Data security protection method and system for large model training

Abstract

The invention discloses a data security protection method and system for large model training. The core of the method is that a data safe box trusted application (TA application) is deployed in a trusted execution environment (TEE) and cooperatively works with a user-defined virtual file system driver at a host machine side, so that an end-to-end secret state data access control link is constructed. Before encryption protection of sensitive data, a measurement value (including code hash, configuration parameters and a runtime state) of a current TEE instance is submitted to a remote verification service through a data safe box TA application, and the verification service only sends the measurement value (including code hash, configuration parameters and a runtime state) of the current TEE instance to the remote verification service after confirming that the TEE environment meets a preset security policy (such as a kernel version, a software version and configuration parameter compliance). And generating a secret key encrypted by a custom virtual file system on a host machine side through derivation. And malicious copying of a system memory and a storage medium is prevented, so that the traditional attack defense capability is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large model training data security, and more particularly, to a data security protection method and system for large model training. BACKGROUND

[0002] With the rapid development of artificial intelligence technology in recent years, large models LLMs have been widely used in sensitive scenarios such as intelligent manufacturing, smart home, enterprise services, etc. However, in the whole life cycle of large model training, fine-tuning, reasoning and deployment, the security of data faces unprecedented challenges. In the application of large models, the risks not only involve the leakage of user privacy and commercial secret data, but also involve data sovereignty and AI intellectual property. These challenges involve every link from the collection of raw data to the output of model services, showing the characteristics of long chain, wide attack surface and blurred trust boundary.

[0003] Specifically, the security risks of large model training are reflected in two aspects. On the one hand, large model training needs to collect personal privacy data and commercial secret data, such as employee personal information and enterprise documents. These data are long-term resident in memory or temporary storage in plaintext form during preprocessing, word segmentation, etc., which are easy targets for attacks. On the other hand, traditional training relies on operating systems, container runtimes, scheduling frameworks, etc. If any component is compromised, the entire training environment will be compromised. Traditional security management solutions lack hardware-level isolation to contain such attacks moving horizontally, and lack effective monitoring and control of the risk of illegal data theft by internal operation and maintenance personnel. SUMMARY

[0004] The present application overcomes the problem of insufficient protection of sensitive data in large data training in the prior art, and provides a data security protection method and system for large model training, in order to solve the problems existing in the prior art.

[0005] To solve the above technical problems, one aspect of the present application provides a data security protection method for large model training:

[0006] A data security protection method for large model training, comprising the following steps:

[0007] After starting the training, a data safe of the model training is created, and the configuration information of the data safe is issued to a trusted application in a trusted execution environment;

[0008] When a user accesses the data safe, the data safe compares and verifies the user information and access information through the trusted application;

[0009] If the verification is successful, an encrypted file directory and a corresponding plaintext path are initialized and created;

[0010] If not successful, the subsequent steps are not performed;

[0011] The trusted application in the trusted execution environment performs key derivation random generation;

[0012] The generated key, application and environment information are stored in an encrypted file directory;

[0013] When the data needs to be used, the data vault generates a storage encryption key through the trusted application, and processes the data that needs to be used in the data vault.

[0014] Further technical solutions are that when the data needs to be used, the data vault generates a storage encryption key through the trusted application, and processes the data that needs to be used in the data vault, which specifically includes the following steps:

[0015] When the user writes data to the data vault, the trusted application of the data vault generates a storage master encryption key, and performs transparent encryption on the data written to the data vault through the storage master encryption key. The encrypted data ciphertext is stored in the data vault;

[0016] When the user reads data from the data vault, the data vault obtains the storage master encryption key, decrypts and restores the data through the storage master encryption key, and stores the restored data in the data vault.

[0017] Further technical solutions are that the storage encryption key generation step is that the data vault trusted application generates a storage encryption key according to the key management service, according to the data vault ID, the vault configuration information, the random number and the salt value.

[0018] Further technical solutions are that when the user logs out of the data vault after the data use is completed, the following steps are performed:

[0019] The data vault background service unregisters the file directory mounted by the data vault, and the original data is stored in the file system.

[0020] Further technical solutions are that before starting the training, the following steps are performed:

[0021] The trusted execution environment is securely started;

[0022] The data vault calls the CPU security instruction provided by the trusted execution environment system to generate a corresponding HASH value, and signs the data through the private key of the data vault to generate an encrypted proof report;

[0023] Send the attestation report generated by the data vault to the sensitive data management platform, and the sensitive data management platform verifies the signature according to the system pre-defined protocol, and verifies the trustworthiness of the server environment;

[0024] After receiving the attestation report of the server, the data vault verifies the success, starts to scan the information of the current system, generates a unique digital identity of the device according to the algorithm, and reports the information of the current system to the sensitive data management platform.

[0025] Further, the specific steps of creating a physical machine server resource are:

[0026] Based on the platform, the system is securely started by the initialization code, configuration and initial measurement value of the trusted execution environment.

[0027] The above steps are in a trusted state from the hardware power-on, preventing the trusted execution environment from being tampered with or malicious code injected by the firmware, bootloader or virtual machine monitor;

[0028] Further, the specific steps of calling the CPU security instruction provided by the trusted execution environment system to generate the corresponding HASH value are:

[0029] Based on the measurement value of the trusted data vault, the corresponding HASH value is generated according to the code of the data vault, the configuration file, the current system version and environment information, and the random number attribute.

[0030] Another aspect of the present application also provides a data security protection system for large model training, comprising:

[0031] The data vault is used to store model data and interact with the trusted application of the trusted execution environment to verify user access behavior, and complete interaction with the user after verification;

[0032] The trusted execution environment is used to install a trusted application program, which is used to interact with the data vault, store configuration information of the data vault, verify user access behavior, randomly generate a key derivation, and generate a storage encryption key;

[0033] Further, the data security protection system for large model training further comprises a sensitive data management platform.

[0034] The sensitive data management platform is used to receive the encrypted attestation report of the data vault, verify the signature, verify the trustworthiness of the server environment, and receive the information of the current system.

[0035] Compared with the prior art, the present application has at least the following beneficial effects: the present application provides a sensitive data protection method for large model training and inference scenarios, the core of which is to deploy a data safe TA application inside a trusted execution environment (TEE) and work with a custom virtual file system driver on the host side to build an end-to-end encrypted data access control link. Before the sensitive data is encrypted and protected, the data safe TA application submits the measurement value (including code hash, configuration parameter and runtime state) of the current TEE instance to the remote verification service, and the verification service only generates a custom virtual file system encryption key on the host side after confirming that the TEE environment meets the preset security policy (such as kernel version, software version, configuration parameter compliance). When the user accesses and writes sensitive data, the data safe TA application dynamically verifies and authorizes the management platform according to the current user information, software information and access path. The encryption and decryption processes of the sensitive data are both performed in the TEE, ensuring that the memory isolation cannot be stolen, and the original data is encrypted and protected in the storage area, so that illegal users cannot obtain the plaintext data. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The schematic flowchart of the embodiment one is shown.

[0037] Figure 2 The schematic flowchart of the embodiment one is shown.

[0038] Figure 3 The structural schematic diagram of the embodiment two is shown. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0040] For subsequent abbreviations, the following explanations are provided:

[0041] LLMs: Large Language Models

[0042] FUSE: Filesystem in Userspace

[0043] TEE:Trusted Execution Environment

[0044] SGX: Software Guard Extensions A set of CPU hardware security extension instructions introduced by Intel

[0045] SEV-SNP: Secure Encrypted Virtualization - Secure Nested Paging A hardware-level virtual machine memory protection and integrity verification technology introduced by AMD in its EPYC server processors

[0046] TrustZone: A system-level hardware security technology developed by ARM

[0047] TA: Trusted Application is a trusted application running in the secure world inside the TEE.

[0048] Embodiment one

[0049] A data security protection method for large model training, see Figure 1 、 Figure 2 , comprising the following steps:

[0050] When the user creates a physical machine server resource before large model training and deployment, that is, the initial stage of the system. It can be understood that the initial stage of the system is not the protection method itself, and after the first setup is completed, there is no need to repeat the preparation steps.

[0051] S01: Trusted execution environment secure boot;

[0052] When the system starts, the secure boot program in the trusted execution environment verifies the system startup boot code, system configuration and initial measurement value, and the system can run normally after successful verification.

[0053] In this embodiment, based on Intel SGX / TDX, AMD SEV-SNP or ARM TrustZone platform, the boot program of the trusted execution environment (TEE environment) verifies the startup code, system configuration and initial measurement value when the system is securely started. From the hardware power-on, it is in a trusted state, preventing firmware, bootloader or virtual machine monitor from tampering with or injecting malicious code into the TEE environment;

[0054] S02: After the data safe box runs, call the CPU security instruction provided by the trusted execution environment system, generate the corresponding HASH value, and sign the data through the private key of the data safe box to generate an encrypted proof report;

[0055] Based on the measurement value of the trusted data vault, a corresponding HASH value is generated according to the code of the data vault, the configuration file, the current system version and environment information, and a random number attribute.

[0056] In this embodiment, the trusted data vault in the TEE calls the CPU security instruction provided by the TEE system after the system runs, generates a corresponding HASH value based on the measurement value of the trusted data vault according to the code of the data vault, the configuration file, the current system version and environment information, and a random number attribute, and signs the data by using the private key of the data vault to generate an encrypted proof report.

[0057] S03: The proof report generated by the data vault is sent to the sensitive data management platform, the sensitive data management platform verifies the signature according to the protocol defined by the system in advance, and verifies the trustworthiness of the server environment.

[0058] S04: The data vault receives the proof report of the server, verifies the success, starts to scan the information of the current system, generates a unique digital identity of the device according to an algorithm, and reports the information of the current system to the sensitive data management platform.

[0059] In this embodiment, the system information specifically includes the version of the current general system, the application program, the IP address, the MAC address, the unique DID of the device generated according to an algorithm, and the system information reported to the sensitive data management platform.

[0060] After the above steps are completed, the physical machine server resource creation is completed, and the related training steps can be started.

[0061] S1: After starting the training, a data vault for model training is created, and the configuration information of the data vault is distributed to the trusted application in the trusted execution environment;

[0062] In this embodiment, a user creates a data vault for model training in the sensitive data management platform, needs to configure the storage file path of the model, the program for accessing data, the user access permission, and the configuration of other information, and the sensitive data management platform distributes the configuration information of the data vault to the data vault trusted application (TA application) in the TEE environment.

[0063] S2: When the user accesses the data vault, the data vault compares and verifies the user information and the access information by using the TA application;

[0064] If the verification is successful, an encrypted file directory and a corresponding plaintext path are initialized and created;

[0065] If the verification is not successful, the subsequent steps are not performed.

[0066] S3: The TA application in the trusted execution environment performs key derivation random generation;

[0067] S4: The generated key, application, and environment information are encrypted and stored in an encrypted file directory;

[0068] The above steps specifically involve calling a FUSE file system initialization to create an encrypted file directory and a corresponding plaintext path according to a configuration file;

[0069] The TA application used for initializing the FUSE file system password is randomly generated by key derivation through a data vault TA application in the TEE, and the application and environment information are encrypted and stored in the directory configuration created by the FUSE file system;

[0070] S5: When data needs to be used, the data vault generates a storage encryption key through a trusted application, and processes the data that needs to be used in the data vault.

[0071] Processing the data that needs to be used specifically includes the following steps:

[0072] When a user writes data to the data vault, the trusted application of the data vault generates a storage encryption key DEK, and the data written to the data vault is transparently encrypted through the storage encryption key DEK and IV. The encrypted data ciphertext is stored in the data vault;

[0073] The storage encryption key generation step is that the data vault trusted application generates a storage encryption key according to the key management service, according to the data vault ID, vault configuration information, random number, and salt value.

[0074] When a user reads data from the data vault, the data vault obtains the storage master encryption key from the data vault TA application in the TEE environment. The data vault TA application obtains the previously generated storage master encryption key according to the data vault ID according to the key management service, and decrypts and restores the data through the storage master encryption key. The restored data is stored in the data vault.

[0075] S6: When the user logs out of the data vault, the data vault background service unmounts the file directory mounted by the data vault, and the original data is encrypted and stored in the file system.

[0076] Through the above steps of verification and protection, when an illegal user accesses, only ciphertext data can be obtained.

[0077] Next, the actual operation steps are further explained in detail:

[0078] Step 1: Remote attestation challenge sending

[0079] Data Locker TA generates a random number RandomA, and performs HASH calculation and signature on the system version, software version, configuration information, and RandomA through the data locker private key Sk to prevent replay attacks.

[0080] Step 2: Remote attestation request generation

[0081] The data locker TA calculates the hash of the current data locker program running code, configuration parameters, system version, and software version for measurement.

[0082] The data locker private key Sk is used to perform SHA256 algorithm HASH calculation and ECC algorithm signature on the system version, software version, configuration information, and RandomA to prevent replay attacks.

[0083] The measurement information + signature information + random number RandomA are sent to the sensitive data management platform for remote attestation.

[0084] Step 3: Verification and decision

[0085] The sensitive data management platform verifies the certificate chain, performs signature verification on the remote attestation request through the data locker public key Pk, and checks the measurement value, and compares the platform configured system version, software version, data locker code HASH value, and configuration parameter information to determine whether they meet the expected value.

[0086] If not successful, the subsequent steps are not executed.

[0087] If the verification is successful, an attestation result is generated, and a structured result in JWT format is output, including the remote attestation result + random number RandomB + remote attestation signature value.

[0088] Step 4: Key derivation

[0089] After successful remote attestation, the data locker derives the encryption key DEK through the key derivation algorithm HMAC-SHA256 algorithm using the master key MasterKey and the salt value Salt.

[0090] DEK = KDF(MasterKey / Password, Salt)

[0091] Step 5: Generate storage encryption key through trusted application and calculate IV

[0092] When the user performs data storage transparent encryption, the SHA256 algorithm is called to perform digest calculation on the data safe configuration parameters ConfigFile and RandomNumberC to obtain HashConfig and HashRandom, and the calculation results are spliced and 12-byte IV is intercepted

[0093] HashConfig = H(ConfigFile)

[0094] HashRandom = H(RandomNumberC)

[0095] IV = Truncate12(HashConfig || HashRandom)

[0096] In the data safe, the data to be used is processed;

[0097] The generated data encryption key DEK and the calculated IV are used, and the user space Fuse transparent encryption is performed according to the AES-256-GCM mode to encrypt the actual data. When the user reads the data, the encrypted data is decrypted according to the AES-256-GCM mode through the encryption key DEK and the calculated IV in the TEE.

[0098] (Ciphertext, Tag) = AES-256-GCM(DEK, IV, Plaintext, AAD)

[0099] The above technical solution first verifies the trustworthiness of the system environment in the TEE trusted execution environment through the data safe TA, then generates a data encryption key DEK using the TEE cryptographic algorithm library, and then combines the hash value of the system environment configuration information and the hash value of the random number to form an IV. The Fuse file system transparent encryption technology encrypts and protects the model and data. It can prevent malicious copying of system memory and storage media, thereby effectively improving the defense against traditional attack capabilities.

[0100] Embodiment two

[0101] A data security protection system for large model training, referring to Figure 3 , comprising:

[0102] A data safe is used to store model data and interact with a trusted application of a trusted execution environment to verify user access behavior, and complete interaction with the user after verification is completed;

[0103] A trusted execution environment for installing a trusted application for interacting with the data vault, storing configuration information of the data vault, verifying user access behavior, key derivation random generation, generating storage encryption keys.

[0104] A sensitive data management platform for receiving an encrypted attestation report of the data vault, verifying the signature, and verifying the trust of the server environment, receiving information of the current system.

[0105] While the application has been described herein with reference to the exemplary embodiments thereof, it is to be understood that the application is capable of many variations and modifications in light of the foregoing teaching. For example, the subject combination layout can be modified in many ways, and / or the layout can be modified. Other uses will be apparent to those of ordinary skill in the art.

Claims

1. A data security protection method for training large models, characterized in that, Includes the following steps: After training begins, a data vault for model training is created, and the configuration information of the data vault is distributed to trusted applications in the trusted execution environment. When a user accesses the data vault, the data vault compares and verifies the user information and access information through a trusted application. If the verification is successful, an encrypted file directory and its corresponding plaintext path will be created during initialization. If unsuccessful, do not proceed with the next steps; The trusted application execution key in a trusted execution environment is derived and randomly generated. The generated key, application, and environment information are encrypted and saved in an encrypted file directory; When data is needed, the data vault generates a storage encryption key through a trusted application, and then processes the data within the data vault.

2. The data security protection method for large model training as described in claim 1, characterized in that, When data is needed, the data vault generates a storage encryption key through a trusted application. The data vault then processes the data to be used, specifically including the following steps: When a user writes data to the data vault, the trusted application of the data vault generates a storage master encryption key, which is used to transparently encrypt the data written to the data vault. The encrypted data is then stored in the data vault. When a user reads data from the data vault, the data vault obtains the master encryption key and uses it to decrypt and restore the data. The restored data is then stored in the data vault.

3. The data security protection method for large model training as described in claim 2, characterized in that, The steps for generating the storage encryption key are as follows: The trusted application of the data vault generates the storage encryption key by performing key derivation based on the key management service, the data vault ID, vault configuration information, random number, and salt value.

4. A data security protection method for large model training as described in any one of claims 1-3, characterized in that, When a user logs out of the data safe after using the data, perform the following steps: The Data Vault backend service unmounts the file directory that the Data Vault is mounted on, and the original data is encrypted and stored in the file system.

5. The data security protection method for large model training as described in claim 1, characterized in that, Before starting training, perform the following steps: Secure boot of the Trusted Execution Environment; The data vault invokes CPU security instructions provided by the trusted execution environment system to generate a corresponding hash value, and signs the data using the data vault's private key to generate an encrypted proof report; The certificate report generated by the data vault is sent to the sensitive data management platform, which verifies the signature according to the predefined protocol of the system and verifies the trustworthiness of the server environment; Once the data vault receives the verification report from the server and successfully verifies it, it begins scanning the current system's information, generates a unique digital identity for the device based on the algorithm, and then reports the current system's information to the sensitive data management platform.

6. The data security protection method for large model training as described in claim 5, characterized in that, The specific steps for creating physical machine server resources are as follows: Based on the platform, secure system startup consists of initialization code, configuration, and initial metrics from a trusted execution environment.

7. The data security protection method for large model training as described in claim 5, characterized in that, The specific steps for invoking the CPU security instructions provided by the trusted execution environment system to generate the corresponding HASH value are as follows: Based on the metrics of the trusted data vault, a corresponding hash value is generated according to the data vault's code, configuration file, current system version and environment information, and random number attributes.

8. A data security protection system for training large models, characterized in that, include: The data safe is used to store model data and, through interaction with trusted applications in a trusted execution environment, to verify user access behavior and complete the interaction with the user after verification. A trusted execution environment is used to install trusted applications, which interact with the data vault, store the data vault's configuration information, verify user access behavior, generate random keys, and generate and store encrypted keys.

9. A data security protection system for large model training as described in claim 8, characterized in that, It also includes a sensitive data management platform; The sensitive data management platform is used to receive encrypted proof reports from the data vault, verify signatures, verify the trustworthiness of the server environment, and receive information from the current system.

Images