each client, the similarity between clients is calculated, and each client is matched with a collaborative client with similar data distribution. This allows the client to sample experience tuples from the collaborative client's experience pool and update its own reinforcement learning model. Receive updated local model parameters uploaded by the client after local training is completed.

[0055] In some embodiments, the local training module is specifically used to define the input state space of the reinforcement learning model, which is a combined vector consisting of the local model classification head weights of the current client and the classification head weights of the global recognition model. Define the output action space of the reinforcement learning model, which is a set of continuous or discrete actions used to adjust the coefficient values ​​of the regularization term; The reinforcement learning model receives a state space vector, calculates and outputs specific actions through its policy network, and maps the specific actions to the coefficient values ​​of the regularization term used in the loss function during this local training process. After the client completes local training based on the selected coefficient values, the immediate reward is calculated based on the performance changes of the local model and the global recognition model on the validation set. The immediate reward is determined by at least the improvement in the accuracy of the local model, the improvement in the accuracy of the global recognition model, and the training round factor. By using immediate rewards and state transition information, the policy network parameters of the reinforcement learning model are updated through the policy gradient algorithm, thereby optimizing its decision-making ability to select the optimal coefficient values ​​for different states.

[0056] In some embodiments, the anomaly detection module is specifically used to construct a user heterogeneous graph with users in the system as nodes and the interaction relationships between users as edges. The feature vector of each node is the multimodal feature of the corresponding user extracted or encoded by the updated global detection model, and each edge is attached with the relationship feature of its corresponding interaction relationship type. For each node in the user heterogeneous graph, sort it according to its similarity with all neighbor nodes in the feature space to generate an initial candidate list of neighbors; Using a reinforcement learning model, the neighbor selection threshold is dynamically optimized for each relationship type. The state of the reinforcement learning model includes the current threshold, the average neighbor feature distance of nodes in the graph, and the node degree. Its actions are to increase or decrease the threshold. Its reward is based on the change in the neighbor aggregation effect before and after the threshold adjustment and the performance index of the auxiliary classification task. The optimized neighbor selection threshold is applied to filter the final set of neighbors for information aggregation from the initial candidate list. Based on the filtered neighbor set, message passing and feature aggregation are performed through a relation-specific graph convolutional network layer to update the embedding representation of each node in the user heterogeneous graph; Based on the updated node embedding representation, a classifier is used to predict whether the corresponding node is a normal user or an abnormal user, and the identification result is output.

[0057] This application also discloses a control device.

[0058] Specifically, the control device includes a memory and a processor, the memory storing a computer program that can be loaded and executed by the processor to perform the aforementioned multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction.

[0059] This application also discloses a computer-readable storage medium.

[0060] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the multi-dimensional risk assessment method for numbers based on dynamic feature extraction described above. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0061] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction, characterized in that, include: Acquire multi-source heterogeneous telecommunications anomaly data from clients, evaluate the computing power of each client, and generate a corresponding computing power score. The computing power score is calculated based on the client's CPU performance, GPU memory capacity, and remaining memory percentage. Based on each client's computing power score, the corresponding data sampling ratio is dynamically determined. For clients with computing power scores below a preset threshold, their local datasets are stratified and sampled according to the data sampling ratio to form a sampled dataset for training. For clients with computing power scores above the preset threshold, their complete local datasets are used. For text data in each client's sampled dataset, a pre-trained language model is used to extract text feature vectors; for image data in each client's sampled dataset, a convolutional neural network is used to extract global and local image features; a generative adversarial network training framework is adopted to align the text feature vectors, global image features, and local image features into a unified semantic space, and to perform weighted fusion through a dynamic attention mechanism to generate multimodal features for each client. A global recognition model is distributed to each client. Each client uses its local sampled dataset and the corresponding multimodal features to train the received global recognition model locally. The loss function of the local training includes a regularization term. The parameter value of the regularization term is dynamically selected by the reinforcement learning model based on the difference between the current client's model classification head weight and the global recognition model's classification head weight. The system receives the local model parameters trained by each client, calculates adaptive aggregation weights based on the data quality scores of each client and the similarity between their local models and the global recognition model, and then performs weighted aggregation on the local model parameters uploaded by all clients to update the global recognition model. Based on the updated global recognition model and the multimodal features, a user heterogeneous graph is constructed, and reinforcement learning is used to dynamically optimize neighbor selection. Information is aggregated through graph neural networks to identify abnormal users.

2. The method for multi-dimensional risk assessment of phone numbers based on dynamic feature extraction according to claim 1, characterized in that, The process of acquiring multi-source heterogeneous telecommunications anomaly data from clients, evaluating the computing power of each client, and generating a corresponding computing power score includes: The system retrieves telecommunications anomaly data stored locally by multiple clients. This telecommunications anomaly data includes text data, image data, and communication behavior data. For each client, collect the hardware performance metrics of its computing node. The hardware performance metrics include at least CPU performance, GPU memory capacity, and current remaining memory capacity. Based on hardware performance indicators, a corresponding computing power score for the client is generated through weighted calculation and normalization. Among them, CPU performance, GPU memory capacity, and the proportion of the remaining memory capacity in the total memory are used as calculation items, and each is configured with a corresponding weight coefficient.

3. The multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction according to claim 1, characterized in that, For text data in each client's sampled dataset, a pre-trained language model is used to extract text feature vectors; for image data in each client's sampled dataset, a convolutional neural network is used to extract global and local image features; a generative adversarial network training framework is employed to align the text feature vectors, global image features, and local image features into a unified semantic space, and a dynamic attention mechanism is used for weighted fusion to generate multimodal features for each client, including: For the text data in the sampled dataset of each client, a pre-trained BERT model is used to encode the text feature vector. For the image data in the sampled dataset of each client, the ResNet model is used to extract features to obtain the global image features of the image; at the same time, features are extracted from multiple preset local regions in the image to obtain multiple local image features. A generative adversarial network (GAN) is constructed, which includes a text encoder, an image encoder, and a discriminator. The text feature vector, the global image features, and multiple local image features are respectively input into the corresponding encoders. The discriminator distinguishes the modality categories to which the features output by different encoders belong. Through adversarial training, the features of different modalities are forced to be mapped to a unified semantic space, thereby generating unified text features and unified image features after semantic space alignment. Based on the unified text features and unified image features after semantic space alignment, the attention weights of the unified text features on the global image features and each of the local image features are calculated through a dynamic attention mechanism. The global image features and local image features are then weighted and summed according to the attention weights to obtain the fused image feature representation. The unified text features are concatenated with the image feature representation to generate the multimodal features corresponding to the client.

4. The multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction according to claim 3, characterized in that, For the image data in the sampled dataset of each client, a ResNet model is used for feature extraction to obtain the global image features; simultaneously, feature extraction is performed on multiple preset local regions in the image to obtain multiple local image features, including: For each input image in the sampled dataset, it is fed into the pre-trained ResNet model; The high-level semantic representation of the input image is extracted through the forward propagation of the ResNet model, which serves as the global image feature of the corresponding image. A preset masking strategy is applied to the input image to generate multiple local image regions. The masking strategy is defined by a binary matrix of the same size as the image. Pixels with a value of 1 in the matrix constitute the local region to be extracted, and pixels with a value of 0 are non-masked regions. Each local image region defined by the mask is input into the ResNet model to extract the high-level semantic representation corresponding to each region, forming a set of local image features.

5. The method for multi-dimensional risk assessment of phone numbers based on dynamic feature extraction according to claim 1, characterized in that, The step of distributing the global recognition model to each client, and each client using its local sampled dataset and the corresponding multimodal features to train the received global recognition model locally, includes: The global recognition model for the current round is distributed to all clients participating in the training, and each client uses the received global recognition model as the initial state of its local model. For each client, observe the current state based on the deployed reinforcement learning model. The current state is the local model classification head weight and the classification head weight of the global recognition model after the previous training round of the corresponding client. The control reinforcement learning model outputs coefficients for the regularization term used in the loss function during the current local training process, based on the current state. Based on the coefficient values ​​selected by the reinforcement learning model, local training is performed, and its loss function is defined as the sum of cross-entropy loss and regularization loss. The regularization loss is determined by the coefficient values ​​and the difference between the global recognition model parameters and the local model parameters. During local training, the control client constructs training experience into experience tuples and stores them in the local experience pool. At the same time, based on the model classification head weights uploaded by each client, the similarity between clients is calculated, and each client is matched with a collaborative client with similar data distribution. This allows the client to sample experience tuples from the collaborative client's experience pool and update its own reinforcement learning model. Receive updated local model parameters uploaded by the client after local training is completed.

6. The multi-dimensional risk assessment method for phone numbers based on dynamic feature extraction according to claim 5, characterized in that, The control reinforcement learning model, based on the current state, outputs coefficients for dynamically selecting the regularization term used in the loss function during the current local training process, including: Define the input state space of the reinforcement learning model, which is a combined vector consisting of the local model classification head weights of the current client and the classification head weights of the global recognition model; Define the output action space of the reinforcement learning model, which is a set of continuous or discrete actions used to adjust the coefficient values ​​of the regularization term; The reinforcement learning model receives a state space vector, calculates and outputs specific actions through its policy network, and maps the specific actions to the coefficient values ​​of the regularization term used in the loss function during this local training process. After the client completes local training based on the selected coefficient values, an immediate reward is calculated based on the performance changes of the local model and the global recognition model on the validation set. The immediate reward is determined by at least the improvement in the accuracy of the local model, the improvement in the accuracy of the global recognition model, and the training round factor. Using the immediate rewards and state transition information, the policy network parameters of the reinforcement learning model are updated through the policy gradient algorithm, thereby optimizing its decision-making ability to select the optimal coefficient values ​​for different states.

7. The method for multi-dimensional risk assessment of phone numbers based on dynamic feature extraction according to claim 1, characterized in that, The process involves constructing a user heterogeneous graph based on the updated global recognition model and the multimodal features, dynamically optimizing neighbor selection using reinforcement learning, and aggregating information through a graph neural network to identify abnormal users, including: Using users in the system as nodes and the interaction relationships between users as edges, a user heterogeneous graph is constructed. The feature vector of each node is the multimodal feature of the corresponding user extracted or encoded by the updated global recognition model, and each edge is attached with the relationship feature of its corresponding interaction relationship type. For each node in the user heterogeneous graph, sort it according to its similarity with all neighbor nodes in the feature space to generate an initial candidate list of neighbors; Using a reinforcement learning model, the neighbor selection threshold is dynamically optimized for each relationship type. The state of the reinforcement learning model includes the current threshold, the average neighbor feature distance of nodes in the graph, and the node degree. Its actions are to increase or decrease the threshold. Its reward is based on the change in the neighbor aggregation effect before and after the threshold adjustment and the performance index of the auxiliary classification task. The optimized neighbor selection threshold is applied to filter the final set of neighbors for information aggregation from the initial neighbor candidate list; Based on the filtered neighbor set, message passing and feature aggregation are performed through a relation-specific graph convolutional network layer to update the embedding representation of each node in the user heterogeneous graph; Based on the updated node embedding representation, a classifier is used to predict whether the corresponding node is a normal user or an abnormal user, and the identification result is output.

8. A multi-dimensional risk assessment device for phone numbers based on dynamic feature extraction, characterized in that, The device includes: The data acquisition module is used to acquire multi-source heterogeneous telecommunications anomaly data from clients, evaluate the computing power of each client, and generate a corresponding computing power score. The computing power score is calculated based on the client's CPU performance, GPU memory capacity, and remaining memory percentage. The data sampling module is used to dynamically determine the corresponding data sampling ratio based on the computing power score of each client. For clients with computing power scores below a preset threshold, their local datasets are stratified and sampled according to the data sampling ratio to form a sampled dataset for training. For clients with computing power scores above the preset threshold, their complete local datasets are used. The feature generation module is used to extract text feature vectors from text data in each client's sampled dataset using a pre-trained language model; and to extract global and local image features from image data in each client's sampled dataset using a convolutional neural network. A generative adversarial network training framework is adopted to align the text feature vectors, global image features, and local image features into a unified semantic space, and to perform weighted fusion through a dynamic attention mechanism to generate multimodal features for each client. The local training module is used to distribute the global recognition model to each client. Each client uses its local sampled dataset and the corresponding multimodal features to train the received global recognition model locally. The loss function of the local training includes a regularization term. The parameter value of the regularization term is dynamically selected by the reinforcement learning model based on the difference between the current client's model classification head weight and the global recognition model classification head weight. The global update module is used to receive the local model parameters trained by each client, calculate the adaptive aggregation weight based on the data quality score of each client and the similarity between its local model and the global recognition model, and perform weighted aggregation on the local model parameters uploaded by all clients accordingly to update the global recognition model. An anomaly detection module is used to construct a user heterogeneous graph based on the updated global detection model and the multimodal features, dynamically optimize neighbor selection using reinforcement learning, and aggregate information through a graph neural network to identify abnormal users.

9. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images