Full-period asset safety guarantee method and device for nuclear power industrial model

Abstract

The invention discloses a full-cycle asset security guarantee method and device for a nuclear power industrial model, and the method comprises the following steps: carrying out the preprocessing of a data set, optimization of model parameters, and setting of a corresponding defense strategy for data poisoning in a model compiling stage, i.e., a model generation stage; in the model transmission stage, a secure transmission protocol is set, a virtual private network is adopted, encryption transmission and uploading are carried out on the model, a secret key is managed and stored, and integrity verification is carried out on transmitted model data; in the model storage stage, encryption and access authority control are performed on the stored model data, and a data backup and recovery strategy is set for the stored model data; performing access authority control in a model operation calling stage, and performing operation calling on the model by adopting isolation and sandbox technologies; access authority control is carried out in the model sharing stage, and a corresponding application program interface is set for model sharing. According to the method, the asset safety of the nuclear power industrial model in the whole life cycle can be guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nuclear power industry model asset security, and particularly relates to a nuclear power industry model full-cycle asset security method and device. BACKGROUND

[0002] With the rapid development of the nuclear power industry and the in-depth promotion of digital nuclear power construction, the demand for informationization and intelligentization application is increasingly booming, and the security threats faced by nuclear power industry models are also increasingly diversified. These models involve a large amount of data interaction, algorithm operation and model simulation, and their security is directly related to the stable operation and overall safety of the nuclear industry control system.

[0003] While nuclear industry enterprises are increasingly strengthening network security awareness and effectively fulfilling the main responsibility for network security, they should use systematic thinking and digital information means to scientifically promote model security protection work and accelerate the construction of a model security protection system. SUMMARY

[0004] Therefore, it is necessary to provide a nuclear power industry model full-cycle asset security method and device to effectively solve the technical problem of inadequate nuclear power industry model security protection.

[0005] The present application provides a nuclear power industry model full-cycle asset security method, comprising the following steps: Step S1, in the model generation stage, the data set relied on by the model generation is preprocessed, the model parameters are optimized, and the corresponding defense strategy is set for data poisoning; Step S2, in the model transmission stage, a secure transmission protocol is set, a virtual private network is used, the model is encrypted and uploaded, the key in the transmission process is hosted and stored, and the integrity of the model data after transmission is verified; Step S3, in the model storage stage, the stored model data is encrypted and access control is performed, the corresponding data backup strategy and data recovery strategy are set for the stored model data; Step S4, in the model running and calling stage, the model is run and called by using isolation and sandbox technology, access control, input data verification and filtering, security vulnerability scanning, security audit and log monitoring are performed on the running and calling of the model; Step S5, in the model sharing stage, access control is performed on the model sharing, the corresponding application program interface is set for the model sharing, and the model is encapsulated and packaged before sharing; a firewall and an intrusion detection defense system are deployed to defend against network threats of the model.

[0006] Preferably, the data relied on by the model generation in step S1 is preprocessed, specifically: Data cleaning, data conversion, data preprocessing and data verification are performed on the data set; The data cleaning includes missing value processing and outlier processing, the outlier processing includes filtering data, identifying and proposing outliers by using statistical methods, and the missing value processing includes processing missing values by using direct deletion, interpolation calculation and machine learning methods; the data preprocessing includes data conversion and feature engineering, the data conversion includes data standardization and feature encoding, the feature encoding is realized by using label encoding and one-hot encoding, the feature engineering includes feature construction and feature selection, the feature construction uses numerical features, text features or interaction features to construct, and the feature selection selects key features through correlation analysis and recursive feature elimination; the data verification divides the data set into a training set, a validation set and a test set by using cross-validation, generates multiple different data sets by using repeated sampling for random sampling, and detects consistency of the data set by using consistency test.

[0007] Preferably, the step S1 optimizes model parameters, specifically: An automatic parameter tuning technology is used to automatically explore the parameter space, a model performance evaluation index is used as a target function, the data set is used for model training and verification, and the optimal parameters are obtained to complete the optimization of model parameters.

[0008] Preferably, the step S1 sets a corresponding defense strategy for data poisoning, specifically: A differential privacy technology is used to protect the privacy of the data set; a disturbance or noise is introduced in the model training process to enhance the resistance of the model to malicious input; an access control mechanism is set for the data set to limit unauthorized users from accessing or modifying the data set; at the same time, real-time log records in the model generation process are audited for security.

[0009] Preferably, the step S2 encrypts and uploads the model, specifically: Different encryption methods are used for encryption according to different types of models and different contents thereof; The model includes an industrial mechanism model and an industrial artificial intelligence model; The industrial mechanism model includes a basic theory model, a process model and a simulation model; The encryption of the basic theoretical model specifically comprises: using an encryption algorithm to encrypt the basic theoretical model, using AES or RSA to encrypt key data in the basic theoretical model, and setting a specific encryption algorithm for specific data in the basic theoretical model; using code obfuscation technology to change variable names, function names and logical structures of the basic theoretical model; using model segmentation technology to split the basic theoretical model into multiple modules, and combining the modules to run the completed basic theoretical model; The encryption mode of the process model specifically comprises: selecting a symmetric encryption algorithm to encrypt the process model, and selecting an asymmetric encryption algorithm to encrypt key information of the process model; for a configuration file of the process model, using a special encryption tool to encrypt; for a parameter model executable by a solver, adding an encryption step in the model generation or calling process to realize encryption; for an FMU model encapsulated in a standard specification, adding an encryption mechanism in the encapsulation process, encrypting the FMU file based on the encryption mechanism to realize encryption of the FMU model; The encryption mode of the simulation model specifically comprises: for a simulation model file, using a file encryption tool to encrypt; for a model code file, encrypting a source code file before compiling; for a binary form simulation model file, using binary encryption to encrypt machine code of the binary file; The industrial artificial intelligence model comprises a machine learning model, a deep learning model and an industry large model; The encryption of the machine learning model specifically comprises: for a machine learning model file, using a symmetric encryption algorithm or an asymmetric encryption algorithm to encrypt; for a model code file, obfuscating a source code to realize encryption; for a binary model file of the machine learning model, using a special encryption tool to encrypt; for a machine learning model taking a container as a carrier, creating a corresponding encrypted file system on a host computer, and mounting the encrypted file system to the container as a data volume, files stored in the data volume are automatically encrypted; The encryption mode of the deep learning model specifically comprises: for a deep learning model file, using a symmetric encryption algorithm or an asymmetric encryption algorithm to encrypt in an overall encryption mode; for a model code file, using a source code obfuscation method to realize encryption, and compiling the source code into bytecode, encrypting the bytecode file, and using dynamic decryption execution for key code segments; for a deep learning model taking a container as a carrier, encrypting the entire image file, and using access control functions provided by a container platform to control access permissions of model files of the deep learning model; The encryption mode of the industry large model is specifically: using an encryption algorithm to convert the model file of the industry large model into encrypted binary data; encrypting the key parameters and weight data of the industry large model; at the same time, implementing access control, input verification and anomaly detection on the inference service engine, and using a secure communication protocol to realize the communication between the inference server and the client.

[0010] Preferably, a secure transmission protocol is set, a virtual private network is adopted, and the key in the transmission process is hosted and stored, and the integrity of the data completed transmission is verified, specifically: All data transmitted between the client and the target server is encrypted using the HTTPS hypertext transfer protocol; the encrypted model data is encrypted and transmitted using the SFTP secure file transfer protocol and identity verification; A VRN virtual private network is adopted to create an encrypted channel on a public network to ensure the security of data when a remote user accesses internal resources of the platform; the VRN virtual private network uses encryption technology to encrypt the transmitted data, converts the data into an unreadable form, and the recipient decrypts and reads the data by using a decryption key; A hardware security module or a third-party server is used to store and manage the key; The data completed transmission is verified for integrity by using a hash value comparison, a digital signature or a checksum; according to the model type, the application scenario and the data characteristics, a specific integrity verification mode is selected.

[0011] Preferably, the step S3 is specifically: Sensitive data including model parameters and weights are protected by using an encryption algorithm; a permission management mechanism is established to manage access to model data based on the permission management mechanism, while limiting direct access to sensitive data, recording user access behavior and performing auditing; based on the importance, update frequency and storage space of the data, a corresponding backup strategy is set; a data recovery mechanism is established, and when the data is problematic, the model is recovered based on the backup data and the data recovery mechanism, and the integrity and availability of the backup data are verified periodically.

[0012] Preferably, the step S4 is specifically: The role-based access control is adopted, a role set is established for various permissions, roles in the role set are assigned to different users, and thus the permission of accessing specific resources in the system is controlled; the model running environment is isolated from other systems by using a virtual machine or a virtual machine technology, and the model running safety is ensured; the debugging model is run in a sandbox environment, and the model running permission is limited; input data in the running and debugging process is subjected to input verification and input filtering; the model is subjected to security vulnerability scanning periodically; the model running log and calling record are monitored in real time, and the model running log and calling record are subjected to security audit periodically.

[0013] Preferably, the step S5 specifically includes: Access permission control is performed on model sharing; an API interface is set, and the API interface is used for model sharing; the model is encapsulated and packaged before model sharing, and the input interface and the output interface of the model are defined in the encapsulation process; after the encapsulation of the model is completed, test verification is performed, the encapsulated model is deployed to a target environment, and is distributed to users.

[0014] The application further provides a nuclear power industry model full-cycle asset security guarantee device, including a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to realize the nuclear power industry model full-cycle asset security guarantee method.

[0015] Compared with the prior art, the application has the beneficial effects that different security guarantee strategies are set for the generation stage, the transmission stage, the storage stage, the running calling stage, the sharing stage and network threats of the nuclear power industry model, and the asset safety in the full life cycle is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application, and do not limit the application in any way. In the drawings: Fig. 1 is a flowchart of an embodiment of the nuclear power industry model full-cycle asset security guarantee method provided by the application; Figure 2 is Figure 1 Fig. 2 is a framework diagram of an embodiment of the nuclear power industry model full-cycle asset security guarantee method in the embodiment shown. DETAILED DESCRIPTION

[0017] The preferred embodiments of the application are specifically described below with reference to the drawings, wherein the drawings constitute a part of this application, and together with the embodiments of the application serve to explain the principles of the application, and are not used to limit the scope of the application.

[0018] Embodiment one Please refer toFigure 1 The full-cycle asset security protection method of a nuclear power industry model in this embodiment includes the following steps: Step S1, in the model generation phase, pre-process the data set relied on by the model generation, optimize the model parameters, and set the corresponding defense strategy for data poisoning; Step S2, in the model transmission phase, set a secure transmission protocol, use a virtual private network to encrypt the model transmission and upload, and host and store the key during transmission, and verify the integrity of the model data after transmission is completed; Step S3, in the model storage phase, encrypt and access control the stored model data, set the corresponding data backup strategy and data recovery strategy for the stored model data; Step S4, in the model running call phase, use isolation and sandbox technology to run and call the model, and access control the model running call; Step S5, in the model sharing phase, access control the model sharing, set the corresponding application program interface for the model sharing, and package and package the model before sharing; deploy a firewall and an intrusion detection defense system to defend against network threats to the model.

[0019] This embodiment sets different security protection strategies in the generation phase, transmission phase, storage phase, running call phase, sharing phase and network threat of the nuclear power industry model, to ensure the asset security throughout the life cycle.

[0020] In the model generation phase, data quality and accuracy problems, improper parameter adjustment, and security vulnerabilities and network attacks can all pose a threat to model security. In the nuclear power industry, the data relied on by the model generation may come from multiple different sources, such as sensors, monitoring devices, operation records, etc. Due to device failure, data transmission error or human error, these data may contain noise, outliers or missing values. These problems will directly affect the training effect of the model, causing the model to be unable to accurately reflect the true state of the nuclear power system. If the data is tampered with or forged by humans, it may also lead to misleading results of the model, posing a threat to the safe operation of the nuclear power system.

[0021] During the model generation process, the adjustment of parameters has an important influence on the performance of the model. If the parameters are not set properly, such as a too large learning rate or a too small regularization parameter, the model training may be unstable, and overfitting or underfitting may occur. This not only affects the prediction accuracy of the model, but also increases the uncertainty of the model in actual application.

[0022] During the model generation process, if there are security vulnerabilities in the software or hardware used, attackers can exploit these vulnerabilities to obtain sensitive data, tamper with the model structure, or inject malicious code. This not only may cause the model generation to fail, but also may leak critical information of the nuclear power system, posing a threat to the safe operation of the nuclear power facility. For example, attackers may attack the model through data poisoning and malicious software injection. In a data poisoning attack, attackers may tamper with the operation data of the nuclear power plant, causing the model to learn incorrect behavior patterns or associations, resulting in incorrect decisions by the model in actual application. At the same time, attackers may also introduce malicious software, such as viruses, Trojans, or ransomware, into the model training process to disrupt the training process of the model or steal sensitive data. These malicious software are hidden in training data, model code, or dependent libraries, making them difficult to detect and remove.

[0023] When transmitting model data from the local to the platform, model encrypted transmission can ensure the security, integrity, and confidentiality of the data. The present embodiment ensures the security of platform model transmission from four aspects of encrypted transmission process, encryption technology, security protocol, and key storage and management.

[0024] The model storage security guarantee method mainly solves how to ensure the security of model data during the storage process, preventing data leakage or tampering. The present embodiment mainly guarantees the security of model storage from three aspects of data encryption, access control, data backup and recovery.

[0025] During the platform model sharing process, it is crucial to ensure the privacy and security of the model content. There may be scenarios where users do not want to see the model kernel after sharing the model. We guarantee the security of platform model sharing from three aspects of setting access control permissions, API interface design, and model packaging.

[0026] Specifically, the step S1 preprocesses the data relied on by the model generation, specifically: The data set is cleaned, preprocessed, and verified. The data cleaning includes filtering the data, identifying and proposing outliers using statistical methods, and processing missing values using interpolation methods and machine learning techniques; the data preprocessing includes data standardization, feature selection, and encoding; the data verification is to divide the data set into training set, validation set and test set by cross-validation, generate multiple different data sets by random sampling with repeated sampling, and detect the consistency of the data set by consistency test.

[0027] Data from different sources, such as sensors, monitoring devices, operational records, etc., need to be cleaned to eliminate noise, outliers, and missing values. At the same time, data validation techniques such as cross-validation and resampling are used to ensure the accuracy and reliability of the model learning data.

[0028] In the process of generating nuclear power industry models, data cleaning is the first step to ensure data quality. By using filtering techniques such as moving average filtering and Kalman filtering, sensor errors and environmental interference noise can be effectively removed, thus smoothing the data. At the same time, using statistical methods to identify and eliminate outliers can prevent equipment failure or human error from interfering with model training. For missing value processing, interpolation methods and machine learning techniques are used to ensure data integrity and accuracy.

[0029] Data preprocessing involves data standardization, feature selection, and encoding. The standardization process allows data from different sources to be compared on the same scale, improving the stability of model training. Feature selection can remove redundant and irrelevant feature information, further improving model performance. Based on methods such as one-hot encoding, classification features can be adapted to the specific requirements of numerical calculations.

[0030] To ensure the reliability of data quality and model performance, various data validation techniques can be used. Through cross-validation, the data set is divided into training set, validation set and test set to fully evaluate the generalization ability of the model. The resampling technique generates multiple different data sets by random sampling, providing rich data samples for model training and reducing the result bias caused by the specific division of the data set. In addition, consistency testing can ensure that data from different sources or different time periods is logically consistent, enhancing the reliability and stability of the model. These measures together constitute an important part of data quality management, providing a solid data foundation for the generation of nuclear power industry models.

[0031] Specifically, the step S1 optimizes the model parameters, specifically: Automatic parameter tuning technology is used to automatically explore the parameter space, with model performance evaluation indicators as the objective function, and the data set is used for model training and verification to obtain the optimal parameters, completing the optimization of model parameters.

[0032] In the generation of nuclear power industry models, parameter adjustment is a crucial step that directly affects the stability and prediction accuracy of the model. To optimize model parameters, automatic parameter tuning technology is used in combination with expert experience to achieve more accurate and efficient parameter adjustment.

[0033] Automatic parameter tuning is a method that uses algorithms to automatically explore the parameter space to find the optimal parameter combination. In this embodiment, grid search is used to find the optimal parameter value that makes the model performance optimal by traversing each possible combination in the parameter space. Although grid search requires a large amount of computation, it can fully explore the parameter space and ensure that no possible excellent parameter combination is missed. Another automatic parameter tuning technique used in this embodiment is random search, which can obtain relatively good results in a limited time by randomly sampling in the parameter space. In addition, Bayesian optimization is also an efficient automatic parameter tuning technique that uses Bayesian theorem to guide the search process and gradually approaches the optimal parameter combination by iteratively updating the parameter distribution.

[0034] During parameter adjustment, model performance evaluation indicators also need to be combined for specific optimization. Accuracy, recall rate and other indicators can directly reflect the ability of the model to identify positive and negative examples. By continuously adjusting parameters and observing the changes in these indicators, the optimal parameter combination that makes the model performance optimal is gradually found. At the same time, in order to more comprehensively evaluate the performance of the model, the data set is divided into multiple subsets in combination with the cross-validation technique in the previous section, and training and verification are performed respectively, effectively avoiding overfitting problem, and the more reliable performance evaluation results are used to guide parameter adjustment.

[0035] Specifically, the step S1 sets the corresponding defense strategy for data poisoning, specifically: Use differential privacy technology to protect the privacy of the data set; introduce disturbance or noise in the model training process to enhance the model's resistance to malicious input; set access control mechanisms for the data set to limit unauthorized users from accessing or modifying the data set; at the same time, security audit of real-time log records during model generation process.

[0036] In the model generation stage of the nuclear power industry, data poisoning attacks are a highly threatening security challenge. Since the nuclear power industry has very high requirements for the accuracy and reliability of the model, any slight data disturbance can have a significant impact on the model performance, and thus pose a serious threat to the safe operation of nuclear power facilities.

[0037] The essence of data poisoning attacks is that attackers tamper with or inject malicious data to disrupt the training process of the model. In the nuclear power industry, these data may involve critical information such as operating parameters of nuclear power plants, sensor readings, device status, etc. Attackers may add fake data, modify real data or delete important data in the training set, etc. to make the model learn incorrect behavior patterns or correlations. Once the model is affected by the poisoning attack, it may make incorrect decisions in actual application, such as misjudging device status, incorrectly adjusting operating parameters, etc., which may cause serious safety accidents.

[0038] To prevent data poisoning attacks, the nuclear power industry needs to take a series of security measures. First, strengthen data cleaning and preprocessing, ensure the authenticity and integrity of training data, use data validation techniques to detect outliers and noise, and eliminate potential data pollution through data filtering and standardization. Second, establish strict data access control mechanisms to limit access and modification of training data only to authorized users or systems. At the same time, implement security audits and log records to trace and trace in the event of security problems.

[0039] In addition, the embodiment uses differential privacy technology to protect the privacy of training data and prevent sensitive information from being leaked to attackers. In addition, the introduction of adversarial training technology enhances the model's resistance to malicious input by introducing perturbations or noise during training, to some extent, improving the model's defense capability against poisoning attacks.

[0040] Step S2 is specifically: Select an encryption algorithm and generate a corresponding key, encrypt the model data based on the encryption algorithm and the generated key; After encryption is complete, use a secure communication protocol to transmit the encrypted model data to the target server; After the target server receives the encrypted model data, it uses the corresponding key to decrypt the encrypted model data; Calculate the checksum of the decrypted model data, verify the integrity of the model data based on the comparison of the checksum; or by calculating the hash value of the decrypted model data, verify the integrity of the model data by comparing the hash value; Store the decrypted and verified model data in the local file system.

[0041] Specifically, the model is encrypted and transmitted and uploaded in step S2, specifically: Different encryption methods are used to encrypt different types of models and different contents therein; The model includes an industrial mechanism model and an industrial artificial intelligence model; The industrial mechanism model includes a basic theory model, a process model, and a simulation model; Encrypting the basic theory model specifically includes: using an encryption algorithm to encrypt the basic theory model, using AES or RSA to encrypt key data in the basic theory model, setting a specific encryption algorithm for specific data in the basic theory model; using code obfuscation technology to change variable names, function names, and logical structures of the basic theory model; using model segmentation technology to split the basic theory model into multiple modules, and combining the modules to run the completed basic theory model; The encryption of the process model specifically includes: using a symmetric encryption algorithm to encrypt the process model, and using an asymmetric encryption algorithm to encrypt the key information of the process model; for the configuration file of the process model, a special encryption tool is used for encryption; for the parameter model executable by the solver, an encryption step is added in the model generation or calling process to realize encryption; for the FMU model encapsulated in a standard specification, an encryption mechanism is added in the encapsulation process, and the FMU file is encrypted based on the encryption mechanism to realize the encryption of the FMU model; The encryption of the simulation model specifically includes: for the simulation model file, a file encryption tool is used for encryption; for the model code file, the source code file is encrypted before compilation; for the binary form of the simulation model file, binary encryption is adopted to encrypt the machine code of the binary file; The industrial artificial intelligence model includes a machine learning model, a deep learning model, and an industry large model; The encryption of the machine learning model specifically includes: for the machine learning model file, a symmetric encryption algorithm or an asymmetric encryption algorithm is used for encryption; for the model code file, the source code is obfuscated to realize encryption; for the binary model file of the machine learning model, a special encryption tool is used for encryption; for the machine learning model with a container as a carrier, a corresponding encrypted file system is created on the host machine, and the encrypted file system is mounted to the container as a data volume, and the files stored in the data volume are automatically encrypted; The encryption of the deep learning model specifically includes: for the deep learning model file, a symmetric encryption algorithm or an asymmetric encryption algorithm is used for encryption in an overall encryption mode; for the model code file, a source code obfuscation method is used to realize encryption, and the source code is compiled into bytecode, and the bytecode file is encrypted, and the key code segment is executed dynamically; for the deep learning model with a container as a carrier, the entire image file is encrypted, and the access control function provided by the container platform is used to control the access permission of the model file of the deep learning model; The encryption of the industry large model specifically includes: using an encryption algorithm to convert the model file of the industry large model into encrypted binary data; encrypting the key parameters and weight data of the industry large model; at the same time, access control, input verification and anomaly detection are implemented on the inference service engine, and a secure communication protocol is used to realize the communication between the inference server and the client.

[0042] In the process of uploading and encrypting the transmission of the model, it is essential to ensure the security and confidentiality of the data. The specific encryption transmission process is described as follows.

[0043] First, the appropriate encryption algorithm needs to be selected based on business requirements and security requirements. The choice of encryption algorithm directly affects the level of data security and transmission efficiency. The encryption algorithms used in this embodiment include AES (Advanced Encryption Standard) and RSA, etc. AES is a symmetric encryption algorithm that uses the same key for encryption and decryption, suitable for large amounts of data encryption. RSA is an asymmetric encryption algorithm that uses a pair of keys for encryption and decryption, where the public key is used for encryption and the private key is used for decryption.

[0044] After selecting the encryption algorithm, the corresponding key needs to be generated. For symmetric encryption algorithms, a unique key needs to be generated and both the client and server need to securely obtain this key. This key needs to be properly managed and cannot be leaked to unauthorized personnel. For asymmetric encryption algorithms, a pair of public and private keys needs to be generated. The public key can be publicly distributed, while the private key must be strictly confidential and only the entity that owns the private key can decrypt encrypted data.

[0045] After that, the model data needs to be encrypted according to the selected encryption algorithm and key. The encryption process converts the original data into unreadable ciphertext, ensuring that even if the data is intercepted during transmission, the original content of the model cannot be directly obtained. Encrypted data will have a high degree of confidentiality, and only entities with the corresponding key can decrypt and access it.

[0046] When encryption is complete, the model data is transmitted to the target server through the network. During transmission, we use secure communication protocols such as HTTPS to protect the security of the data. The HTTPS protocol uses SSL / TLS encryption technology at the transmission layer to encrypt and verify communication data, ensuring data confidentiality and integrity. In this way, even if the data is intercepted during transmission, attackers cannot decrypt and read its contents.

[0047] When the encrypted data reaches the server, the server uses the corresponding key to decrypt the received encrypted data. For symmetric encryption algorithms, the server uses the same key as the client for decryption. For asymmetric encryption algorithms, the server uses its own private key for decryption. The decryption process restores the ciphertext to the original model data, allowing the server to correctly store and use the model.

[0048] After decryption, verification is also needed to ensure the integrity of the data. This can be done by calculating a checksum or using a hash function. A checksum is a method of calculating a value for a block of data using a specific algorithm and transmitting it along with the original data block. The receiver calculates the checksum for the received data block using the same algorithm and compares it with the checksum provided by the sender to verify the integrity of the data. A hash function, on the other hand, can map data of any length to a fixed-length hash value, with one-way and collision resistance. By calculating the hash value of the original data and comparing it with the hash value of the received data, we can verify whether the data has been tampered with.

[0049] Finally, once the model data is successfully decrypted and passes the integrity verification, the server can store it in the local file system or directly load the model into memory for use. When storing model data, the server needs to take appropriate security measures such as access control, backup and recovery, etc. to ensure the security and availability of model data.

[0050] Through the above encryption transmission process, the security and confidentiality of model data during the upload process can be ensured. Selecting appropriate encryption algorithms, generating and properly managing keys, using secure communication protocols, and verifying data integrity are all important steps to ensure the security of model data. At the same time, it is also necessary to regularly update the key, monitor any abnormal activities, and take other security measures to cope with the changing security threats.

[0051] In the encryption transmission of model upload, various encryption algorithms are usually used. These algorithms convert the original data plaintext into unreadable ciphertext format by encrypting the data, thus protecting the confidentiality of the data. Encryption technology can be divided into symmetric encryption and asymmetric encryption.

[0052] The biggest feature of symmetric encryption is that the same key is used for encryption and decryption. If a key is used to encrypt data, the same key must be used to decrypt the data. Since there is only one key, the distribution and management of the key become a challenge for symmetric encryption.

[0053] This embodiment selects AES as one of the encryption algorithms, supports multiple key lengths, and has an efficient algorithm design, while strengthening key management to ensure data security.

[0054] The asymmetric encryption algorithm used in this embodiment is: Rivest-Shamir-Adleman, abbreviated as RSA, and Elliptic Curve Cryptography, abbreviated as ECC.

[0055] Symmetric encryption is generally faster than asymmetric encryption, but asymmetric encryption is more secure in terms of key management, as public keys can be made public while private keys remain secret. Symmetric encryption, on the other hand, requires ensuring that both parties share the same key, which can be a challenge in distributed systems.

[0056] Different encryption methods are used for different types of models and their different contents.

[0057] Encryption of basic theoretical models is an important step in ensuring the security of industrial mechanism models. Since basic theoretical models are mainly based on fluid mechanics, thermodynamics, aerodynamics, and other basic theoretical knowledge, and are expressed in the form of expressions, functions, calculation formulas, and theoretical rules, the encryption process needs to balance data security and model usability.

[0058] First, use encryption algorithms to encrypt the model. Use strong encryption algorithms such as AES or RSA to encrypt the key data in the basic theoretical model, ensuring that even if the model file is illegally obtained, it is difficult to easily parse sensitive information. At the same time, specific encryption algorithms are designed for specific data in the model to further improve security.

[0059] Second, use code obfuscation techniques to increase the difficulty of reading for illegal visitors. By changing variable names, function names, logic structures, and other ways, the expressions, functions, and calculation formulas in the model become difficult to understand and reverse engineer. Not only can it improve the security of the model, but it also reduces the risk of being illegally copied or spread.

[0060] Finally, use model segmentation techniques to encrypt the basic theoretical model. Split the basic theoretical model into multiple modules or components, each containing only part of the data and logic. By combining these modules to run the complete model, the risk of a single module being leaked is reduced, while the flexibility and maintainability of the model are improved.

[0061] Process model encryption, process models contain important process, formula, parameter and other information in the production process, once these information is leaked, it may cause significant loss, so it is necessary to encrypt the process model to ensure the security of the process model.

[0062] According to the importance and sensitivity of the process model, first of all, we can choose symmetric encryption algorithm or asymmetric encryption algorithm. Symmetric encryption algorithm such as AES has high efficiency and simplicity, suitable for encryption of large amount of data; asymmetric encryption algorithm such as RSA, although the operation cost is higher, but provides higher security, suitable for encryption of key information.

[0063] For process model configuration files stored in XML, TXT, and other formats, use a specialized XML encryption library or file encryption tool for encryption. At the same time, use the XML encryption standard to encrypt XML files to ensure that the content is not stolen or tampered with during transmission. These tools and encryption standards provide powerful encryption algorithms and flexible encryption options to ensure the security of file content.

[0064] For parameter models that solvers can run, add an encryption step during model generation or invocation. For example, when exporting the model, use an encryption algorithm to encrypt the model parameters, and decrypt them when importing the model. This ensures the security of parameter models during transmission and use.

[0065] For FMU models packaged in FMI, OSLC, and other standard specifications, add an encryption mechanism during packaging and select a specific encryption method. By encrypting the FMU file, unauthorized access and modification are prevented. At the same time, decryption is also required when the model is invoked to ensure correct operation of the model.

[0066] Simulation model encryption is an important part of ensuring the security of simulation modeling data for industrial scene physical and chemical changes.

[0067] For simulation model files with ".m", ".mdl", ".sxl" suffixes, use file encryption tools or algorithms to encrypt these files. Ensure that only users with the correct key can decrypt and access the file content.

[0068] For model code files with ".c" suffix, the source code file needs to be encrypted before compilation and decrypted for execution at runtime to prevent unauthorized access and modification.

[0069] For simulation model files in the form of ".so", ".dll", etc., which are dynamic link libraries or shared objects, binary encryption is performed on the binary file, and the machine code of the binary file is encrypted to ensure that decryption is required when loading and executing. By using a specialized encryption tool or library, the binary file is encrypted and decrypted. When the model is loaded, it is first decrypted, and then normal linking and execution are performed.

[0070] Machine learning model encryption, for machine learning model files with ".pkl" and ".pmml" suffixes, use the above symmetric encryption and asymmetric encryption algorithms for encryption.

[0071] For model code files with ".py" suffix, the code is obfuscated to make it difficult to understand and reverse engineer. By changing variable names, function names, comments, etc., the difficulty of analyzing the code by attackers is increased. At the same time, source code protection tools and source code encryption tools are used to encrypt the source code file, and the encrypted code is embedded in the program. During program execution, the source code is restored and executed through decryption algorithm. In this way, direct access and viewing of the source code can be prevented.

[0072] For machine learning model files with "jar" suffix, use professional encryption tools or libraries such as xjar to encrypt the entire JAR package. Such tools usually provide encryption options, allowing you to set encryption passwords and other related parameters. The encrypted JAR package will become an executable binary file. Before execution, you need to configure the corresponding decryption environment or Java agent to decrypt and execute the code in it at runtime. The overall encryption of the JAR package ensures the security of the entire package by setting encryption passwords and parameters, preventing unauthorized access and modification.

[0073] For machine learning models in container format, create a corresponding encrypted file system on the host machine and mount it to the container as a data volume. All files stored in the data volume will be automatically encrypted. The encrypted file system of this embodiment includes eCryptfs, LUKS, etc. In the container, run encryption tools such as GPG, OpenSSL, etc. to encrypt and decrypt files stored in the data volume. This method requires developers to integrate encryption logic into the application and ensure correct handling of encryption and decryption operations.

[0074] Deep learning model encryption, for ".onnx", ".tflite", ".pth" suffix deep learning model files, use the overall encryption method, based on symmetric encryption algorithm or asymmetric encryption algorithm to encrypt the entire model file. Through the above encryption method, the model file is converted into unreadable binary data, and only in the case of holding the correct key can the model be decrypted and loaded. In addition, in order to increase security, we also perform separate encryption processing on the model parameters, ensuring that even if the model file is accessed illegally, attackers cannot directly obtain the key parameters of the model.

[0075] For model code files with the.py suffix, use source code obfuscation methods and change variable names, function names, delete comments, etc. to make the code logic difficult to understand. In addition, compile Python source code into bytecode and encrypt these bytecode files. Even if the attacker can obtain the.pyc file, it needs to be decrypted to restore it to executable code. For key code segments such as model loading and parameter initialization, dynamic decryption execution is also used to further protect the core logic of the model.

[0076] For deep learning models using containers as carriers, encrypt the entire image file to ensure that even if the image file is stolen, attackers cannot directly run the container. In addition, using the access control function provided by the container platform, we limit access to model files to ensure that only authenticated and authorized users can access and run models in containers.

[0077] Industry model encryption is a crucial task aimed at ensuring the security and integrity of model data. The encryption process needs to consider multiple aspects to build a comprehensive security system.

[0078] First, encrypting the industry model file itself is a key step. By using encryption algorithms such as AES or RSA, model files can be converted into encrypted binary data, ensuring the security of the model during storage and transmission. In this way, even if the model file is illegally obtained, without the correct key, the model cannot be decrypted and used.

[0079] At the same time, encrypting key parameters and weight data in the model can also support the security of the model. These parameters and weights are the core components of the model and have a significant impact on the performance and functionality of the model. By encrypting these sensitive data, unauthorized access and malicious modification can be prevented, ensuring the integrity and accuracy of the model.

[0080] In addition, the inference service engine also needs to be secured. As the manifestation of the industry model, the security of the inference service engine is also important. By implementing access control, input validation, and anomaly detection, etc. security measures, malicious attacks and unauthorized access can be prevented. At the same time, use a secure communication protocol such as HTTPS to protect the communication between the inference service engine and the client, ensuring the confidentiality and integrity of the data during transmission.

[0081] Specifically, set up a secure transmission protocol, use a virtual private network, and host and store the keys during transmission. Complete the integrity verification of the data transmitted, which is: All data transmitted between the client and the target server is encrypted using the HTTPS Hypertext Transfer Protocol; the encrypted model data is encrypted and transmitted using the SFTP Secure File Transfer Protocol and identity verification; A VRN virtual private network is used to create an encrypted channel on a public network, ensuring the security of data when remote users access internal resources of the platform; the VRN virtual private network uses encryption technology to encrypt data transmission, converting data into unreadable form, which is decrypted and read by the recipient through a decryption key; A hardware security module or third-party server is used to store and manage the key; Hash value comparison, digital signature or checksum are used to verify the integrity of the transmitted data; the specific integrity verification method is selected according to the model type, application scenario and data characteristics.

[0082] This embodiment uses HTTP and SFTP to encrypt the transmitted data.

[0083] HTTPS uses SSL / TLS protocol to encrypt all data transmitted between the client and the server, so even if the data is intercepted during transmission, the attacker cannot decrypt and read the content. At the same time, HTTPS also verifies the integrity of the transmitted data, ensuring that the data is not tampered with or damaged during transmission. If the data is modified during transmission, the receiving party detects and rejects the tampered data. Finally, HTTPS is also used to authenticate the server and client. The server proves its identity by providing SSL / TLS certificates, and the client, such as a web browser, verifies the validity of these certificates to ensure communication with the correct server and prevent man-in-the-middle attacks, data theft or tampering.

[0084] SFTP Secure File Transfer Protocol is used in this embodiment to ensure that the content of the transmitted file and authentication information such as username and password are encrypted during transmission, preventing unauthorized personnel from reading or eavesdropping, ensuring the confidentiality of the data. Users need to provide a username and password for identity verification. In addition, SSH key pairs can also be used for identity verification based on secure matching of public and private keys.

[0085] The embodiment selects VPN virtual private network to create an encrypted channel on the public network, ensuring the security of data when remote users access internal resources of the platform. Specifically, various encryption technologies are used to encrypt the transmitted data, converting the data into an unreadable form that can only be decrypted and read by the recipient who has the decryption key. The encrypted channel ensures that data is always transmitted through an encrypted VPN connection from one place to another. By using VPN, remote users can securely access internal resources of the platform on the public network, ensuring that data is not intercepted or tampered with during transmission, protecting the confidentiality and integrity of data. VPN not only encrypts the communication data between the user and the server, but also encrypts the connection data between the user and the VPN server, so even if the data is intercepted during transmission, the attacker cannot decrypt and read its content.

[0086] Key management and storage are key links to ensure the security of encrypted transmission. The key management method of the embodiment is key escrow, which stores the key in a trusted third-party service. The third-party service generates, stores and manages the key, ensuring the security of the key. Since the sender and receiver do not need to handle the key directly, the risk of key leakage is greatly reduced, the key management process is simplified, and the operation difficulty of the sender and receiver is reduced. Because the user does not need to handle the key directly, the security risk caused by human error is reduced. However, the disadvantages of key escrow are also obvious, that is, the reliability of the third-party service is critical. If the third-party service has security vulnerabilities or is attacked, it may lead to a large number of key leaks, causing serious security risks.

[0087] Therefore, the embodiment also uses another method, which is to use a hardware security module HSM. HSM uses multiple security measures such as encrypted storage, access control and physical security to prevent unauthorized access or tampering of the key, and has high physical security. At the same time, the embodiment will also regularly review the key management strategy to ensure the security of the key, and regularly select and adjust the key management and storage strategy according to specific security requirements and scenarios.

[0088] Integrity verification methods primarily involve confirming that data has not been tampered with or damaged during transmission or storage. The integrity verification methods used in this embodiment primarily include hash value comparison, digital signatures, and checksums. The aforementioned hash value comparison, digital signature, and checksum methods can be applied to various types of data and models. However, different model types and application scenarios may require different integrity verification strategies, so this embodiment selects different integrity verification strategies for different model types and application scenarios. For example, for machine learning models, a higher level of security is required, so methods such as digital signatures may be required to ensure the integrity and origin of the data. For some simple statistical models, basic integrity verification using methods such as checksums may be sufficient. Therefore, when selecting an integrity verification method, consider the specific model type, application scenario, and data characteristics, and select the most suitable integrity verification method.

[0089] Specifically, the step S3 is as follows: An encryption algorithm is used to protect sensitive data, including model parameters and weights; a permission management mechanism is established to manage access to model data based on the permission management mechanism, while restricting direct access to sensitive data, recording user access behavior and conducting audits; corresponding backup strategies are set based on the importance of the data, update frequency and storage space; a data recovery mechanism is established to restore model operation based on backup data and the data recovery mechanism when data problems occur, and the integrity and availability of the backup data are regularly verified.

[0090] Data encryption is a key part of model storage security. It involves encrypting model data to prevent it from being illegally obtained and abused during storage and transmission.

[0091] In model storage security, the application of data encryption focuses on protecting sensitive data, such as model parameters and weights. This sensitive data is a core component of the model. Once leaked or tampered with, it can lead to a decline in model performance and even threaten the security of the entire system.

[0092] In order to achieve encryption protection of these sensitive data, encryption algorithms such as symmetric encryption algorithms or asymmetric encryption algorithms can be used, and the storage model can be protected by comparing it with the corresponding model type.

[0093] Access control can manage access rights to model data, preventing unauthorized access and data leakage. In model storage security, the main goal of access control is to ensure that only authorized personnel can access and modify model data. To achieve this goal, a strict permission management mechanism is established, and user identity verification and permission verification are performed. In addition, access control also enhances security by taking other measures, such as limiting direct access to sensitive data, recording user access behavior and conducting audits, etc. These measures help to discover and handle unauthorized access behavior in a timely manner, ensuring the security of model data. Data backup and recovery, as an integral part of the model storage security system, focuses on regularly backing up model data and quickly and accurately recovering when needed, thereby avoiding model service interruptions caused by data loss or damage. The implementation of this link is crucial to ensure the continuous availability of model data and the continuity of business.

[0094] In model storage security, the main goal of data backup and recovery is to ensure that the model can quickly recover normal operation when data is lost, damaged, or tampered with. To achieve this goal, a clear data backup and recovery strategy needs to be developed, and regular backup operations need to be performed.

[0095] Backup strategies should take into account factors such as data importance, update frequency, and storage space. For important model data, regular backups should be used to ensure data timeliness and integrity. At the same time, backup data is stored in a safe and reliable place, such as a dedicated backup server or cloud storage service, to prevent data loss or tampering.

[0096] In terms of data recovery, the embodiment establishes a rapid recovery mechanism to quickly recover the normal operation of the model when data problems occur. This includes regularly verifying the integrity and availability of backup data to ensure that data can be successfully recovered when needed. In addition, emergency plans are developed to quickly respond and handle in the event of an emergency.

[0097] Specifically, the step S4 is specifically: Role-based access control is used to establish a role set for various permissions, and roles in the role set are assigned to different users to control access to specific resources in the system. Virtual machines or virtual machine technology are used to isolate the model running environment from other systems to ensure model running security. Debugging models are run in a sandbox environment to limit model running permissions. Input data during the debugging process is verified and filtered. Regular security vulnerability scans are performed on models. Real-time monitoring of model running logs and call records is performed, and regular security audits of model running logs and call records are performed.

[0098] The security guarantee method of platform model running invocation involves multiple aspects, including access control, security audit, isolation technology, input verification, vulnerability scanning, and monitoring and alarm. By comprehensively using these methods, the security of the model in the running and calling process can be greatly improved, and the system can be protected from potential attacks and threats.

[0099] Access control and permission management are the core components of ensuring the security of platform model running invocation. Through strict access control and identity verification measures, access to model running invocation can be limited, and only authorized users or systems can perform related operations.

[0100] Role-based access control (RBAC) controls who has access to specific resources in the system by establishing a set of roles for various permissions of system operations and then assigning these roles to different users. RBAC is a security policy based on the concept of "roles", which are a collection of permissions. Once a user is assigned an appropriate role, they have all the operational permissions associated with that role. The advantage of this method is that permission management becomes simpler and more centralized, as permissions are assigned to roles rather than directly to users. In this way, when a user's role changes, only their role assignment needs to be adjusted, rather than modifying the permissions of multiple users. In addition, RBAC also supports hierarchical relationships between roles, such as senior roles inheriting the permissions of junior roles, thereby achieving more granular permission control. This mechanism is very useful in large organizations or systems, as it provides an effective, flexible, and easy-to-manage access control solution.

[0101] Attribute-based access control (ABAC) is different from role-based access control (RBAC), as ABAC does not directly assign permissions to roles or users, but rather determines access permissions based on the attributes of entities such as users, resources, and environments. In ABAC, permissions are not predefined, but are dynamically calculated according to actual needs. Therefore, it provides higher flexibility and adaptability, and can meet complex access control requirements. For example, in an enterprise network, employees' access to specific resources can be restricted based on their department, position, and working hours.

[0102] In addition, ABAC also supports the existence of multiple attribute authorities, which are responsible for defining and managing the attributes of entities. When access control decisions need to be made, the policy decision point will obtain relevant attributes from multiple attribute authorities and evaluate whether to grant access permissions according to predefined policies. Attribute-based access control is a powerful and flexible access control method that dynamically calculates access permissions based on entity attributes and context, thereby meeting complex security requirements.

[0103] Isolation and sandboxing techniques are important means to improve the security of platform model execution calls. These techniques create independent and restricted execution environments to ensure that models do not introduce potential security risks when interacting with other systems.

[0104] Isolation techniques such as containerization or virtual machines can isolate the model execution environment from other systems. Containerization technology encapsulates the model and its dependencies through lightweight containers, ensuring that the model runs independently in the container without affecting the security of the host machine or other containers. Virtual machines further enhance security by simulating an entire operating system to isolate the model's execution environment. Sandbox technology is a security mechanism used to restrict model execution permissions. In a sandbox environment, models are limited to specific resources and usage permissions and cannot access or modify critical system resources, reducing the risk of malicious exploitation. Sandbox technology can control model execution permissions by limiting input and output, monitoring model behavior, and restricting network access. In this way, even if there are potential security vulnerabilities in the model, attackers will find it difficult to exploit these vulnerabilities to harm the entire system.

[0105] Input validation and filtering are important aspects of ensuring the security of platform model execution calls. By strictly validating and filtering input data, malicious inputs can be prevented from attacking the model, ensuring the security and stability of the system.

[0106] First, input validation refers to checking input data to ensure it meets expected formats, types, and ranges. This can be achieved through a series of validation rules, such as checking whether input data conforms to specific data formats, such as dates, email addresses, etc., or whether it is within the allowed range. If the input data does not meet the validation rules, the system rejects it and returns an error message.

[0107] Second, input filtering refers to cleaning and processing input data to eliminate potential security risks. This can include removing or escaping special characters, deleting unnecessary white space characters, and limiting input data length. By filtering input data, malicious users can be prevented from injecting attack code or special characters to interfere with the normal operation of the model.

[0108] Security vulnerability scanning and repair are ongoing processes to ensure the security of platform model execution calls. By performing these tasks regularly, organizations can identify and address potential security risks in a timely manner, protecting the integrity of their data and systems.

[0109] Firstly, regular security vulnerability scanning is crucial. This involves using specialized vulnerability scanning tools to check the platform model for potential weaknesses. These tools can automatically detect known vulnerabilities in the model code, configuration, and dependencies, and generate reports for the security team to review. By reviewing these reports, the organization can understand the risks the model may face and take appropriate measures to reduce these risks.

[0110] A security vulnerability scanning tool is a software tool used to automatically detect security vulnerabilities in computer systems, networks, or applications. These tools work by simulating the behavior of an attacker, scanning and testing the target system to find potential security weaknesses.

[0111] Once security vulnerabilities are discovered, timely remediation becomes crucial. Remediation may involve code changes, configuration adjustments, or patch application. In some cases, remediation requires intervention from the model's development or maintenance team. Therefore, maintaining close communication with these teams is key to ensuring timely remediation. Additionally, keeping the software and systems that the model runs on up to date is equally important. Updates to software and systems often contain security patches that fix known security vulnerabilities.

[0112] Security audits and log monitoring are important steps in ensuring the security of platform model run calls. Through these two means, potential security threats and abnormal behaviors can be discovered and addressed in a timely manner.

[0113] Firstly, security audits are the process of evaluating and checking the security of a system. In the context of model run calls, security audits should focus on the model's run logs and call records. These logs record when the model is called, who the caller is, the parameters and results of the call, and other information. By regularly checking these logs, any abnormal behavior or potential security threats can be discovered. For example, if it is found that a certain user frequently calls the model, or the parameters of the call do not match the usual operation, these may be signs of security problems. Security audits can also include reviews of the model itself to ensure that it is designed and implemented in accordance with security standards and that no malicious code or vulnerabilities have been implanted.

[0114] Secondly, log monitoring and alarm handling are key means of real-time detection and response to security incidents. Log monitoring systems can monitor the model's run and call logs in real time, and through pre-set rules and algorithms, identify any unauthorized or abnormal operations. Once abnormal behavior is detected, the system immediately triggers an alarm to notify relevant personnel to respond and handle. The alarm handling mechanism includes steps and processes for quick response, such as notifying the security team to investigate, blocking unauthorized access, rolling back tampered data, etc.

[0115] Specifically, the step S5 is specifically: Access control for model sharing; set up API interface, use API interface for model sharing; package and share the model after packaging, define the input and output interfaces of the model in the packaging process, test and verify after completing the model packaging, deploy the packaged model to the target environment and distribute it to users.

[0116] Access control is a fundamental and critical method to ensure the security of the platform's model sharing process. It involves limiting and controlling user access to model resources, ensuring that only authorized users or organizations can access and use the model. The platform requires users to register and log in before using the platform's model sharing function. By verifying the user's identity, it ensures that only legitimate users can access the model. For high-level or sensitive model access, multi-factor authentication is used, such as mobile phone verification code, fingerprint recognition, etc., to improve the security of the login process.

[0117] At the same time, different roles are defined in the system, such as ordinary users, model administrators, model owners, etc., each role corresponds to different access permissions. According to the user's role and responsibility, assign the corresponding model access permissions. For example, model owners have complete control, while ordinary users only have the right to view and use. The system should record all user model access logs, including access time, accessed models, operation types, etc. Through regular auditing of access logs, any abnormal access behavior can be detected and found, and appropriate security measures can be taken in a timely manner. In addition, clear access control policies should be established, such as IP white list, time limit, etc., to ensure that only users who meet certain conditions can access the model. According to the sensitivity of the model and the behavior of the user, dynamically adjust the user's access permissions, such as limiting access frequency, temporarily locking the account, etc.

[0118] Using API application interface to share platform models is an effective way to ensure the security of model content while allowing users to access model prediction functions through programming. API allows users to send input data and receive model prediction results without exposing the internal structure and source code of the model.

[0119] First, the model needs to be deployed as a service that can receive input and return output. Integrate the model into a backend service that can handle requests from the API. Related services can be deployed on platform servers, cloud servers, or containerized environments.

[0120] When designing API interfaces, ensure that the interface is simple, easy to use, and complies with RESTful design principles. API should provide clear endpoints for receiving input data and returning prediction results. At the same time, consider the format of the data, as well as any necessary parameters, such as model version, input data format, etc.

[0121] APIs should use authentication mechanisms, such as API keys, to verify the identity of the requester. Authorization mechanisms ensure that only authorized users or applications can call the API. To prevent malicious input or attacks, API input should be strictly validated and filtered, including checking data types, size limits, and formats. Additionally, HTTPS should be used to encrypt API requests and responses to prevent data theft or tampering during transmission.

[0122] In terms of API usage, provide detailed API documentation that explains how to use the API, the meaning of parameters, the format of return values, etc. This will help users understand and correctly use the API and reduce the risk of misuse or abuse. Implement API monitoring and logging to track API usage, performance, and any abnormal behavior, and promptly identify and address potential security issues. To prevent API abuse, implement rate limiting and quota strategies. For example, limit the request frequency for each user or application, set a daily request quota, etc. Regularly update model versions, including fixing known security vulnerabilities and updating authentication mechanisms, to maintain the stability and security of the API interface.

[0123] Model encapsulation and packaging is an effective way to protect model content from unauthorized users. By encapsulating the model in a "black box," users are restricted from accessing the model's internal structure and implementation, allowing them to interact with the model only through predefined input and output interfaces.

[0124] First, you need to choose a suitable model packaging tool or platform. These tools or platforms typically provide the ability to convert models into executable files or containerized applications. Popular choices include TensorFlow Serving, ONNX Runtime, and PyInstaller. Before packaging, the model needs to be converted and optimized. This includes converting the model from training frameworks like TensorFlow and PyTorch to a common model format like ONNX, as well as optimizing the model through pruning and quantization to improve inference performance. Using the selected tool or platform, package the converted model into an executable file or containerized application, bundling the model files, dependent libraries, and other necessary resources to ensure that the model can run without the original source code.

[0125] During the packaging process, define the input and output interfaces of the model. These interfaces should be clear and allow users to send input data to the model through standard data formats such as CSV, JSON, image files, etc., and receive the output results of the model. To further improve the security of the model, some additional measures can be taken. For example, encrypt the model file to ensure that even if someone obtains the model file, they cannot directly view the contents of the model. In addition, access permissions to the model can be restricted to allow only authorized users or applications to call the model.

[0126] After completing the model packaging, detailed testing and verification are extremely important to ensure that the model can run correctly in the new packaging environment, check the function and performance of the input and output interfaces. Finally, deploy the packaged model to the target environment and distribute it to users. The model can be deployed to a platform server, cloud platform or containerized environment so that users can access the model through the network.

[0127] Through model packaging and packaging, the security and privacy of the model content can be effectively protected, while allowing users to interact with the model through simple input and output interfaces. This is very effective for sharing models without exposing the internal structure or implementation details of the model.

[0128] The model network threat protection method mainly involves ensuring the security of the model in the network environment, preventing network attacks and malicious intrusions. Building a firewall and intrusion detection system, and developing common network attack defense strategies are of great significance to protect the model from network risk threats.

[0129] Firewall and intrusion detection system IDS / IPS are key tools for resisting network risk threats and protecting the security of the model. By deploying a firewall and intrusion detection system, a multi-level security protection system is built to effectively monitor network traffic and prevent malicious access and attacks.

[0130] The firewall is the first line of defense for protecting the model's security. It can deploy the model in an isolated network environment, limiting access to the model to authorized users only. The firewall can inspect all traffic entering and leaving the network and filter it according to pre-set security rules, ensuring that only data packets that meet the rules can pass through, effectively preventing unauthorized access and known malicious traffic, providing basic protection for the model.

[0131] However, relying solely on a firewall is not enough to deal with all threats. To more comprehensively protect the model, the assistance of an intrusion detection system (IDS) and an intrusion prevention system (IPS) is also needed. IDS can monitor network traffic in real time, using pattern matching, statistical analysis and other methods to detect abnormal behavior and potential malicious activities, identify known attack patterns, malware, vulnerability exploits and other threats, and issue alerts in a timely manner, allowing administrators to respond quickly.

[0132] In comparison to IDS, IPS not only has the ability to detect threats, but also actively takes measures to prevent or mitigate attacks. It can discard malicious packets, reset connections, block attacker IPs, etc. when detecting malicious traffic, thus more actively protecting the model's security. Many IPS devices also integrate firewall functions, capable of performing traffic filtering and network isolation, providing more comprehensive security protection.

[0133] When deploying these systems, appropriate hardware and software need to be selected to ensure they can adapt to the model's network architecture and traffic requirements. At the same time, it is crucial to keep firewalls and IDS / IPS updated and upgraded to address new threats and vulnerabilities. In addition, it is necessary to ensure that these systems can seamlessly integrate and work with existing network management systems to achieve unified security management.

[0134] By deploying firewalls and intrusion detection and prevention systems IDS / IPS, the model's security can be greatly improved, real-time monitoring of network traffic, identifying abnormal behavior, and taking timely measures to prevent potential threats.

[0135] Embodiment Two The embodiment provides a nuclear power industry model full-cycle asset security device, comprising a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to realize the nuclear power industry model full-cycle asset security method in the embodiment one.

[0136] The nuclear power industry model full-cycle asset security device provided in the embodiment is used to realize the nuclear power industry model full-cycle asset security method, so the nuclear power industry model full-cycle asset security device also has the technical effects of the nuclear power industry model full-cycle asset security method, which will not be described here.

[0137] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the present application.

Claims

1. A full-cycle asset security assurance method for a nuclear power industry model, characterized by, The method comprises the following steps: Step S1, in the model generation stage, the data set relied on by the model generation is preprocessed, the model parameters are optimized, and the corresponding defense strategy is set for the data poisoning; Step S2, in the model transmission stage, a secure transmission protocol is set, a virtual private network is used, the model is encrypted and transmitted and uploaded, the key in the transmission process is hosted and stored, and the integrity of the model data after transmission is verified; Step S3, in the model storage stage, the stored model data is encrypted and access control is performed, the corresponding data backup strategy and data recovery strategy are set for the stored model data; Step S4, in the model running calling stage, the model is run and called by using isolation and sandbox technology, and the running and calling of the model is controlled by access control; Step S5, in the model sharing stage, the access control of model sharing is controlled, the corresponding application program interface is set for model sharing, and the model is packaged and packaged before sharing; a firewall and an intrusion detection defense system are deployed to defend against network threats of the model.

2. The full-cycle asset security assurance method for a nuclear power industry model of claim 1, wherein, The data relied on by the model in the step S1 is preprocessed, specifically: The data set is cleaned, converted, preprocessed and verified; The data cleaning includes missing value processing and outlier processing, the outlier processing includes filtering the data, identifying and proposing outliers by using statistical methods, the missing value processing includes deleting, interpolating and machine learning methods to process missing values; the data preprocessing includes data conversion and feature engineering, the data conversion includes data standardization and feature encoding, the feature encoding is realized by label encoding and one-hot encoding, the feature engineering includes feature construction and feature selection, the feature construction uses numerical features, text features or interaction features to construct, and the feature selection selects key features through correlation analysis and recursive feature elimination; the data verification is to divide the data set into training set, validation set and test set by using cross-validation, to generate multiple different data sets by using repeated sampling for random sampling, and to detect the consistency of the data set by using consistency test.

3. The full-cycle asset security method for nuclear power industry models according to claim 1, characterized in that, The model parameters are optimized in the step S1, specifically: Automatic parameter tuning technology is used to automatically explore the parameter space, the model performance evaluation index is used as the objective function, the data set is used for model training and verification, the optimal parameters are obtained, and the optimization of the model parameters is completed.

4. The full-cycle asset protection method for nuclear power industry models according to claim 1, characterized in that, In the step S1, the corresponding defense strategy is set for the data poisoning, specifically: The differential privacy technology is used to protect the privacy of the data set; in the model training process, disturbance or noise is introduced to enhance the resistance of the model to malicious input; an access control mechanism is set for the data set to limit unauthorized users from accessing or modifying the data set; at the same time, the real-time log records in the model generation process are audited.

5. The full-cycle asset protection method for nuclear power industry models according to claim 1, characterized in that, In the step S2, the model is encrypted and transmitted and uploaded, specifically: Different encryption methods are used for encryption for different types of models and different contents therein; The model includes an industrial mechanism model and an industrial artificial intelligence model; The industrial mechanism model comprises a basic theory model, a process model and a simulation model; The encryption mode of the basic theory model is that the basic theory model is encrypted by using an encryption algorithm, key data in the basic theory model is encrypted by using AES or RSA, a specific encryption algorithm is set for specific data in the basic theory model, a code obfuscation technology is used to change variable names, function names and logical structures of the basic theory model, the basic theory model is split into multiple modules by using a model splitting technology, and the completed basic theory model is run by combining the modules; The encryption mode of the process model is that the process model is encrypted by using a symmetric encryption algorithm, key information of the process model is encrypted by using an asymmetric encryption algorithm, a special encryption tool is used to encrypt a configuration file of the process model, an encryption step is added in a model generation or calling process for a parameter model executable by a solver, encryption is realized, an encryption mechanism is added in an encapsulation process for an FMU model encapsulated according to a standard specification, and the FMU model is encrypted based on the encryption mechanism; The encryption mode of the simulation model is that a file encryption tool is used to encrypt a simulation model file, source code files are encrypted before compilation for model code files, and binary encryption is adopted for binary simulation model files to encrypt machine code of the binary files; The industrial artificial intelligence model comprises a machine learning model, a deep learning model and an industry large model; The encryption mode of the machine learning model is that a symmetric encryption algorithm or an asymmetric encryption algorithm is used to encrypt a machine learning model file, source code is obfuscated to realize encryption for model code files, a special encryption tool is used to encrypt binary model files of the machine learning model, and a corresponding encrypted file system is created on a host computer for a machine learning model in a container as a data volume, and files stored in the data volume are automatically encrypted; The encryption mode of the deep learning model is that a symmetric encryption algorithm or an asymmetric encryption algorithm is used to encrypt a deep learning model file by using an overall encryption mode, a source code obfuscation method is used to realize encryption for model code files, source code is compiled into bytecode, the bytecode file is encrypted, and dynamic decryption execution is adopted for key code segments, and an entire image file is encrypted for a deep learning model in a container, and access control functions provided by a container platform are used to control access permissions of model files of the deep learning model. The encryption method of the industry large model is specifically: using an encryption algorithm to convert the model file of the industry large model into encrypted binary data; encrypting the key parameters and weight data of the industry large model; at the same time, implementing access control, input verification and anomaly detection on the inference service engine, and using a secure communication protocol to realize communication between the inference server and the client.

6. The full-cycle asset protection method for nuclear power industry models according to claim 1, characterized in that, A secure transmission protocol is set, a virtual private network is used, and the key during transmission is hosted and stored, and the integrity of the data transmitted is verified, specifically: All data transmitted between the client and the target server is encrypted using the HTTPS Hypertext Transfer Protocol; the encrypted model data is encrypted and transmitted using the SFTP secure file transfer protocol and identity verification; A VRN virtual private network is used to create an encrypted channel on a public network to ensure the security of data when remote users access internal resources of the platform; the VRN virtual private network uses encryption technology to encrypt transmitted data, converting the data into an unreadable form, which is decrypted and read by the recipient through a decryption key; A hardware security module or third-party server is used to store and manage the key; The integrity of the transmitted data is verified using a hash value comparison, digital signature or checksum; the specific integrity verification method is selected according to the model type, application scenario and data characteristics.

7. The full-cycle asset protection method for nuclear power industry models according to claim 1, characterized in that, The step S3 is specifically: Sensitive data including model parameters and weights are protected using an encryption algorithm; a permission management mechanism is established to manage access to model data based on the permission management mechanism, while limiting direct access to sensitive data, recording user access behavior and conducting audits; based on the importance, update frequency and storage space of the data, set the corresponding backup strategy; A data recovery mechanism is established, and when the data is problematic, the model is recovered based on the backup data and the data recovery mechanism, and the integrity and availability of the backup data are verified periodically.

8. The full-cycle asset protection method for nuclear power industry models according to claim 1, characterized in that, The step S4 is specifically: Role-based access control is used to establish a role set for various permissions, assign roles in the role set to different users, and thus control the permission to access specific resources in the system; a virtual machine or virtual machine technology is used to isolate the model running environment from other systems to ensure model running security; debug the model in a sandbox environment to limit model running permissions; input validation and input filtering are performed on input data during the running and debugging process; security vulnerabilities are scanned periodically; The running logs and call records of the model are monitored in real time, and the running logs and call records of the model are audited periodically.

9. The full-cycle asset protection method for nuclear power industry models according to claim 1, characterized in that, The step S5 is specifically: Access permission control is performed on model sharing; an API interface is set up, and the API interface is used for model sharing; the model is encapsulated and packaged before model sharing, and the input and output interfaces of the model are defined during encapsulation; after encapsulation, the encapsulated model is deployed to the target environment and distributed to users.

10. A full-cycle asset security assurance device for a nuclear power industry model, characterized by, The computer program product comprises a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to implement the full-cycle asset security guarantee method of the nuclear power industry model according to any one of claims 1-9.

Images