A code hosting platform risk control system and method based on deep learning

Abstract

A code hosting platform risk control system and method based on deep learning, comprising a risk control modeling platform and a risk control operation platform, the risk control modeling platform and the risk control operation platform are built on a cloud infrastructure, and the risk control of the whole life cycle of the code is realized based on the code hosting platform and the user center. The cloud infrastructure aggregates a large amount of computing, storage and network resources, and provides machine learning, deep learning, big data and other service capabilities; the code hosting platform provides code hosting, code review, progress management, quality management, problem management, continuous integration and other functions, realizes code hosting and collaborative development and continuous release; the user center stores the basic information of code visitors, submitters and collaborators, and combines the identity information to carry out user portrait and form an identity tag. The application can effectively identify the security risk of the source code, timely carry out risk early warning, continuously improve the model, improve the risk control accuracy and enhance the ability to discover unknown risks.

Description

Technical Field

[0001] The present invention belongs to the fields of information security, machine learning and deep learning technologies, and specifically relates to a code hosting platform risk control system and method based on deep learning. Background Art

[0002] With the rapid development of deep learning technology and the support of massive data and efficient computing power in the era of the Internet and cloud computing, deep learning technology represented by CNN convolutional neural networks has constructed large-scale neural networks similar to the structure of the human brain through training, and has made breakthrough progress in computer vision, speech recognition, natural language understanding and other fields, bringing disruptive changes to the entire society.

[0003] In recent years, businesses have accelerated their digital transformation, posing new challenges for the software industry. IT companies are adopting new development methods such as agile and DevOps to rapidly respond to evolving business needs. New-generation code collaboration platforms, such as GitHub, have become a mainstream collaborative development method, and code hosting platforms have become essential infrastructure for software development. Frequent security incidents such as code leaks, malicious code modifications, malicious code deletions, and vulnerability injections pose significant risks to businesses. The importance of source code is self-evident, placing higher demands on code hosting platforms. Furthermore, with the rapid development of next-generation information technology, the value of open source has become increasingly prominent, necessitating the security of the open source supply chain for the open source code repositories relied upon. In this context, how to effectively leverage deep learning technologies, combined with existing expert knowledge, to implement risk control in the code collaboration process and ensure source code security through code hosting platforms has become a pressing issue. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned existing technologies, the purpose of the present invention is to provide a code hosting platform risk control system and method based on deep learning. By utilizing deep learning technology and based on the log data recorded during the code collaboration process on the code hosting platform, a neural network model is designed. Various factors affecting code security are comprehensively considered, and the deep connections between the behaviors of code submitters are explored. At the same time, combined with existing risk control methods, a targeted risk control model is formed to effectively identify source code security risks, provide timely risk warnings, and continuously improve the model to enhance risk control accuracy and the ability to discover unknown risks.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A deep learning-based risk control system for a code hosting platform, including a risk control modeling platform and a risk control operations platform. These platforms are built on cloud infrastructure and implement risk control throughout the code lifecycle based on the code hosting platform and user center.

[0007] The user center includes basic information of users including code visitors, submitters, and collaborators, and combines their identity information to create user profiles and form identity tags;

[0008] The risk control modeling platform runs a rule-based risk control model, a random forest risk control model, a deep learning risk control model, a data fusion risk control model, and a dynamic optimization model;

[0009] The rule-based risk control model is used to extract the data required by the rules from the code hosting platform and the user center data to perform risk assessment;

[0010] The random forest risk control model is used to extract set data variables from the code hosting platform and the user center data to perform risk assessment;

[0011] The deep learning risk control model is used to directly perform risk assessment using the raw data of the code hosting platform and the user center as input;

[0012] The data fusion risk control model is used to comprehensively judge the data of the code hosting platform and the user center to make risk assessments;

[0013] The dynamic optimization model is used to timely adjust model parameters for the code hosting platform and the data of the user center;

[0014] The risk control operation platform performs risk control determination based on the rule determination risk control model, random forest risk control model, deep learning risk control model, data fusion risk control model and dynamic optimization model;

[0015] The cloud infrastructure aggregates a large amount of computing, storage, and network resources, and provides service capabilities such as machine learning, deep learning, and big data; the code hosting platform runs on the cloud infrastructure, providing functions such as code hosting, code review, progress management, quality management, problem management, and continuous integration, realizing code hosting, collaborative development, and continuous release.

[0016] The risk control modeling platform is mainly used for risk control model modeling, and includes three sub-platforms: model platform, rule platform and process platform;

[0017] The rule platform is responsible for generating rules and providing preset rules based on domain experts;

[0018] The process platform is responsible for the creation and setting of risk control detection processes;

[0019] The model platform is based on the general machine learning, deep learning training and reasoning services provided by the cloud infrastructure, and combines the code hosting business to form rule judgment risk control models, random forest risk control models, deep learning risk control models, data fusion risk control models and dynamic optimization models. At the same time, it provides intelligent recommendations of relevant rule thresholds, selection of judgment rules and intelligent combination recommendations for the rule platform, and provides process combination recommendation services for the process platform.

[0020] The rule-based risk control model is a risk control model formed based on the rule platform and the process platform; the deep learning risk control model is an end-to-end model based on the original operation log behavior data provided by the code hosting platform, and adopts a CNN neural network model with a self-attention mechanism to process the time-sequential behavior data and identify risks in combination with the personal identity information of the code visitor or submitter; the core of the data fusion risk control model is a neural network model, which performs comprehensive analysis and processing based on the judgment results of the rule-based risk control model, the random forest risk control model and the deep learning risk control model, and combines the user identity information and the attributes of the code project to obtain the final risk judgment result.

[0021] The dynamic optimization model adopts a neural network model, which timely adjusts the model parameters of the rule-based risk control model, the random forest risk control model, the deep learning risk control model, and the data fusion risk control model based on the current input. It is composed of corresponding model sub-modules, and timely adjusts the risk control model to improve the response speed of risk anomalies;

[0022] The risk control operation model mainly includes a rule process risk control engine, a random forest risk control engine, a deep learning risk control engine, a fusion judgment module, a dynamic optimization module, a data processing module and an early warning module. The rule process risk control engine extracts the data required by the rules from the code hosting platform and the user center based on the rule judgment risk control model, and performs risk judgment. The random forest risk control engine extracts the set data variables from the data from the code hosting platform and the user center based on the random forest risk control model, and performs risk judgment. The deep learning risk control engine directly performs risk judgment on the original data from the code hosting platform and the user center as input based on the deep learning risk control model. The fusion judgment module performs risk judgment on the data from the code hosting platform and the user center based on the data fusion risk control model and the judgment results of the rule process risk control engine, the random forest risk control engine and the deep learning risk control engine.

[0023] The dynamic optimization module is based on the dynamic optimization model and the data from the code hosting platform and the user center, and timely adjusts and sets the model parameters of the rule judgment risk control engine, the random forest risk control engine, the deep learning risk control engine and the data fusion risk control module; the data processing module is responsible for collecting the data required for cleaning the risk control platform and preparing the data according to the requirements of the model, while responding to the relevant parameter setting changes from the dynamic optimization engine in real time and providing the required data; the early warning module responds to the risk judgment results in a timely manner, records logs, and notifies relevant management personnel or directly performs operations with limited access rights.

[0024] A method for a code hosting platform risk control system based on deep learning, used for model training of a risk control model, comprising the following steps:

[0025] Step 101: Collect massive amounts of code hosting platform log data and user center data, prepare training data, set risk levels, for example, four risk levels: D / Level 1 (major risk), C / Level 2 (high risk), B / Level 3 (average risk), and A / Level 4 (low risk), and label the risk data based on domain expert knowledge.

[0026] Step 102: Based on the labeled data, a gradient descent optimization algorithm is used to train the process recommendation service model, the threshold recommendation service model, and the rule selection service recommendation model to obtain the recommendation service model required for risk control modeling.

[0027] The gradient descent optimization algorithm first sets the objective function, then performs forward calculations on the training samples, and then calculates the error backpropagation until the function converges.

[0028] Step 103: Combine the rule processes and train the risk control model based on the labeled data.

[0029] Step 104: Based on expert knowledge, attributes influencing risk assessment are pre-set, including code project attributes (project number, project name, project domain, project security level, number of project versions, latest project release date, total number of project issues, number of latest project version issues, number of project submitters, latest submission date, number of submissions in the last day, number of submissions in the last month, number of visitors in the last day, number of visitors in the last month, etc.), visitor attributes (visitor ID, visitor identity, visitor IP address, content of this visit, content of this submission, last three login IP addresses, last three visited projects, content of last three visits, projects to which last three submissions belong, content of last three submissions, etc.), and the random forest risk control model is constructed using the random forest algorithm based on the labeled data.

[0030] The random forest algorithm is specifically:

[0031] Traverse the random forest K times: Sampling with replacement from the training set T, sample N times to form a new sub-training set D, randomly select m features, where m < M, use the new training set D and m features to learn a complete decision tree to obtain a random forest.

[0032] Step 105: Design a network model and train the deep learning risk control model using the gradient descent method based on the labeled data;

[0033] Step 106: Fix the model parameters of the discrimination rule risk control model, the random forest risk control model, and the deep learning risk control model, and train the data fusion model using the gradient descent method based on the labeled data;

[0034] Step 107: Fix the model parameters of the discrimination rule risk control model and the random forest risk control model, and slightly adjust the network parameters of the deep learning risk control model and the data fusion model;

[0035] The micro-adjustment still uses the optimization algorithm and back propagation. Micro-adjustment mainly refers to small changes in the model parameters. You can refer to the BERT pre-trained language model. The model will be adjusted according to actual business needs, but the specific adjustment method is given differently from the training time.

[0036] Step 108: Fix the model parameters of the discrimination rule risk control model and train the corresponding submodule of the dynamic optimization model based on the labeled data;

[0037] Step 109: Fix the model parameters of the random forest risk control model and train the corresponding submodule of the dynamic optimization model based on the labeled data;

[0038] Step 110: Fix the model parameters of the deep learning risk control model and train the corresponding submodules of the dynamic optimization model based on the labeled data;

[0039] Step 111 , fixing the model parameters of the data fusion model, and training the corresponding submodules of the dynamic optimization model based on the labeled data, can achieve dynamic optimization of the four models in steps 108 - 111 .

[0040] A code hosting platform risk control method based on deep learning, used for risk control of code hosting platforms, comprising the following steps:

[0041] Step 201: Using the risk control modeling platform, generate an initial risk control model through steps 101-111, including the discrimination rule risk control model, the random forest risk control model, the deep learning risk control model, the data fusion model, and the dynamic optimization model;

[0042] Step 202: Load the initial risk control model into the static model memory to form a service, python: Model new_model = load_model('risk control model.nn'), in the engine or module corresponding to the risk control operation platform;

[0043] Step 203: Continuously collect data from the code hosting platform and the user center, including real-time data, log data, user data, user profiles, and other information. The data processing module of the risk control operation platform extracts, cleans, filters, and processes the data to meet the data requirements of the rule-based risk control engine, the random forest risk control engine, the deep learning risk control engine, the fusion judgment module, and the dynamic optimization module.

[0044] Step 204: The rule-based risk control engine determines the risks based on the set processes and rules (such as IP blacklisted, permission errors, etc.), and submits the results to the fusion determination module.

[0045] Step 205: The random forest risk control engine uses a random forest algorithm to determine the risk and submits the result to the fusion determination module;

[0046] The determination is based on the following attributes, and a random forest algorithm is executed to perform classification;

[0047] Code project category (project number, project name, project field, project security level, number of project versions, latest project release date, total number of project issues, number of latest project version issues, number of project committers, latest commit date, number of commits in the last day, number of commits in the last month, number of visitors in the last day, number of visitors in the last month, etc.), visitor category (visitor ID, visitor identity, visitor IP, content of this visit, content of this submission, last three login IP addresses, last three visited projects, content of last three visits, projects to which last three submissions belonged, content of last three submissions, etc.) and other attributes.

[0048] Step 206: The deep learning risk control engine uses the raw data from the code hosting platform and the user center as input to directly perform risk assessment (according to the model's input requirements, the raw data is directly entered into the model), and submits the results to the fusion assessment module.

[0049] Step 207: The fusion determination module integrates the determination results of the rule-based risk control engine, the random forest risk control engine, and the deep learning risk control engine to perform risk determination and notify the early warning module of the result.

[0050] Step 208: The early warning module promptly responds to the risk assessment result (recording a log, issuing an alarm, or directly restricting user operations, notifying an administrator based on the risk level), records the log, and notifies relevant administrators or directly restricts access rights.

[0051] Step 209: The dynamic optimization module optimizes and adjusts the model parameters and input data items of the rule-based risk control engine, the random forest risk control engine, the deep learning risk control engine, and the data fusion risk control module based on the project to which the code belongs and the collected data (the optimization model dynamically adjusts the parameters and input items of the other four models based on different input data, such as the information of the logged-in person, the operation, and the content accessed) to accurately respond to new risk changes.

[0052] Step 210: Continue to collect log data and result feedback from the code hosting platform and risk control platform, continuously optimize and improve the model, improve risk control accuracy, and enhance the ability to discover unknown risks.

[0053] Beneficial effects of the present invention:

[0054] The present invention provides a code hosting platform risk control method based on deep learning, which fully considers the main factors affecting code security, forms a risk control modeling and operation platform based on the existing cloud infrastructure, combines traditional process rule decision-making and machine learning risk control management methods, utilizes deep learning technology, and designs a neural network model based on the log data of the code collaboration process recorded by the code hosting platform to explore the behavior patterns of code submitters and visitors, find the deep connection between various elements of code collaboration risk control, more effectively identify source code security risks, and better deal with risk challenges such as code leakage, malicious code modification, malicious code deletion, and vulnerability implantation. Compared with the traditional rule-based or machine learning approach, the use of deep learning technology can better discover potential risks. At the same time, it takes into account the identity data of users participating in code collaboration from the user center, adds external user portraits to the risk control system, and forms a personalized risk control model, which improves the accuracy of risk control and the ability to warn and respond to risks. In addition, it adopts a dynamic optimization strategy to adjust the risk control model in a timely manner, and integrates traditional process rule decision-making and machine learning risk control management methods. At the same time, it makes intelligent recommendations, improves the dynamic response speed of risk anomalies, enhances the ability to discover unknown risks, and continuously improves the model to further improve the accuracy and adaptability of risk identification and response, meeting the needs of personalized development of code collaboration platform projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a schematic diagram of the code hosting platform risk control system.

[0056] Figure 2 This is a schematic diagram of the risk control model of the code hosting platform. DETAILED DESCRIPTION

[0057] The present invention will be described in further detail below with reference to the accompanying drawings.

[0058] In one embodiment, Figure 1 、 2 As shown, a risk control modeling and operation platform is formed based on the existing cloud infrastructure, making full use of the massive log data recorded in the code hosting platform during the code collaboration process, while considering the identity data of users participating in the code collaboration from the user center, combining traditional process rule decision-making and machine learning risk control management methods, and using deep learning technology to design a neural network risk control model to explore the deep connection between the various elements of code collaboration risk control, and make intelligent recommendations for traditional risk control methods to form a personalized risk control model, effectively identify source code security risks, and provide timely risk warnings. In addition, a dynamic optimization strategy is used to adjust the risk control model in a timely manner, improve the risk anomaly response speed, enhance the ability to discover unknown risks, and continuously improve the model to improve risk control accuracy. Among them,

[0059] The cloud infrastructure gathers a large amount of computing, storage, and network resources, and provides service capabilities such as machine learning, deep learning, and big data; the code hosting platform runs on the cloud infrastructure, providing functions such as code hosting, code review, progress management, quality management, problem management, and continuous integration, realizing code hosting and collaborative development and continuous release; the user center stores basic information of users such as code visitors, submitters, and collaborators, and uses their identity information to create user profiles and form identity tags; the risk control modeling platform is mainly responsible for risk control model modeling, including three main sub-platforms: model platform, rule platform, and process platform; the rule platform is responsible for generating rules and providing preset rules based on domain experts. Then; the process platform is responsible for the creation and setting of risk control detection processes; the model platform is based on the general machine learning, deep learning training and reasoning services provided by the cloud infrastructure, combined with the code hosting business to form rule-based risk control models, random forest risk control models, deep learning risk control models, data fusion risk control models and dynamic optimization models, while providing the rule platform with intelligent recommendations of relevant rule thresholds, selection of judgment rules and intelligent combination recommendations, and providing process combination recommendation services for the process platform; the rule-based risk control model is a risk control model formed based on the rule platform and the process platform; the deep learning risk control model is based on the original operation log behavior data provided by the code hosting platform for end-to-end The core of the modeling is the CNN neural network model using the self-attention mechanism, which processes the behavioral data in time sequence and identifies risks in combination with the personal identity information of the code visitor or submitter; the core of the data fusion risk control model is the neural network model, which is based on the judgment results of the rule judgment risk control model, the random forest risk control model and the deep learning risk control model, and combines the user identity information and the attributes of the code project to perform comprehensive analysis and processing to obtain the final risk judgment result; the core of the dynamic optimization model is the neural network model, which timely adjusts the model parameters of the rule judgment risk control model, the random forest risk control model, the deep learning risk control model and the data fusion risk control model according to the current input. The risk control model is composed of corresponding model sub-modules, which can timely adjust the risk control model and improve the response speed of risk anomalies; the risk control operation model is based on the model formed by the model platform to perform risk control judgment, mainly including the rule process risk control engine, random forest risk control engine, deep learning risk control engine, fusion judgment module, dynamic optimization module, data processing module and early warning module; the rule process risk control engine is based on the rule judgment risk control model to extract the data required by the rules from the data of the code hosting platform and the user center, and perform risk judgment; the random forest risk control engine is based on the random forest risk control model to extract the set data variables from the data of the code hosting platform and the user center, and perform risk judgment;The deep learning risk control engine directly performs risk assessments based on the deep learning risk control model, using raw data from the code hosting platform and the user center as input. The fusion assessment module, based on the data fusion risk control model, performs risk assessments based on the data from the code hosting platform and the user center, integrating the assessment results of the rule-based risk control engine, the random forest risk control engine, and the deep learning risk control engine. The dynamic optimization module, based on the dynamic optimization model and data from the code hosting platform and the user center, promptly adjusts and sets the model parameters of the rule-based risk control engine, the random forest risk control engine, the deep learning risk control engine, and the data fusion risk control module. The data processing module is responsible for collecting and cleaning the data required for the risk control platform, preparing the data according to the model requirements, and responding in real time to changes in the relevant parameter settings from the dynamic optimization engine to provide the required data. The early warning module promptly responds to the risk assessment results, records logs, and notifies relevant management personnel or directly performs operations such as access rights restriction.

[0060] The method provided by the present invention will be described in detail below with reference to specific embodiments.

[0061] 1. Model Training of Risk Control Model

[0062] The training of the risk control model includes the following steps:

[0063] Step 101: Collect massive amounts of log data from the code hosting platform and data from the user center, prepare training data, set risk levels, and label risk data based on domain expert knowledge.

[0064] Step 102: Based on the labeled data, a gradient descent optimization algorithm is used to train the process recommendation service model, the threshold recommendation service model, and the rule selection service recommendation model to obtain a recommendation service model required for risk control modeling.

[0065] Step 103: Combine the rule processes and train the risk control model based on the labeled data.

[0066] Step 104: Based on expert knowledge, attributes influencing risk assessment are pre-set, including code project attributes (project number, project name, project domain, project security level, number of project versions, latest project release date, total number of project issues, number of latest project version issues, number of project submitters, latest submission date, number of submissions in the last day, number of submissions in the last month, number of visitors in the last day, number of visitors in the last month, etc.), visitor attributes (visitor ID, visitor identity, visitor IP address, content of this visit, content of this submission, last three login IP addresses, last three visited projects, content of last three visits, projects to which last three submissions belong, content of last three submissions, etc.), and the random forest risk control model is constructed using the random forest algorithm based on the labeled data.

[0067] Step 105: Design a network model and train the deep learning risk control model using the gradient descent method based on the labeled data;

[0068] Step 106: Fix the model parameters of the discrimination rule risk control model, the random forest risk control model, and the deep learning risk control model, and train the data fusion model using the gradient descent method based on the labeled data;

[0069] Step 107: Fix the model parameters of the discrimination rule risk control model and the random forest risk control model, and slightly adjust the network parameters of the deep learning risk control model and the data fusion model;

[0070] Step 108: Fix the model parameters of the discrimination rule risk control model and train the corresponding submodule of the dynamic optimization model based on the labeled data;

[0071] Step 109: Fix the model parameters of the random forest risk control model and train the corresponding submodule of the dynamic optimization model based on the labeled data;

[0072] Step 110: Fix the model parameters of the deep learning risk control model and train the corresponding submodules of the dynamic optimization model based on the labeled data;

[0073] Step 111: Fix the model parameters of the data fusion model and train the corresponding submodules of the dynamic optimization model based on the labeled data;

[0074] 2. Risk Control of Code Hosting Platforms

[0075] Risk control of the code hosting platform includes the following steps:

[0076] The present invention provides a code hosting platform risk control method based on deep learning, which includes:

[0077] Step 201: Generate an initial risk control model using the risk control modeling platform, including the rule-based risk control model, the random forest risk control model, the deep learning risk control model, the data fusion model, and the dynamic optimization model.

[0078] Step 202: Load the initial risk control model into the corresponding engine or module of the risk control operation platform;

[0079] Step 203: Continuously collect data from the code hosting platform and the user center, including real-time data, log data, user data, user profiles, and other information. The data processing module of the risk control operation platform extracts, cleans, filters, and processes the data to meet the data requirements of the rule-based risk control engine, the random forest risk control engine, the deep learning risk control engine, the fusion judgment module, and the dynamic optimization module.

[0080] Step 204: The rule-based risk control engine determines the risk based on the set processes and rules, and submits the results to the fusion determination module.

[0081] Step 205: The random forest risk control engine uses a random forest algorithm to determine the risk and submits the result to the fusion determination module;

[0082] Step 206: The deep learning risk control engine directly performs risk assessment using the raw data from the code hosting platform and the user center as input, and submits the results to the fusion assessment module.

[0083] Step 207: The fusion determination module integrates the determination results of the rule-based risk control engine, the random forest risk control engine, and the deep learning risk control engine to perform risk determination and notify the early warning module of the result.

[0084] Step 208: The early warning module promptly responds to the risk assessment result, records the result in a log, and notifies relevant management personnel or directly performs operations such as access rights restriction.

[0085] Step 209: The dynamic optimization module optimizes and adjusts the model parameters and input data items of the rule-based risk control engine, the random forest risk control engine, the deep learning risk control engine, and the data fusion risk control module based on the project to which the code belongs and the collected data, to accurately respond to new risk changes.

[0086] Step 210: Continue to collect log data and result feedback from the code hosting platform and risk control platform, continuously optimize and improve the model, improve risk control accuracy, and enhance the ability to discover unknown risks.

[0087] The embodiment described above is only one specific implementation of the present invention. Common changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A code hosting platform risk control system based on deep learning, including a risk control modeling platform and a risk control operation platform, characterized in that: The risk control modeling platform and risk control operation platform are built on cloud infrastructure, and implement risk control throughout the code life cycle based on the code hosting platform and user center. The user center stores basic information of users including code visitors, submitters, and collaborators, and creates user profiles based on their identity information to form identity tags; The risk control modeling platform runs a rule-based risk control model, a random forest risk control model, a deep learning risk control model, a data fusion risk control model, and a dynamic optimization model; The rule-based risk control model is used to extract the data required by the rules from the code hosting platform and the user center data to perform risk assessment; The random forest risk control model is used to extract set data variables from the code hosting platform and the user center data to perform risk assessment; The deep learning risk control model is used to directly perform risk assessment using the raw data of the code hosting platform and the user center as input; The data fusion risk control model is used to comprehensively judge the data of the code hosting platform and the user center to make risk assessments; The dynamic optimization model is used to timely adjust model parameters for the code hosting platform and the data of the user center; The risk control operation platform performs risk control determination based on the rule determination risk control model, random forest risk control model, deep learning risk control model, data fusion risk control model and dynamic optimization model; The cloud infrastructure aggregates a large amount of computing, storage, and network resources, and provides machine learning, deep learning, and big data service capabilities; the code hosting platform runs on the cloud infrastructure, providing code hosting, code review, progress management, quality management, problem management, and continuous integration functions, realizing code hosting, collaborative development, and continuous release.

2. The code hosting platform risk control system based on deep learning according to claim 1 is characterized in that: The risk control modeling platform is responsible for risk control model modeling, including three sub-platforms: model platform, rule platform and process platform; The rule platform is responsible for generating rules and providing preset rules based on domain experts; The process platform is responsible for the creation and setting of risk control detection processes; The model platform is based on the general machine learning, deep learning training and reasoning services provided by the cloud infrastructure, and combines the code hosting business to form rule judgment risk control models, random forest risk control models, deep learning risk control models, data fusion risk control models and dynamic optimization models. At the same time, it provides intelligent recommendations of relevant rule thresholds, selection of judgment rules and intelligent combination recommendations for the rule platform, and provides process combination recommendation services for the process platform.

3. The code hosting platform risk control system based on deep learning according to claim 2 is characterized in that: The rule-based risk control model is a risk control model formed based on the rule platform and the process platform; the deep learning risk control model is an end-to-end model based on the original operation log behavior data provided by the code hosting platform. The core is a CNN neural network model that uses a self-attention mechanism to process time-sequential behavior data and identify risks in combination with the personal identity information of the code visitor or submitter; the core of the data fusion risk control model is a neural network model, which is based on the judgment results of the rule-based risk control model, the random forest risk control model and the deep learning risk control model, and combined with the user identity information and the attributes of the code project to perform comprehensive analysis and processing to obtain the final risk judgment result.

4. The code hosting platform risk control system based on deep learning according to claim 2 is characterized in that: The core of the dynamic optimization model is the neural network model, which timely adjusts the model parameters of the rule-based risk control model, the random forest risk control model, the deep learning risk control model, and the data fusion risk control model based on the current input. It is composed of corresponding model sub-modules, and timely adjusts the risk control model to improve the response speed to risk anomalies.

5. The code hosting platform risk control system based on deep learning according to claim 2 is characterized in that: The risk control operation model is a model that performs risk control judgment based on the model formed by the model platform. The risk control operation model mainly includes a rule process risk control engine, a random forest risk control engine, a deep learning risk control engine, a fusion judgment module, a dynamic optimization module, a data processing module and an early warning module. The rule process risk control engine extracts the data required by the rules from the data from the code hosting platform and the user center based on the rule judgment risk control model and performs risk judgment. The random forest risk control engine extracts the set data variables from the data from the code hosting platform and the user center based on the random forest risk control model and performs risk judgment. The deep learning risk control engine directly performs risk judgment based on the raw data from the code hosting platform and the user center as input based on the deep learning risk control model. The fusion judgment module is based on the data fusion risk control model, and performs risk judgment on the data from the code hosting platform and the user center by integrating the judgment results of the rule process risk control engine, the random forest risk control engine and the deep learning risk control engine.

6. The code hosting platform risk control system based on deep learning according to claim 5 is characterized in that: The dynamic optimization module is based on the dynamic optimization model and the data from the code hosting platform and the user center, and timely adjusts and sets the model parameters of the rule judgment risk control engine, the random forest risk control engine, the deep learning risk control engine and the fusion judgment module; the data processing module is responsible for collecting the data required for cleaning the risk control platform and preparing the data according to the requirements of the model, while responding to the relevant parameter setting changes from the dynamic optimization module in real time and providing the required data; the early warning module responds to the risk judgment results in a timely manner, records logs, and notifies relevant management personnel or directly performs operations with limited access rights.

7. The method of the code hosting platform risk control system based on deep learning according to claim 6 is characterized in that: Model training for risk control models includes the following steps: Step 101: Collect massive code hosting platform log data and user center data, prepare training data, set risk levels, and label risk data based on domain expert knowledge; Step 102: Based on the labeled data, a gradient descent optimization algorithm is used to train a process recommendation service model, a threshold recommendation service model, and a rule selection service recommendation model to obtain a recommendation service model required for risk control modeling. Step 103: Combine the rule processes and train the rule-based risk control model based on the labeled data; Step 104: Based on expert knowledge, attributes that affect risk assessment are pre-set, including code project and visitor attributes. Based on the labeled data, a random forest algorithm is used to construct the random forest risk control model. Step 105: Design a network model and train the deep learning risk control model using the gradient descent method based on the labeled data; Step 106: Fix the model parameters of the rule-based risk control model, the random forest risk control model, and the deep learning risk control model, and train the data fusion risk control model using a gradient descent method based on the labeled data. Step 107: Fix the model parameters of the rule-based risk control model and the random forest risk control model, and slightly adjust the network parameters of the deep learning risk control model and the data fusion risk control model; Step 108: Fix the model parameters of the rule-based risk control model and train the corresponding submodules of the dynamic optimization model based on the labeled data; Step 109: Fix the model parameters of the random forest risk control model and train the corresponding submodule of the dynamic optimization model based on the labeled data; Step 110: Fix the model parameters of the deep learning risk control model and train the corresponding submodules of the dynamic optimization model based on the labeled data; Step 111: Fix the model parameters of the data fusion risk control model, and train the corresponding sub-modules of the dynamic optimization model based on the labeled data to achieve dynamic optimization of the four models in steps 108-111.

8. The method of a code hosting platform risk control system based on deep learning according to claim 7, characterized in that: Risk control for code hosting platforms includes the following steps: Step 201: Using the risk control modeling platform, generate an initial risk control model through steps 101-111, including the rule-based risk control model, the random forest risk control model, the deep learning risk control model, the data fusion risk control model, and the dynamic optimization model; Step 202: Load the initial risk control model into the static model memory to form a service, and then load it into the engine or module corresponding to the risk control operation platform; Step 203: Continuously collect data from the code hosting platform and the user center, including real-time data, log data, user data, and user profile information. The data processing module of the risk control operation platform extracts, cleans, and filters the data to meet the data requirements of the rule-based risk control engine, the random forest risk control engine, the deep learning risk control engine, the fusion judgment module, and the dynamic optimization module. Step 204: The rule-based risk control engine determines the risk based on the set processes and rules, and submits the results to the fusion determination module. Step 205: The random forest risk control engine uses a random forest algorithm to determine the risk and submits the result to the fusion determination module; Step 206: The deep learning risk control engine directly performs risk assessment using the raw data from the code hosting platform and the user center as input, and submits the results to the fusion assessment module. Step 207: The fusion determination module integrates the determination results of the rule-based risk control engine, the random forest risk control engine, and the deep learning risk control engine to perform risk determination and notify the early warning module of the result. Step 208: The early warning module promptly responds to the risk determination result, records the result in a log, and notifies relevant management personnel or directly performs access permission restriction operations; Step 209: The dynamic optimization module optimizes and adjusts the model parameters and input data items of the rule-based risk control engine, the random forest risk control engine, the deep learning risk control engine, and the fusion decision module based on the project to which the code belongs and the collected data, so as to accurately respond to new risk changes. Step 210: Continue to collect log data and result feedback from the code hosting platform and risk control platform, continuously optimize and improve the model, improve risk control accuracy, and enhance the ability to discover unknown risks.

Images