Law supervision model multi-level hybrid parallel optimization technology based on business process

Abstract

The invention discloses a legal supervision model multi-level hybrid parallel optimization technology based on a business process, and relates to the technical field of big data processing. In a traditional centralized model training mode, data of each judicial supervision node is required to be transmitted to a central server in a centralized manner for model training, so that not only is an information reconstruction risk caused, but also data transmission delay is caused, and the overall training efficiency is influenced. Constructing a big data law supervision model; a training data set is obtained and preprocessed; training a big data legal supervision model, segmenting network layers by using a GLoSGA algorithm, horizontally segmenting the big data legal supervision model according to layers or modules in the training process, and deploying the segmented network model on each judicial supervision node; and each judicial supervision node sends the gradient parameters subjected to privacy processing to a central server, and the central server summarizes the parameters of each judicial supervision node and then reversely updates the parameters to each judicial supervision node so as to update the network model of each node.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and in particular to a multi-level hybrid parallel optimization technology for legal supervision models based on business processes. Background Technology

[0002] Legal oversight, at its core, involves the rigorous monitoring and management of every detail and procedure in the implementation of the law. With the rapid development of information technology, especially the widespread application of big data technology, massive and complex data resources have been accumulated in the field of legal oversight. This data includes detailed records of every stage from the acceptance of a legal case to its conclusion, covering information from the personal information of the parties involved to the trial process and judgment results. Therefore, big data legal oversight technology is being widely applied in various fields such as judicial supervision, risk assessment, evidence analysis, and public opinion monitoring.

[0003] Some models used in big data legal supervision, such as traditional machine learning and deep learning models, suffer from lengthy training processes. Furthermore, traditional centralized model training methods present a series of problems, including centralized data transmission, leakage of sensitive information, and computational resource bottlenecks. Specifically, traditional centralized model training requires the centralized transmission of data from various judicial supervision nodes to a central server for model training. However, judicial data often contains sensitive information, such as case details and personal identification information. Traditional centralized training, which requires data aggregation, is prone to violating privacy regulations such as GDPR. This not only raises the risk of information reconstruction but also leads to data transmission delays, thus affecting overall training efficiency. Summary of the Invention

[0004] In view of this, the present invention provides a multi-level hybrid parallel optimization technology for legal supervision models based on business processes. It can use a parallel training strategy that combines data parallelism and model parallelism to train a deep learning model for legal supervision of big data, thereby improving the model's generalization ability and significantly increasing the model's training speed.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A multi-level hybrid parallel optimization technique for legal supervision models based on business processes includes the following optimization steps:

[0007] S1, Constructing a big data legal supervision model;

[0008] S2, Obtain the training dataset: This involves obtaining the raw data from each judicial supervision node locally. As training datasets for their respective network models;

[0009] S3 preprocesses the training dataset to obtain an enhanced training dataset. ;

[0010] S4, Training the Big Data Legal Supervision Model: The GLoSGA algorithm is used to segment the network layers. During the training process, the big data legal supervision model is horizontally segmented according to layers or modules, and each judicial supervision node completes its local training task. The segmented network model is then deployed on each judicial supervision node.

[0011] S5, each judicial oversight node utilizes an enhanced training dataset. The network model deployed is trained simultaneously; when the network model is updated, the gradient parameters of the network model of each judicial supervision node are pruned and noise is injected.

[0012] S6, each judicial supervision node sends the privacy-processed gradient parameters to the central server. The central server then aggregates the parameters from each judicial supervision node and updates them back to each judicial supervision node to update the network model of each judicial supervision node.

[0013] Furthermore, the big data legal supervision model in S1 is obtained by introducing a network module for extracting minute features of the data, based on CNN, VGG, or MLP models.

[0014] Furthermore, the network module is one or more combinations of depthwise separable convolution, SE, residual, or attention mechanisms.

[0015] Furthermore, the method for preprocessing the training dataset in S3 is as follows:

[0016] S31, each judicial supervision node's local raw data Perform one or more data augmentation operations, including scaling, random affine transformation, random perspective transformation, grayscale conversion, Gaussian blur, random horizontal flip, or random cropping, to obtain the augmented data. The augmented dataset is then normalized to the mean and then standardized.

[0017] Furthermore, in S5, the specific steps for pruning and injecting noise into the gradient parameters of the network model for each judicial supervision node are as follows:

[0018] The network model performs gradient calculations during forward propagation at each node to obtain the gradient parameters:

[0019]

[0020] Where g is the gradient, Let θ be the gradient operator, θ be the model parameters, L be the loss function, and D be the training dataset;

[0021] Gradient norm clipping of gradient parameters:

[0022]

[0023] in, This is the clipping threshold;

[0024] Noise injection is performed on the clipped gradient parameters:

[0025]

[0026] ;

[0027] in, The gradient after injecting noise is used to update the model parameters; This is a noise vector, used to inject into the gradient to increase the randomness of the training process; Let the noise distribution be a function with a mean of 0 and a covariance matrix of . The multivariate Gaussian distribution;

[0028] Set a privacy budget and update it in stages:

[0029]

[0030]

[0031] in, δ represents the upper limit of privacy loss, indicating the strength of privacy protection; n is the upper limit of the probability of privacy protection failure; ← represents the learning rate; ← indicates the update operation, which assigns the value on the right to the variable on the left.

[0032] The beneficial effects of this invention compared to the prior art are:

[0033] 1. This invention utilizes the distributed computing resources of various judicial supervision nodes. Each judicial supervision node has its own network model, and each node trains its own network model using local data. This allows each judicial supervision node to complete data preprocessing and model training locally, avoiding the transmission of sensitive data to the central server, thereby significantly reducing the risk of data leakage and meeting the strict data confidentiality requirements of the judicial field. Simultaneously, a parallel training strategy combining data parallelism and model parallelism is used to train the big data legal supervision deep learning model. That is, judicial institutions jointly build a global model without sharing data, improving the model's generalization ability and significantly increasing the model training speed. After each model update during training, each judicial supervision node injects noise into the updated model to ensure that the parameters that need to be updated are not leaked when transmitted to the central server. The central server aggregates the parameters from each judicial supervision node and updates them back to each judicial supervision node, preventing the derivation of the original data content through parameter updates. In other words, the network models of each judicial supervision node, through localized data training and parameter exchange, can avoid the cross-domain transmission of original data, meeting the strict privacy protection requirements of the judicial field. Furthermore, it can further improve the security and consistency of training results among nodes in the case of heterogeneous data distribution.

[0034] 2. The big data legal supervision model of the present invention includes improvements in depthwise separable convolution, residual connections, and SE attention mechanism, which effectively enhance the model's ability to capture key features and improve the overall accuracy and robustness of the supervision task.

[0035] 3. The dynamic optimization strategy and historical data recording mechanism on the server side not only provide reliable data support for system traceability, but also provide detailed basis for subsequent model optimization and judicial evidence analysis, further improving the credibility and maintainability of the entire system.

[0036] 4. This invention provides a comprehensive solution to the problems of low training efficiency, data security risks and insufficient model accuracy in the training process of big data legal supervision machine learning models, and can be widely applied to multiple fields such as judicial supervision, risk assessment, evidence analysis and public opinion monitoring. Attached Figure Description

[0037] The accompanying drawings, which form part of this application, are provided to further illustrate the invention.

[0038] Figure 1 A flowchart for constructing a big data legal supervision model.

[0039] Figure 2 A flowchart illustrating the process of augmenting training data.

[0040] Figure 3This diagram illustrates a hybrid parallel training strategy that combines data parallelism and model parallelism.

[0041] Figure 4 This is a schematic diagram illustrating the application of pruning and injecting noise into the gradient parameters of the network model for each judicial supervision node.

[0042] Figure 5 This is a schematic diagram of the network model of each judicial supervision node and the global model of the central server.

[0043] Figure 6 This is a diagram comparing the big data legal supervision model before and after optimization. Detailed Implementation

[0044] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] See Figures 1 to 6 This embodiment proposes a multi-level hybrid parallel optimization technique for a legal supervision model based on business processes. This method utilizes the distributed computing resources of various judicial supervision nodes (including courts, procuratorates, public security organs, and other judicial departments). Each judicial supervision node has its own network model, and each node trains its own network model using local data. A parallel training strategy combining data parallelism and model parallelism is used to train the big data legal supervision deep learning model. In other words, judicial institutions jointly build a global model without sharing data, improving the model's generalization ability and significantly increasing training speed. After each model update during training, each judicial supervision node injects noise into the updated model to ensure that the parameters to be updated are not leaked when transmitted to the central server. The central server aggregates the parameters from each judicial supervision node and updates them back to each judicial supervision node, preventing the derivation of the original data content through parameter updates. In other words, by training the network models of each judicial supervision node with localized data and exchanging parameters, the cross-domain transmission of original data can be avoided, meeting the strict privacy protection requirements of the judicial field. Furthermore, it can further improve the security and consistency of training results among nodes in the case of heterogeneous data distribution.

[0046] Specifically, the optimization steps include the following:

[0047] S1. Constructing a Big Data Legal Supervision Model: Based on CNN, VGG, or MLP models, a network module for extracting subtle features from the data is introduced to obtain the big data legal supervision model. This network module enhances the model's ability to capture key information, further improving the overall accuracy and robustness of the supervision task. The network module is one or more combinations of depthwise separable convolution, SE, residual, or attention mechanisms.

[0048] S2, Obtain the training dataset: This involves obtaining the raw data from each judicial supervision node locally. As training datasets for their respective network models, these judicial departments store their local raw data independently with significant differences in format, so the judicial supervision nodes do not share data.

[0049] S3, Preprocessing the training dataset:

[0050] S31, each judicial supervision node's local raw data Scaling Random affine transformation Random perspective transformation Grayscale conversion Gaussian blur Random horizontal flip Alternatively, one or more data augmentation operations can be performed during random pruning to obtain augmented data. This embodiment utilizes data augmentation techniques or the data processing methods proprietary to the judicial supervision node to process the training dataset, thereby increasing the diversity of the dataset and improving the model's generalization ability and robustness, while also preventing attacks or leaks during data transmission to the central server.

[0051] S32, perform mean normalization on the enhanced dataset to ensure all data are on the same order of magnitude:

[0052]

[0053] in, , Representing data The maximum and minimum values ​​in;

[0054] S33, standardize the data to ensure data quality and model generalization ability;

[0055]

[0056] in, This indicates the data after mean normalization. , ϵ represents the mean and variance; γ is a constant; β is a scaling parameter; and β is an offset parameter.

[0057] S4, Training the Big Data Legal Supervision Model: Utilizing the distributed computing resources of each judicial supervision node, a parallel training strategy combining data parallelism and model parallelism is employed to train the big data legal supervision model, thereby significantly improving the model training speed. The training steps are as follows:

[0058] S41 uses the GLoSGA algorithm to divide the network layers. During the training process, the big data legal supervision model is horizontally divided into layers or modules, and each judicial supervision node completes its local training task.

[0059]

[0060] S42 deploys the segmented network model at each judicial supervision node.

[0061] S5, each judicial oversight node utilizes an enhanced training dataset. The network models deployed are trained simultaneously to achieve parallel optimization of the models at each judicial supervision node. When the network model is updated, the gradient parameters of the network models at each judicial supervision node are pruned and noise is injected to ensure that the parameters that need to be updated are not leaked when transmitted to the central server. The specific operation steps are as follows:

[0062] The network model performs gradient calculations during forward propagation at each node to obtain the gradient parameters:

[0063]

[0064] Where g is the gradient, Let θ be the gradient operator, θ be the model parameters, L be the loss function, and D be the training dataset;

[0065] Gradient norm clipping of gradient parameters:

[0066]

[0067] in, This is the clipping threshold;

[0068] Noise injection is performed on the clipped gradient parameters:

[0069]

[0070] ;

[0071] in, The gradient after injecting noise is used to update the model parameters; This is a noise vector, used to inject into the gradient to increase the randomness of the training process; Let the noise distribution be a function with a mean of 0 and a covariance matrix of . The multivariate Gaussian distribution;

[0072] Set a privacy budget and update it in stages:

[0073]

[0074]

[0075] in, δ represents the upper limit of privacy loss, indicating the strength of privacy protection; n is the upper limit of the probability of privacy protection failure; ← represents the learning rate; ← indicates the update operation, which assigns the value on the right to the variable on the left.

[0076] In step S6, each judicial oversight node sends its privacy-processed gradient parameters to the central server. The central server then aggregates the parameters from all judicial oversight nodes and updates them in reverse, updating the network models of each node. This achieves the goal of parameter exchange among the judicial oversight nodes. This design addresses the limited data volume and significant regional characteristics of a single judicial institution. By aggregating data features from multiple sources and updating them in reverse at each judicial oversight node through the central server, the small sample size problem is mitigated. Furthermore, by customizing the models of each judicial oversight node, the system adapts to the specificities of different judicial jurisdictions. This prevents the derivation of original data content through parameter updates and further enhances the security and consistency of training results among nodes in heterogeneous data distribution scenarios. The server-side dynamic optimization strategy and historical data recording mechanism not only provide reliable data support for system traceability but also provide detailed evidence for subsequent model optimization and judicial evidence analysis, further improving the credibility and maintainability of the entire system.

[0077] The big data legal supervision model in this embodiment can be used for micro-expression recognition tasks, trace recognition tasks, or evidence and witness identification tasks.

[0078] Example 1:

[0079] The following example uses the application of computer vision tasks related to legal supervision, specifically "micro-expression recognition," to illustrate the technical solution of this invention in detail. The model selected in this embodiment is a CNN model.

[0080] The multi-level hybrid parallel optimization technique for the business process-based legal supervision model in this embodiment specifically includes the following optimization steps:

[0081] S1, Constructing a Big Data Legal Supervision Model: This involves improving the CNN deep learning network structure used in the computer vision task for legal supervision, introducing a network module for extracting minute features from the data, and obtaining a big data legal supervision model. Figure 1 The improved structure diagram for building a big data legal supervision model is obtained through the following process;

[0082] S11, embedding depthwise separable convolutions in deep learning network models;

[0083] S12, embedding residual modules in deep learning network models;

[0084] S13 embeds the SE structure into a deep learning network model, ultimately resulting in an improved big data legal supervision model.

[0085] S2, Obtain the training dataset: This involves obtaining the raw data from each judicial supervision node locally. As training datasets for their respective network models;

[0086] S3, Preprocessing the training dataset: Figure 2 This is a schematic diagram of the data augmentation process in this embodiment, illustrating the local augmentation of the original data at each node. The process for data augmentation is as follows:

[0087] S31, for local raw data Scaling ;

[0088] S32, for local raw data Perform random affine transformation ;

[0089] S33, for local raw data Perform random perspective transformation ;

[0090] S34, for local raw data Perform grayscale conversion ;

[0091] S35, for local raw data Gaussian blur ;

[0092] S36, for local raw data Perform random horizontal flipping Thus, enhanced data is obtained. ;

[0093] S37, Perform mean normalization on the enhanced dataset to ensure all data are on the same order of magnitude:

[0094]

[0095] in, , Representing data The maximum and minimum values ​​in;

[0096] S38, standardize the data to ensure data quality and model generalization ability;

[0097]

[0098] in, This indicates the data after mean normalization. , ϵ represents the mean and variance; γ is a constant; β is a scaling parameter; and β is an offset parameter.

[0099] S4, Training the Big Data Legal Supervision Model: The improved deep learning network model is sliced ​​using the GLoSGA algorithm, which facilitates the distributed deployment of the model to various judicial supervision nodes.

[0100] S5 involves deploying the sliced ​​deep learning network model to each judicial supervision node. For details, see [link to method]. Figure 3 As shown.

[0101] S6. After deployment, the model training task is started. After the model training is completed on each judicial supervision node, the gradient parameters are pruned and noise is injected to prevent the gradient from being maliciously restored to the original data during the backpropagation process. After the operation is completed, the parameters are uniformly aggregated to the central server, processed by the central server, and then returned to each judicial supervision node for the backpropagation process. The specific steps are as follows:

[0102] S61, the model performs forward propagation to calculate gradients and obtain gradient parameters;

[0103] S62, perform gradient norm clipping on the gradient parameters;

[0104] S63, Noise injection is performed on the gradient parameters. The noise injection process is as follows: Figure 4 As shown;

[0105] S64 sends the privacy-processed gradient parameters to the central server;

[0106] In S65, after receiving the parameters, the central server summarizes them and then propagates them back to each judicial node to update the network model. Figure 5 This is a schematic diagram of the model after all steps have been completed in an embodiment of the present invention.

[0107] To verify the model's effectiveness in this embodiment, tests were conducted on various parameters of the model before and after optimization. Specific results are as follows: Figure 6 As shown, the optimized model exhibits significant improvements in data security, training speed, model accuracy, computational resource utilization, and privacy protection. The trained big data legal supervision model obtained in this embodiment is primarily used for micro-expression recognition, trace evidence recognition, or witness identification tasks.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions created by the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions created by the present invention without departing from the essence and scope of the technical solutions created by the present invention.

Claims

1. A multi-level hybrid parallel optimization technique for a legal supervision model based on business processes, characterized in that: The optimization steps include the following: S1, Constructing a big data legal supervision model; S2, Obtain the training dataset: This involves obtaining the raw data from each judicial supervision node locally. As training datasets for their respective network models; S3 preprocesses the training dataset to obtain an enhanced training dataset. ; S4, Training the Big Data Legal Supervision Model: The GLoSGA algorithm is used to segment the network layers. During the training process, the big data legal supervision model is horizontally segmented according to layers or modules, and each judicial supervision node completes its local training task. The segmented network model is then deployed on each judicial supervision node. S5, each judicial oversight node utilizes an enhanced training dataset. The network model deployed is trained simultaneously; when the network model is updated, the gradient parameters of the network model of each judicial supervision node are pruned and noise is injected. S6, each judicial supervision node sends the privacy-processed gradient parameters to the central server. The central server then aggregates the parameters from each judicial supervision node and updates them back to each judicial supervision node to update the network model of each judicial supervision node.

2. The multi-level hybrid parallel optimization technology for the business process-based legal supervision model according to claim 1, characterized in that, The big data legal supervision model in S1 is obtained by introducing a network module for extracting minute features of the data, based on CNN, VGG, or MLP models.

3. The multi-level hybrid parallel optimization technology for the business process-based legal supervision model according to claim 2, characterized in that, The network module is one or more of depthwise separable convolution, SE, residual, or attention mechanisms.

4. The multi-level hybrid parallel optimization technology for the business process-based legal supervision model according to claim 1, characterized in that, The method for preprocessing the training dataset in S3 is as follows: Each judicial supervision node processes its local raw data... Perform one or more data augmentation operations, including scaling, random affine transformation, random perspective transformation, grayscale conversion, Gaussian blur, random horizontal flip, or random cropping, to obtain the augmented data. The augmented dataset is then normalized to the mean and then standardized.

5. The multi-level hybrid parallel optimization technology for the business process-based legal supervision model according to claim 1, characterized in that, In step S5, the specific steps for pruning the gradient parameters and injecting noise into the network model of each judicial supervision node are as follows: The network model performs gradient calculations during forward propagation at each node to obtain the gradient parameters: Where g is the gradient, Let θ be the gradient operator, θ be the model parameters, L be the loss function, and D be the training dataset; Gradient norm clipping of gradient parameters: in, This is the clipping threshold; Noise injection is performed on the clipped gradient parameters: ； in, The gradient after injecting noise is used to update the model parameters; This is a noise vector, used to inject into the gradient to increase the randomness of the training process; Let the noise distribution be a function with a mean of 0 and a covariance matrix of . The multivariate Gaussian distribution; Set a privacy budget and update it in stages: in, δ represents the upper limit of privacy loss, indicating the strength of privacy protection; n is the upper limit of the probability of privacy protection failure; ← represents the learning rate; ← indicates the update operation, which assigns the value on the right to the variable on the left.

Images