Model training method and device and storage medium

Abstract

The invention provides a model training method and device and a storage medium, relates to the technical field of computers, and can protect data privacy of model training. The method comprises the following steps: receiving request information, wherein the request information is used for requesting instantiation of virtual resources required by model training; based on the request information, virtual resources are created on a target multi-access edge computing (MEC) host, an MEC application program is installed in the virtual resources, the MEC application program is used for model training, and the target MEC host has a trusted execution environment (TEE) capability; and configuring the MEC application program, and performing model training through the configured MEC application program.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a model training method, device and storage medium. Background Art

[0002] As artificial intelligence (AI) big model technology matures, AI big models are gradually being applied to multiple fields such as customer service, content creation and media, medical care, manufacturing, and transportation. However, since the training data or inference data of AI big models may involve highly sensitive data, it is necessary to use multi-access edge computing (MEC) to deploy and fine-tune AI big models at or near the data source, thereby controlling the area where data is used to reduce the risk of data leakage.

[0003] However, MEC is deployed at the edge of the network in a complex environment, and the infrastructure and applications belong to different owners. As a result, there is still a risk of data leakage when fine-tuning and training large AI models in MEC. Summary of the invention

[0004] The present application provides a model training method, device and storage medium, which can protect the data privacy of model training.

[0005] In order to achieve the above objectives, this application adopts the following technical solutions:

[0006] In a first aspect, the present application provides a model training method, which includes: receiving request information, the request information is used to request the instantiation of virtual resources required for model training; based on the request information, creating virtual resources on a target MEC host, and installing a MEC application in the virtual resources, the MEC application is used for model training, and the target MEC host has a trusted execution environment TEE capability; configuring the MEC application, and performing model training through the configured MEC application.

[0007] In a possible implementation, based on the request information, virtual resources are created on the target MEC host, and the MEC application is installed in the virtual resources, including: determining the virtualization infrastructure manager VIM that manages the target MEC host according to the request information; creating a virtual container or a virtual machine on the target MEC host through the VIM, and installing the MEC application in the virtual container or the virtual machine.

[0008] In one possible implementation, the MEC platform is configured with multiple MEC hosts, and the method further includes: determining the requirement information for model training based on the request information, the requirement information including the virtual resources required for the model training and the required TEE capabilities of the MEC host; and determining a target MEC host that meets the requirement information from among the multiple MEC hosts.

[0009] In a possible implementation, configuring the MEC application includes: configuring the MEC application when the operating environment of the MEC application passes verification.

[0010] In one possible implementation, the method also includes: sending remote certification information of the operating environment of the MEC application to the verifier, so that the verifier performs security verification on the operating environment of the MEC application; receiving response information from the verifier, and the verification response information is used to indicate whether the operating environment of the MEC application has passed the verification.

[0011] In a possible implementation, model training is performed through a configured MEC application, including: obtaining model training parameters; wherein the model training parameters include: model parameters, model training data, model training strategies, and model verification parameters; and model training is performed in the configured MEC application according to the model training parameters.

[0012] In the second aspect, the present application provides a model training device, which includes: a communication unit and a processing unit; the communication unit is used to receive request information, the request information is used to request the instantiation of virtual resources required for model training; the processing unit is used to create virtual resources on the target MEC host based on the request information, and install MEC applications in the virtual resources, the MEC applications are used for model training, and the target MEC host has a trusted execution environment TEE capability; the processing unit is also used to configure the MEC application and perform model training through the configured MEC application.

[0013] In a possible implementation, the processing unit is further used to determine the virtualization infrastructure manager VIM that manages the target MEC host according to the request information; the processing unit is further used to create a virtual container or a virtual machine on the target MEC host through the VIM, and install the MEC application in the virtual container or the virtual machine.

[0014] In one possible implementation, the processing unit is further used to determine the requirement information for model training based on the request information, where the requirement information includes the virtual resources required for the model training and the required TEE capabilities of the MEC host; the processing unit is further used to determine a target MEC host that meets the requirement information from multiple MEC hosts.

[0015] In a possible implementation, when the operating environment of the MEC application is verified, the processing unit is also used to configure the MEC application.

[0016] In one possible implementation, the communication unit is also used to send remote certification information of the operating environment of the MEC application to the verifier, so that the verifier can perform security verification on the operating environment of the MEC application; the communication unit is also used to receive response information from the verifier, and the verification response information is used to indicate whether the operating environment of the MEC application has passed the verification.

[0017] In a possible implementation, the communication unit is also used to obtain model training parameters; wherein the model training parameters include: model parameters, model training data, model training strategies, and model verification parameters; the processing unit is also used to perform model training in the configured MEC application according to the model training parameters.

[0018] In a third aspect, the present application provides a model training device, comprising: a processor and a communication interface; the communication interface and the processor are coupled, and the processor is used to run a computer program or instructions to implement the model training method described in the first aspect and any possible implementation method of the first aspect.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium, which stores instructions. When the instructions are executed on a terminal, the terminal executes the model training method described in the first aspect and any possible implementation of the first aspect.

[0020] In a fifth aspect, the present application provides a computer program product comprising instructions, which, when the computer program product is run on a model training device, enables the model training device to perform the model training method described in the first aspect and any possible implementation of the first aspect.

[0021] In a sixth aspect, the present application provides a chip, comprising a processor and a communication interface, wherein the communication interface and the processor are coupled, and the processor is used to run a computer program or instructions to implement the model training method described in the first aspect and any possible implementation method of the first aspect.

[0022] Specifically, the chip provided in the present application also includes a memory for storing computer programs or instructions.

[0023] The above technical solution brings at least the following beneficial effects: based on the received request information, a virtual resource is created on the target MEC host, a MEC application is installed in the virtual resource, the MEC application is configured, and the model is trained through the configured MEC application. In other words, the model training method provided in the embodiment of the present application is to create a virtual resource on a host with a trusted execution environment capability and install the MEC application in the virtual resource, that is, the MEC application is in a trusted execution environment, thereby being able to protect the data privacy of the model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of a MEC system architecture provided in an embodiment of the present application;

[0025] Figure 2 A schematic diagram of the structure of a model training system provided in an embodiment of the present application;

[0026] Figure 3 A schematic diagram of the structure of another model training system provided in an embodiment of the present application;

[0027] Figure 4 A schematic diagram of the composition of a model training device provided in an embodiment of the present application;

[0028] Figure 5 A flowchart of a model training method provided in an embodiment of the present application;

[0029] Figure 6 A flowchart of another model training method provided in an embodiment of the present application;

[0030] Figure 7 A flowchart of another model training method provided in an embodiment of the present application;

[0031] Figure 8 A flowchart of another model training method provided in an embodiment of the present application;

[0032] Fig. 9 A schematic diagram of the structure of a model training device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0033] The model training method, device and storage medium provided in the embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0034] The term "and / or" in this article is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0035] The terms "first" and "second" and the like in the specification and drawings of this application are used to distinguish different objects, or to distinguish different processing of the same object, rather than to describe a specific order of objects.

[0036] In addition, the terms "including" and "having" and any variations thereof mentioned in the description of the present application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0037] It should be noted that, in the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0038] In the description of the present application, unless otherwise specified, “plurality” means two or more.

[0039] The following explains the terms involved in the embodiments of the present application to facilitate the reader's understanding.

[0040] (1) Confidential computing and trusted execution environment (TEE)

[0041] Confidential computing is a computing paradigm that protects the data and code in use by performing data calculations in a specific environment (for example, a hardware-based trusted execution environment). The credibility of the TEE instance can be evaluated based on a remote attestation process. TEE-based confidential computing is considered to be a scalable solution that can be used to achieve "data is available but not accessible". This solution protects the privacy of data and code by inputting sensitive data and program code into the TEE in ciphertext form to complete the calculation and output the calculation results.

[0042] TEE is an execution environment that enforces that only authorized program code can be executed within the processor, and the data used by the program code cannot be read or tampered with by code outside the TEE. Specifically, TEE is a secure area within the central processing unit (CPU) or graphics processing unit (GPU). TEE runs in an independent environment and runs in parallel with the operating system. The CPU or GPU ensures that the confidentiality and integrity of the code and data in the TEE are protected. Trusted applications running in the TEE can access the full functionality of the device's main processor and memory, while hardware isolation protects the TEE from user-installed applications running in the main operating system. Therefore, data-sensitive application processing can be performed in the TEE, such as key generation and storage, payment account authentication, privacy data processing, and algorithm intellectual property protection.

[0043] (2) AI large models and large model fine-tuning

[0044] AI big models refer to large parameter models trained using large-scale data and powerful computing power, such as generative pre-trained transformer (GPT), vision transformer (VIT), and Baidu Wenxin Yiyan. AI big models are widely used in natural language processing, computer vision, speech recognition, and intelligent recommendation. For example, in the field of natural language processing, they can achieve text generation, machine translation, and intelligent question and answer, in the field of computer vision, they can achieve image classification, target detection, and face recognition, and in the field of intelligent recommendation, they can provide personalized recommendations based on user behavior and preferences.

[0045] Since AI big models have a large number of parameters, require a lot of computing resources and rich data resources, and are costly, some small and medium-sized enterprises tend to use AI big models that have been trained by third parties, and fine-tune them based on their own data to meet the professional needs of their own business or industry characteristics. Big model fine-tuning refers to further training the model based on a pre-trained large language model or other type of large model using relatively less labeled data.

[0046] The above is a brief introduction to some concepts involved in the embodiments of this application.

[0047] With the deployment of the fifth generation mobile communication technology (5G), especially the deployment of 5G private networks, the demand for data processing at the edge of the network is gradually increasing. On the one hand, due to the low latency requirements of specific services, the services are required to provide agile and fast responses. On the other hand, due to data management policy requirements, business data is required to be processed within the park. MEC can provide IT services and cloud computing capabilities in data centers close to the edge of the network, and has gradually become a necessary infrastructure for Internet companies and operators to build edge business ecosystems.

[0048] As mentioned in the background technology section, AI big models are gradually being applied to multiple fields, such as chatbots, content generation, healthcare, and industrial robots. However, the training data or inference data of AI big models may involve highly sensitive data, such as medical data, industrial production data, or human activity data, including audio commands, gestures, and location information in the home. Therefore, it is necessary to use MEC to deploy and fine-tune AI big models at or near the data source, and then control the area where data is used to reduce the risk of data leakage.

[0049] MEC is a system that provides IT services and cloud computing capabilities at the edge of the access network that includes one or more access technologies. Because MEC is close to users, it can handle services with low latency requirements and high privacy protection requirements. MEC receives user-side requirements through a multi-access edge orchestrator (MEO), uniformly orchestrates MEC infrastructure as a service (IaaS) resources, manages the life cycle of MEC applications (APPs), and manages virtualized infrastructure managers (VIMs) and multi-access edge platform managers (MEPMs). VIM is used to manage IaaS resources in local data centers, including resource reporting, resource allocation, and the launch or deregistration of virtual machines (VMs) and virtual containers. In addition, VIM is also used to manage MEC hosts, which run multi-access edge platforms (MEPs), and MEPM is used to manage MEPs.

[0050] For example, Figure 1 A schematic diagram of a MEC system architecture provided in an embodiment of the present application. Figure 1 As shown in Figure 1, the MEC system includes a mobile edge system level and a mobile edge host level. The mobile edge system level includes a customer facing services (CFS) portal, a user equipment application, a user app life cycle management (LCM) proxy, an operations support system (OSS), and a mobile edge orchestrator.

[0051] The mobile edge host layer includes other mobile edge hosts, mobile edge hosts, virtualized infrastructure managers, and mobile edge platform managers. Among them, other mobile edge hosts include other mobile edge platforms. The mobile edge host includes a mobile edge platform, a virtualized infrastructure, and multiple mobile edge applications. The mobile edge platform includes traffic rules control, domain name system (DNS) configuration (handling), a service registry, and mobile edge services. The virtualized infrastructure includes a data plane. The mobile edge application includes services. The mobile edge platform manager includes mobile edge platform element management (platform element management), mobile edge application rule and requirement management, and mobile edge application lifecycle management.

[0052] also, Figure 1 The Mx1 and Mx2 interfaces are interfaces with external entities of MEC. Mp1, Mp2, and Mp3 are interfaces with the MEC platform. Mm1 to Mm9 are interfaces with the management unit of the MEC platform.

[0053] However, MEC is deployed at the edge of the network, where the environment is complex, and the infrastructure and applications belong to different owners. As a result, there is still a risk of leakage of fine-tuning data of large AI models in the shared virtual environment built on general x86.

[0054] In view of this, an embodiment of the present application provides a model training method, which includes: based on the received request information, creating a virtual resource on the target MEC host, installing a MEC application in the virtual resource, configuring the MEC application, and performing model training through the configured MEC application. In other words, the model training method provided by the embodiment of the present application is to create a virtual resource on a host with a trusted execution environment capability, and install the MEC application in the virtual resource, that is, the MEC application is in a trusted execution environment, thereby being able to protect the data privacy of the model training.

[0055] For example, Figure 2 The structural diagram of a model training system 20 provided in an embodiment of the present application is shown. The model training system 20 may include a model training party 201 and an MEC platform 202. The model training party 201 may be connected to the MEC platform 202 for communication.

[0056] The model training party 201 refers to an entity / platform that has model fine-tuning training requirements and is used to send request information to the MEC platform 202.

[0057] The MEC platform 202 is used to provide virtual resources and perform model training through virtual resources.

[0058] Optionally, the model training system 20 may further include a verifier 203 , which may communicate with the MEC platform 202 .

[0059] The verifier 203 is used to perform security verification on the environment where the virtual resources are located.

[0060] It should be noted that Figure 2 This is just an exemplary framework diagram. Figure 2 The number of nodes included in the Figure 2 In addition to the functional nodes shown, the model training system 20 may also include other nodes, and this application does not impose any restrictions on this.

[0061] Furthermore, Figure 3 FIG. 3 is a schematic diagram of another model training system 30 provided in an embodiment of the present application. Figure 3 As shown, the MEC platform 202 may include OSS 301 , MEO 302 , MEPM 303 , MEP 304 , MEC APP 305 , and VIM 306 .

[0062] OSS301 is used to receive and forward request information and response information.

[0063] MEO302, through interaction with MEPM303, coordinates the allocation of resources and the instantiation of applications.

[0064] MEPM303 is used to manage computing resources, storage resources, and network resources.

[0065] MEP304 is used to configure MEC APP305.

[0066] MEC APP305, used for model training.

[0067] VIM306, used to create virtual resources.

[0068] The application scenarios of the embodiments of the present application are not limited. The system architecture and business scenarios described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. It is known to those skilled in the art that with the evolution of the network architecture and the emergence of new business scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0069] When implementing it specifically, Figure 2 and Figure 3 The equipment in Figure 4 The structure shown, or including Figure 4 Parts shown. Figure 4 A schematic diagram of the composition of a model training device 40 provided in an embodiment of the present application. Figure 4 As shown, the model training device 40 may include a processor 401 and a bus 402.

[0070] Furthermore, the model training device 40 may also include a communication interface 403 and a memory 404. The processor 401, the memory 404 and the communication interface 403 may be connected via a bus 402.

[0071] The processor 401 is a CPU, a general processor, a network processor (NP), a digital signal processor (DSP), a microprocessor, a microcontroller, a programmable logic device (PLD), or any combination thereof. The processor 401 may also be other devices with processing functions, such as circuits, devices, or software modules, without limitation.

[0072] The bus 402 is used to transmit information between the components included in the model training device 40.

[0073] The communication interface 403 is used to communicate with other devices or other communication networks. The other communication networks may be Ethernet, radio access network (RAN), wireless local area network (WLAN), etc. The communication interface 403 may be a module, a circuit, a communication interface or any device capable of achieving communication.

[0074] The memory 404 is used to store instructions, where the instructions may be computer programs.

[0075] Among them, the memory 404 can be a read-only memory (ROM) or other types of static storage devices that can store static information and / or instructions, or a random access memory (RAM) or other types of dynamic storage devices that can store information and / or instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), magnetic disk storage medium or other magnetic storage devices, etc., without limitation.

[0076] It should be noted that the memory 404 can exist independently of the processor 401, or can be integrated with the processor 401. The memory 404 can be used to store instructions or program codes or some data, etc. The memory 404 can be located in the model training device 40, or can be located outside the model training device 40, without limitation. The processor 401 is used to execute the instructions stored in the memory 404 to implement the model training method provided in the following embodiments of the present application.

[0077] In an example, the processor 401 may include one or more CPUs, for example, CPU0 and CPU1 (not shown in the figure).

[0078] As an optional implementation, the model training device 40 includes multiple processors.

[0079] As an optional implementation, the model training device 40 further includes an output device and an input device. Exemplarily, the input device is a keyboard, a mouse, a microphone, a joystick, and the like, and the output device is a display screen, a speaker, and the like.

[0080] It should be noted that the model training device 40 can be a desktop computer, a portable computer, a network server, a mobile phone, a tablet computer, a wireless terminal, an embedded device, a chip system or a Figure 4 In addition, Figure 4 The composition shown in the Figure 2 , Figure 3 as well as Figure 4 The limitations of each device in Figure 4 In addition to the parts shown, Figure 2 , Figure 3 as well as Figure 4 The various devices in the drawings may include more or fewer components than shown, or combine certain components, or arrange the components differently.

[0081] In the embodiment of the present application, the chip system may be composed of a chip, or may include a chip and other discrete devices.

[0082] In addition, the actions, terms, etc. involved in the various embodiments of the present application can refer to each other without limitation. The message name or parameter name in the message exchanged between the various devices in the embodiments of the present application is only an example, and other names can also be used in the specific implementation without limitation.

[0083] The model training method provided by the embodiment of the present application is described below in conjunction with the accompanying drawings. Among them, the actions, terms, etc. involved in the various embodiments of the present application can refer to each other without limitation. The message name or parameter name in the message exchanged between the various devices in the embodiment of the present application is only an example, and other names can also be used in the specific implementation without limitation. The actions involved in the various embodiments of the present application are only an example, and other names can also be used in the specific implementation, such as: "included in" in the embodiment of the present application can also be replaced by "carried on" or "carried in", etc.

[0084] like Figure 5 As shown, the embodiment of the present application proposes a model training method, which includes:

[0085] S501: The model training party sends a request message to the MEC platform. Correspondingly, the MEC platform receives the request message from the model training party.

[0086] The request information is used to request the instantiation of virtual resources required for model training. The model training party refers to the entity / platform that has the model fine-tuning training requirements. The MEC platform is used to provide edge infrastructure, including virtual resources or physical host resources, and the edge infrastructure supports confidential computing capabilities based on hardware TEE. Hardware TEE can be based on CPU or GPU.

[0087] Exemplarily, the virtual resources required for the above model training include computing resources, storage resources, and network resources. Among them, the computing resources include CPU TEE instance resources and / or GPU TEE instance resources.

[0088] Exemplarily, the request information may include a MEC APP description file and MEC APP software image information. The MEC APP description file includes virtual resource requirement information and TEE capability requirement information. The MEC APP software image information may be image address information.

[0089] S502: The MEC platform creates virtual resources on the target MEC host based on the request information, and installs the MEC application in the virtual resources.

[0090] Among them, MEC applications are used for model training. The target MEC host has the trusted execution environment TEE capability.

[0091] In one possible implementation, the MEC platform reads and parses the MEC APP description file in the request information, obtains the virtual resource requirements and TEE capability requirements, selects the target MEC host that matches the virtual resource requirements and TEE capability requirements, and creates virtual resources on the target MEC host. The MEC platform installs the MEC application in the virtual resource based on the MEC APP software image information in the request information.

[0092] It should be noted that the prerequisite of this application is that the MEC platform supports the discovery, management and orchestration of TEE-based confidential computing resources in the edge infrastructure.

[0093] S503. The MEC platform configures the MEC application.

[0094] Exemplarily, the above configuration content may include at least one of the following: DNS rules, service discovery rules, data flow rules or TEE remote verification rules.

[0095] DNS rules refer to regulations or policies used to manage and control DNS queries and responses.

[0096] Service discovery rules refer to the regulations or policies used to manage and control service discovery and registration.

[0097] Data flow rules refer to the rules or policies used to manage and control the flow of data. These rules define the transmission mode, path, format and processing of data between different components, services or applications.

[0098] TEE remote verification rules refer to the verification and authentication rules formulated in TEE to ensure the security of remote communication and data processing.

[0099] Of course, the above is only an exemplary description of the configuration content, and the above configuration content may also be other content, and the embodiments of the present application do not impose any limitations on this.

[0100] S504. The MEC platform performs model training through the configured MEC application.

[0101] In one possible implementation, the MEC platform performs model training on the configured MEC application based on model parameters, model training data, model training strategy, and model verification parameters.

[0102] The model provider refers to an entity / platform that can provide large AI models. The data provider refers to an entity / platform that can provide training data sets for model fine-tuning training.

[0103] It should be noted that the configured MEC application can meet the installation and operation of the deep learning framework (for example, the PyTorch framework or the TensorFlow framework) and its related toolkits and dependencies.

[0104] In the model training method provided in this application, the MEC platform creates virtual resources on the target MEC host based on the received request information, installs the MEC application in the virtual resources, configures the MEC application, and performs model training through the configured MEC application. In other words, the MEC platform creates virtual resources on a host with a trusted execution environment capability and installs the MEC application in the virtual resources, that is, the MEC application is in a trusted execution environment, thereby protecting the data privacy of the model training.

[0105] In one embodiment, Figure 6 As shown, the above S504 can be specifically determined through the following S601 to S604.

[0106] S601: The model provider sends model parameters to the MEC platform. Correspondingly, the MEC platform receives the model parameters from the model provider.

[0107] In a possible implementation, when it is determined that the operating environment of the MEC application has passed the verification, the model provider sends the model parameters to the MEC platform.

[0108] It is understandable that when the operating environment of the MEC APP is verified, the model provider loads the model parameters into the MEC APP, which can protect the intellectual property rights of the model.

[0109] S602: The data provider sends model training data to the MEC platform. Correspondingly, the MEC platform receives the model training data from the model trainer.

[0110] In a possible implementation, when it is determined that the operating environment of the MEC application has passed the verification, the data provider sends the model training data to the MEC platform.

[0111] Exemplarily, the model training data may be image data, video data, voice data, or text data that has been cleaned, labeled, and normalized.

[0112] It should be noted that the data provider can also be the model trainer.

[0113] S603: The model training party sends the model training strategy and model verification parameters to the MEC platform. Correspondingly, the MEC platform receives the training strategy and model verification parameters from the model training party.

[0114] In a possible implementation, when it is determined that the operating environment of the MEC application has passed the verification, the model training direction sends the model training strategy and model verification parameters to the MEC platform.

[0115] Exemplarily, the above-mentioned model training strategy can be a parameter efficient fine-tuning (PEFT) strategy, such as a low-rank adaptation (LoRA) strategy, a quantization strategy, a pruning strategy, an adapter tuning strategy, parallel split learning, or split federated learning.

[0116] It is understandable that large model fine-tuning training consumes a lot of computing resources. The above-mentioned efficient parameter fine-tuning strategy can reduce the computing resources and memory resources required for large AI model fine-tuning training, and thus can be used for large AI model fine-tuning training on edge devices with limited MEC resources. For example, a large model with 10 billion parameters can be fine-tuned.

[0117] S604. The MEC platform performs model training on the configured MEC application according to the model parameters, model training data, model training strategy, and model verification parameters.

[0118] In one embodiment, in combination with the above Figure 3 ,like Figure 7 As shown, the above S501 to S503 can be specifically determined by the following S701 to S717.

[0119] S701: The model training party sends a request message to the OSS. Correspondingly, the OSS receives the request message from the model training party.

[0120] The request information is used to request the instantiation of virtual resources required for model training.

[0121] S702: OSS sends request information to MEO. Correspondingly, MEO receives the request information from OSS.

[0122] S703: The MEO determines the VIM that manages the target MEC host and the MEPM associated with the VIM according to the request information.

[0123] In one possible implementation, MEO reads and parses the MEC APP description file in the request information to obtain the virtual resource requirements and TEE capability requirements. MEO evaluates the resources of the currently available MEC hosts, selects the target MEC host that matches the virtual resource requirements and TEE capability requirements, and determines the VIM that manages the target MEC host and the MEPM associated with the VIM.

[0124] S704: MEO sends request information to MEPM. Correspondingly, MEPM receives the request information from MEO.

[0125] The MEPM is the MEPM associated with the VIM determined in the above S703.

[0126] S705: MEPM sends resource allocation request information to VIM. Correspondingly, VIM receives resource allocation request information from MEPM.

[0127] The resource allocation request information is used to request the creation of virtual resources on the MEC host. The resource allocation request includes the MEC APP software image information.

[0128] S706. The VIM creates a virtual container or a virtual machine on the target MEC host, and installs the MEC application in the virtual container or the virtual machine.

[0129] In a possible implementation, the VIM creates a virtual container or a virtual machine on the target MEC host according to the virtual resource requirements, and installs the MEC APP in the created virtual container or virtual machine according to the MEC APP software image information.

[0130] S707: VIM sends resource allocation response information to MEPM. Correspondingly, MEPM receives resource allocation response information from VIM.

[0131] The resource allocation response information includes resource allocation information. The resource allocation information refers to information for creating a virtual container or a virtual machine.

[0132] S708: MEPM sends configuration request information to MEP. Correspondingly, MEP receives the configuration request information from MEPM.

[0133] The configuration request information is used to request configuration of the MEC APP.

[0134] S709, MEP sends acquisition request information to MEC APP. Correspondingly, MEC APP receives the acquisition request information from MEP.

[0135] The acquisition request information is used to request remote attestation information. The remote attestation information is used to describe the environmental parameters of the TEE instance used by the MEC APP.

[0136] Exemplarily, the remote attestation information may include the activation certificate of the TEE instance, the TEE instance identifier (ID), the TEE security version number (SVN), the CPU ID or GPU ID of the MEC host, the trusted computing base (TCB) information, and the signature of the private key corresponding to the activation certificate of the CPU / GPU TEE instance.

[0137] S710: The MEC APP sends remote attestation information to the MEP. Correspondingly, the MEP receives the remote attestation information from the MEC APP.

[0138] S711, the MEP sends a verification request message to the authenticator. Correspondingly, the authenticator receives the verification request message from the MEP.

[0139] The verification request information is used to instruct to verify the remote attestation information. The verification request information includes the remote attestation information. The verifier refers to an entity that can verify the remote attestation information.

[0140] Exemplarily, the remote attestation and verification service can be a remote attestation and verification service for TEE instances provided by CPU or GPU manufacturers, or a local remote attestation and verification service in a data center provided by an infrastructure provider, or a local remote attestation and verification service provided by an MEC operator.

[0141] S712: The authenticator sends a trusted authentication response message to the MEP. Correspondingly, the MEP receives the trusted authentication response message from the authenticator.

[0142] The trusted verification response information is used to indicate that the operating environment of the MEC APP has passed the verification.

[0143] In one possible implementation, the verifier verifies the remote attestation information based on the preset verification information pre-registered by the locally stored TEE instance, and sends a trusted verification response message to the MEP if the remote attestation information is consistent with the preset verification information.

[0144] Exemplarily, the above preset verification information may include the CPU ID or GPU ID of the MEC host, TCB information, TEE SVN, and signature verification.

[0145] Furthermore, when the remote attestation information is inconsistent with the preset verification information, the verification party sends an untrusted verification response message to the MEP. If the MEP receives an untrusted verification response message from the verification party, the MEP terminates the interaction with the MEC APP.

[0146] S713. MEP configures MEC APP.

[0147] In one possible implementation, the MEP configures DNS rules, service discovery, data flow rules, and TEE remote verification for the MEC APP.

[0148] It is understandable that configuring DNS rules and data flow rules for MEC APP can improve network access efficiency and security. Configuring service discovery for MEC APP can promote efficient discovery and utilization of services. Configuring TEE remote verification for MECAPP can ensure the secure execution of applications and data privacy protection.

[0149] S714: MEP sends configuration response information to MEPM. Correspondingly, MEPM receives configuration response information from MEP.

[0150] The configuration response information is used to respond to the configuration of the MEC APP.

[0151] S715: MEPM sends a response message to MEO. Correspondingly, MEO receives the response message from MEPM.

[0152] The response information is used to respond to the instantiation of virtual resources required for the completed model training, and includes resource allocation information.

[0153] S716: MEO sends a response message to OSS. Correspondingly, OSS receives the response message from MEO.

[0154] S717, OSS sends response information to the model training party. Correspondingly, the model training party receives the response information from OSS.

[0155] Furthermore, after S717, the MEC APP needs to obtain model training parameters so that it can subsequently perform model training in the configured MEC application according to the model training parameters. The model training parameters include model parameters, model training data, model training strategy, and model verification parameters. In view of this, Figure 8 As shown, the model training method described in the embodiment of the present application may also include the following steps.

[0156] S801: The model training party sends a first request message to the model provider. Correspondingly, the model provider receives the first request message from the model training party.

[0157] The first request information is used to request to load the model parameters into the MEC APP. The first request information includes the entry information of the MEC APP.

[0158] S802: The model provider sends an acquisition request message to the MEC APP. Correspondingly, the MEC APP receives the acquisition request message from the model provider.

[0159] The acquisition request information is used to request remote certification information.

[0160] S803, MEC APP sends remote certification information to the model provider. Correspondingly, the model provider receives the remote certification information from MEC APP.

[0161] S804: The model provider sends a verification request message to the verifier. Correspondingly, the verifier receives the verification request message from the model provider.

[0162] The verification request information is used to instruct to verify the remote certification information. The verification request information includes the remote certification information.

[0163] S805: The verifier sends a trusted verification response message to the model provider. Correspondingly, the model provider receives the trusted verification response message from the verifier.

[0164] The trusted verification response information is used to indicate that the operating environment of the MEC APP has passed the verification.

[0165] S806: The model provider loads the model parameters into the MEC APP.

[0166] It is understandable that when the operating environment of the MEC APP is verified, the model provider loads the model parameters into the MEC APP, which can protect the intellectual property rights of the model.

[0167] S807: The model provider sends first response information to the model trainer. Correspondingly, the model trainer receives the first response information from the model provider.

[0168] The first response information is used to respond that the model parameters have been loaded into the MEC APP.

[0169] S808: The model training party sends a second request message to the data provider. Correspondingly, the data provider receives the first request message from the model training party.

[0170] The second request information is used to request to load the model training data into the MEC APP. The second request information includes the entry information of the MEC APP.

[0171] It should be noted that the model training party may also be the data provider. In the case where the model training party is the data provider, there is no need to execute the above S808, and S809 is directly executed.

[0172] S809: The data provider sends an acquisition request message to the MEC APP. Correspondingly, the MEC APP receives the acquisition request message from the data provider.

[0173] The acquisition request information is used to request remote certification information.

[0174] S810, MEC APP sends remote certification information to the data provider. Correspondingly, the data provider receives the remote certification information from MEC APP.

[0175] S811: The data provider sends a verification request to the verifier. Correspondingly, the verifier receives the verification request from the data provider.

[0176] The verification request information is used to instruct to verify the remote certification information. The verification request information includes the remote certification information.

[0177] S812: The verifier sends a trusted verification response message to the data provider. Correspondingly, the data provider receives the trusted verification response message from the verifier.

[0178] The trusted verification response information is used to indicate that the operating environment of the MEC APP has passed the verification.

[0179] S813. The data provider loads the model training data into the MEC APP.

[0180] Exemplarily, the model training data may be image data, video data, voice data, or text data that has been cleaned, labeled, and normalized.

[0181] It is understandable that when the operating environment of the MEC APP is verified, the data provider loads the model training data into the MEC APP, which can protect the privacy of the model training data.

[0182] S814: The data provider sends second response information to the model trainer. Correspondingly, the model trainer receives the second response information from the data provider.

[0183] The second response information is used to respond that the model training data has been loaded into the MEC APP.

[0184] S815: The model training party sends an acquisition request message to the MEC APP. Correspondingly, the MEC APP receives the acquisition request message from the model training party.

[0185] The acquisition request information is used to request remote certification information.

[0186] S816, MEC APP sends remote certification information to the model training party. Correspondingly, the model training party receives the remote certification information from MEC APP.

[0187] S817: The model trainer sends a verification request message to the verifier. Correspondingly, the verifier receives the verification request message from the model trainer.

[0188] The verification request information is used to instruct to verify the remote certification information. The verification request information includes the remote certification information.

[0189] S818: The verification party sends a trusted verification response message to the model training party. Correspondingly, the model training party receives the trusted verification response message from the verification party.

[0190] The trusted verification response information is used to indicate that the operating environment of the MEC APP has passed the verification.

[0191] S819: The model training party loads the model training strategy and model verification parameters into the MEC APP.

[0192] It should be noted that the order of the above steps S801 to S807 and steps S808 to S814 is not limited, and the two can be performed in parallel or alternately. In addition, model fine-tuning training can be performed in parallel on multiple MEC APPs, that is, the above steps S801 to S819 can be performed in parallel on multiple MEC APPs at the same time.

[0193] Further, after S819, MEC APP performs model fine-tuning training according to model parameters, model training data, model training strategy, and model verification parameters to obtain a fine-tuned model, and sends the fine-tuned model to the model training party. Correspondingly, the model training party receives the fine-tuned model from MEC APP. The model training party has data ownership of the fine-tuned model.

[0194] To sum up, by adopting the model training method provided in the embodiment of the present application, MEC tenants, system administrators, and infrastructure providers cannot obtain the data and algorithms used for fine-tuning training of large AI models, thereby protecting the data privacy of model fine-tuning training on the one hand, and protecting the intellectual property rights of large AI models on the other.

[0195] It is understandable that the above-mentioned model training method can be implemented by a model training device. In order to realize the above-mentioned functions, the model training device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the modules and algorithm steps of each example described in the embodiments disclosed herein, the embodiments disclosed in this application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the embodiments disclosed in this application.

[0196] The embodiments disclosed in the present application can divide the functional modules according to the model training device generated by the above method example. For example, each functional module can be divided corresponding to each function, or two or more functions can be integrated into one processing module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiments disclosed in the present application is schematic and is only a logical functional division. There may be other division methods in actual implementation.

[0197] Fig. 9 This is a schematic diagram of the structure of a model training device provided in an embodiment of the present application. Fig. 9 As shown, the model training device 90 can be used to perform Figure 5-Figure 8 The model training method shown. The model training device 90 includes: a communication unit 901 and a processing unit 902.

[0198] The communication unit 901 is used to receive request information, where the request information is used to request the instantiation of virtual resources required for model training; the processing unit 902 is used to create virtual resources on the target MEC host based on the request information, and install MEC applications in the virtual resources, where the MEC applications are used for model training, and the target MEC host has a trusted execution environment TEE capability; the processing unit 902 is also used to configure the MEC application and perform model training through the configured MEC application.

[0199] In a possible implementation, the processing unit 902 is further used to determine the virtualization infrastructure manager VIM that manages the target MEC host based on the request information; the processing unit 902 is further used to create a virtual container or a virtual machine on the target MEC host through the VIM, and install the MEC application in the virtual container or the virtual machine.

[0200] In one possible implementation, the processing unit 902 is further used to determine the requirement information for model training based on the request information, where the requirement information includes the virtual resources required for the model training and the required TEE capabilities of the MEC host; the processing unit 902 is further used to determine a target MEC host that meets the requirement information from multiple MEC hosts.

[0201] In a possible implementation, when the operating environment of the MEC application is verified, the processing unit 902 is also used to configure the MEC application.

[0202] In one possible implementation, the communication unit 901 is also used to send remote certification information of the operating environment of the MEC application to the verifier, so that the verifier can perform security verification on the operating environment of the MEC application; the communication unit 901 is also used to receive response information from the verifier, and the verification response information is used to indicate whether the operating environment of the MEC application has passed the verification.

[0203] In one possible implementation, the communication unit 901 is also used to obtain model training parameters, wherein the model training parameters include: model parameters, model training data, model training strategies, and model verification parameters; the processing unit 902 is also used to perform model training in the configured MEC application according to the model training parameters.

[0204] Through the description of the above implementation methods, technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device and unit described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0205] The present disclosure also provides a computer-readable storage medium having instructions stored thereon. When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the model training method provided by the above-mentioned embodiment of the present disclosure.

[0206] The embodiments of the present disclosure also provide a computer program product containing instructions, which, when executed on an electronic device, enables the electronic device to execute the model training method provided by the embodiments of the present disclosure.

[0207] Among them, the computer readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a register, a hard disk, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above, or any other form of computer readable storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application-specific integrated circuit (ASIC). In the embodiments of the present application, a computer-readable storage medium may be any tangible medium that contains or stores a program, which may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0208] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto, and any changes or substitutions within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A model training method, characterized in that: Applied to a multi-access edge computing MEC platform, the method includes: Receive request information, where the request information is used to request instantiation of virtual resources required for model training; Based on the request information, create a virtual resource on the target MEC host, and install a MEC application in the virtual resource, wherein the MEC application is used for model training, and the target MEC host has a trusted execution environment TEE capability; The MEC application is configured, and model training is performed through the configured MEC application.

2. The method according to claim 1, characterized in that The step of creating a virtual resource on a target MEC host based on the request information and installing a MEC application in the virtual resource includes: Determine, according to the request information, a virtualization infrastructure manager VIM that manages the target MEC host; A virtual container or a virtual machine is created on the target MEC host through the VIM, and the MEC application is installed in the virtual container or the virtual machine.

3. The method according to claim 2, characterized in that The MEC platform is configured with multiple MEC hosts, and the method further includes: Determine the requirement information of the model training according to the request information, where the requirement information includes the virtual resources required for the model training and the required TEE capabilities of the MEC host; From the multiple MEC hosts, determine the target MEC host that meets the requirement information.

4. The method according to claim 1, characterized in that: The configuring of the MEC application includes: When the operating environment of the MEC application passes the verification, the MEC application is configured.

5. The method according to claim 4, characterized in that The method further comprises: Sending remote attestation information of the operating environment of the MEC application to the verifier, so that the verifier performs security verification on the operating environment of the MEC application; Receive response information from the verifier, where the verification response information is used to indicate whether the operating environment of the MEC application has passed the verification.

6. The method according to any one of claims 1 to 5, characterized in that The model training is performed by the configured MEC application, including: Obtain model training parameters; wherein the model training parameters include: model parameters, model training data, model training strategy and model verification parameters; According to the model training parameters, the MEC application after the configuration performs model training.

7. A model training device, characterized in that: The device comprises: a communication unit and a processing unit; The communication unit is used to receive request information, where the request information is used to request instantiation of virtual resources required for model training; The processing unit is used to create a virtual resource on a target MEC host based on the request information, and install a MEC application in the virtual resource, wherein the MEC application is used for model training, and the target MEC host has a trusted execution environment TEE capability; The processing unit is also used to configure the MEC application and perform model training through the configured MEC application.

8. A model training device, characterized in that: include: Memory and processor; The memory is coupled to the processor; The memory is used to store instructions executable by the processor; When the processor executes the instructions, it performs the model training method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and when the computer instructions are executed on a computer, the computer executes the model training method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product includes computer program instructions, which, when executed by a processor, implement the model training method as described in any one of claims 1 to 6.

Images