Model training method and system based on dma-mass federated learning platform

Abstract

The application provides a model training method and system based on a DMA-MaaS federated learning platform, which comprises the following steps: uploading training data to the federated learning platform, checking the training data by the federated learning platform, and adding the training data to a data pool; uploading a task to the federated learning platform, adding the public task to a public task pool by the federated learning platform so as to be selected by other user terminals; selecting a task initiated by oneself or selecting a task in the public task pool, judging whether the selected task type is federated learning, if yes, performing federated learning based on the training data locally on a user device where the user terminal is located, returning model parameters and results obtained by learning to the federated learning platform for parameter aggregation, until the aggregated model reaches the required performance, otherwise, performing non-federated learning on the cloud based on the training data by the federated learning platform. Through the MaaS function, the heterogeneity of the user side of the federated platform is relieved, the management and creation of data, tasks, algorithms and models are completed, and the value of the federated model is played.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a model training method and system based on the DMA-MaaS federated learning platform. Background Technology

[0002] In recent years, with the development of machine learning and deep learning technologies, they have been widely applied in various fields. However, due to factors such as competition between companies, security issues, and complex approval processes, there are significant obstacles to data cooperation between companies and even data interconnection within companies, leading to the phenomenon of "data silos." At the same time, with some large enterprises experiencing user privacy data breaches, data privacy protection has become a global concern, and countries are continuously improving and strengthening relevant laws and regulations on data privacy. This poses a significant challenge to the development of machine learning technologies that rely on data.

[0003] To address the two major issues of "data silos" and "data privacy," numerous scholars and institutions have proposed a variety of solutions, among which the Federated Learning algorithm framework, proposed in 2016, has received widespread attention. With the rise of Federated Learning, after several years of rapid development, the technology and industrialization are gradually maturing, giving rise to numerous platforms and products, and beginning large-scale commercial deployment. It is currently being applied in many fields, including finance, healthcare, and smart cities. However, Federated Learning platforms face a new model that differs from traditional machine learning platforms that only provide training and inference services. Federated platforms suffer from heterogeneity in user devices (such as data heterogeneity, model heterogeneity, and execution environment heterogeneity). Therefore, the data, models, and algorithms of federated platforms require new methods for management, solving problems while ensuring user convenience.

[0004] Current federated learning platforms primarily focus on communication, privacy, and processes, while lacking scientific and reasonable methods for managing data, tasks, algorithms, and models to address the heterogeneity problem in federated learning. Summary of the Invention

[0005] This invention focuses on the scientific management of data, tasks, algorithms, and models within a federated platform. It designs a service platform based on DMA (Data / Mission / Algorithm Pool) and MaaS (Model as a Service) to implement an object-oriented federation mechanism. This allows for the direct provision of more lightweight and user-friendly federated services to various users. The federated platform architecture of this invention enables developers to complete their work with simple operations such as clicks, easily creating and assembling tasks and allowing federated models to realize their value. To achieve the above technical effects, this invention includes the following key technical points:

[0006] Key point 1: Adopt a federated platform architecture design based on DMA-MaaS service platform;

[0007] Key Point 2: Data management methods for implementing data quality assessment and incentive mechanisms within the data pool;

[0008] Key point 3: After evaluating the cluster user environment in the Mission Pool, set up personalized federated or asynchronous federated tasks for heterogeneous federated learning.

[0009] Key point 4: The federated platform algorithm management method that divides algorithms into federated and non-federated algorithms in the algorithm pool;

[0010] Key Point 5: A federated platform model management method based on MaaS (Model as a Service) for incremental learning routing control of model flow and model uploading, downloading, use, and billing.

[0011] Specifically, this invention proposes a model training method based on the DMA-MaaS federated learning platform, which includes:

[0012] The data preparation steps involve the user uploading training data to the federated learning platform and setting whether the training data is public or not. The federated learning platform then checks the training data and adds it to the data pool.

[0013] The task initiation process involves the user client uploading the task to the federated learning platform. The task content includes the task name, whether the task is public, whether the task type is federated learning, the algorithm and parameters required for the task, the model and parameters required for the task, and the number of participants. The federated learning platform adds the public tasks to the public task pool for other users to choose from.

[0014] In the task preparation step, the user selects a task they initiated or a task from the public task pool. The system determines whether the selected task type is federated learning. If it is, the federated learning step is executed; otherwise, the non-federated learning step is executed.

[0015] This federated learning step involves performing federated learning locally on the user device where the user is located, based on the training data, and returning the learned model parameters and results to the federated learning platform for parameter aggregation until the aggregated model reaches the required performance.

[0016] This non-federated learning step involves the federated learning platform performing non-federated learning in the cloud based on the training data to obtain a model that meets the required performance.

[0017] The model training method based on the DMA-MaaS federated learning platform includes an algorithm pool containing federated learning algorithms and non-federated algorithms. The federated learning algorithms include horizontal federated learning algorithms, vertical federated learning algorithms, user-preset federated learning algorithms, and secure aggregation methods. The non-federated learning algorithms include a stochastic gradient descent model optimization algorithm.

[0018] The model training method based on the DMA-MaaS federated learning platform includes a model pool containing various preset models. The model management of the federated learning platform includes model uploading, model downloading, model usage, model billing, model marketplace, and model promotion. Model uploading, downloading, usage, billing, and promotion are all conducted in the model marketplace, providing users with diverse model management options.

[0019] Within the existing data, tasks, algorithms, and models on this federated learning platform, tasks can be initiated using the following methods. The platform will also recommend appropriate parameter configurations based on user frequency:

[0020] Initiating from the data pool: After users see the desired public data in the data pool, they can initiate a task on the data page. At this point, they also need to select the corresponding algorithm and model, or choose an algorithm and model recommended by the platform; or

[0021] Initiating from the algorithm pool: Users can see the algorithms they want to train in the algorithm pool and initiate the task on the algorithm page. At this time, they also need to select the corresponding data and model, or select data and models recommended by the platform; or

[0022] Starting from the model marketplace: After seeing the model they want to use in the model marketplace, users can initiate a task on the model page, either an inference task or a federated task, as needed. In an inference task, users only need to select the data they want to use to perform model inference and obtain the inference result. In a federated task, users also need to select the data and algorithm to use to initiate the task, or select the data and algorithm recommended by the platform.

[0023] The model training method based on the DMA-MaaS federated learning platform is described above, wherein the model is an image classification model, the training data consists of images with labeled categories, and the task is an image classification task.

[0024] It also includes a model execution step, in which the image to be identified is input into the image classification model to perform the image classification task and obtain the category to which the image to be identified belongs.

[0025] This invention also proposes a model training system based on the DMA-MaaS federated learning platform, which includes:

[0026] The data preparation module is used to upload training data to the federated learning platform and set whether the training data is public. The federated learning platform checks the training data and adds it to the data pool.

[0027] The task initiation module is used to upload tasks to the federated learning platform. The task content includes the task name, whether the task is public, whether the task type is federated learning, the algorithm and parameters required for the task, the model and parameters required for the task, and the number of participants in the task. The federated learning platform adds public tasks to the public task pool for other users to choose from.

[0028] The task preparation module is used to determine whether the selected task type is federated learning by selecting a task initiated by itself or by selecting a task from the public task pool. If it is, the federated learning module is executed; otherwise, the non-federated learning module is executed.

[0029] This federated learning module is used to perform federated learning locally on the user device where the user terminal is located, based on the training data, and return the learned model parameters and results to the federated learning platform for parameter aggregation until the aggregated model reaches the required performance.

[0030] This unfederated learning module enables the federated learning platform to perform unfederated learning in the cloud based on the training data, thereby obtaining a model that meets the required performance.

[0031] The model training system based on the DMA-MaaS federated learning platform includes an algorithm pool containing federated learning algorithms and non-federated algorithms. The federated learning algorithms include horizontal federated learning algorithms, vertical federated learning algorithms, user-preset federated learning algorithms, and secure aggregation methods. The non-federated learning algorithms include a stochastic gradient descent model optimization algorithm.

[0032] The model training system based on the DMA-MaaS federated learning platform includes a model pool containing various preset models. The model management of the federated learning platform includes model uploading, model downloading, model usage, model billing, model marketplace, and model promotion. Model uploading, downloading, usage, billing, and promotion are all conducted in the model marketplace, providing users with diverse model management options.

[0033] Within the existing data, tasks, algorithms, and models on this federated learning platform, tasks can be initiated using the following methods. The platform will also recommend appropriate parameter configurations based on user frequency:

[0034] Initiating from the data pool: After users see the desired public data in the data pool, they can initiate a task on the data page. At this point, they also need to select the corresponding algorithm and model, or choose an algorithm and model recommended by the platform; or

[0035] Initiating from the algorithm pool: Users can see the algorithms they want to train in the algorithm pool and initiate the task on the algorithm page. At this time, they also need to select the corresponding data and model, or select data and models recommended by the platform; or

[0036] Starting from the model marketplace: After seeing the model they want to use in the model marketplace, users can initiate a task on the model page, either an inference task or a federated task, as needed. In an inference task, users only need to select the data they want to use to perform model inference and obtain the inference result. In a federated task, users also need to select the data and algorithm to use to initiate the task, or select the data and algorithm recommended by the platform.

[0037] The model training system based on the DMA-MaaS federated learning platform, wherein the model is an image classification model, the training data is images with labeled categories, and the task is an image classification task;

[0038] It also includes a model execution module, which inputs the image to be identified into the image classification model to perform the image classification task and obtain the category to which the image to be identified belongs.

[0039] The present invention also proposes a storage medium for storing a program that executes a model training method based on any of the DMA-MaaS federated learning platforms as described in claims 1 to 4.

[0040] The present invention also proposes a client for a model training system based on any one of the DMA-MaaS federated learning platforms as described in claims 5 to 8.

[0041] As can be seen from the above solutions, the advantages of the present invention are:

[0042] Compared to existing technologies, the present invention provides a DMA-MaaS federated platform architecture design, DMA-related optimizations for heterogeneous scenarios of federated learning, and Model-as-a-Service (MaaS) for incremental learning routing control and model management of model streams. Attached Figure Description

[0043] Figure 1 A design diagram of a federated platform architecture based on DMA-MaaS;

[0044] Figure 2 This is a block diagram of the overall system of the present invention;

[0045] Figure 3 This is a flowchart of the federated learning method of the present invention;

[0046] Figure 4 This is a schematic diagram of the data processing device of the present invention. Detailed Implementation

[0047] The purpose of this invention is to address the shortcomings of existing federated platforms in terms of their insufficient targeting of federated models and the lack of new management models for data, tasks, algorithms, and models. It proposes a federated platform architecture design based on DMA-MaaS. Through innovative data pools, mission pools, and algorithm pools, and their interconnected service platform and model marketplace, which provide Model as a Service (MaaS) functions such as model fusion and model increment, the heterogeneity of the federated platform on the user side can be alleviated. This enables the management and rapid creation of data, tasks, algorithms, and models, and maximizes the value of federated models.

[0048] This invention is a novel management method for data, tasks, algorithms, and models based on DMA-MaaS, situated between the underlying driver service and the application layer of a federated platform, addressing the heterogeneity of federated systems. The overall architecture is as follows: Figure 1 As shown.

[0049] The federated platform architecture based on DMA-MaaS consists of a DMA service platform and a MaaS service layer, as well as the interaction between the two, in addition to the underlying drivers and upper-layer applications.

[0050] 1. The DMA service platform consists of three parts: a data pool, a task pool, and an algorithm pool.

[0051] 1) Data Pool: The data pool is mainly used for user data management and contains data metadata and federated data processing methods.

[0052] The metadata of data includes data name, data type, data permissions, etc. In table data, it includes information such as data quantity, column name, column data type, etc. In image data, it includes information such as image name, image size, and label.

[0053] Federated data suffers from heterogeneity, such as significant differences in data distribution and quantity. This can lead to poor performance, insufficient generalization ability, and unfairness in federated models. Therefore, quality assessment and further processing of federated data are necessary. Methods such as data quality quantification based on prior and posterior fusion can be used to ensure the generalization ability and fairness of federated models.

[0054] 2) Mission Pool: The mission pool contains various tasks of the user and their corresponding federation modes.

[0055] The tasks in the task pool include task initiation, task modification, task preparation, task distribution, task execution, and task completion.

[0056] Simultaneously, different federated task modes need to be provided during task construction for different algorithms. To address the issue of data heterogeneity, personalized federated tasks are required when improving model performance on single user data. When the execution environments of different federated parties vary significantly, such as due to network, computing power, or interruptions, asynchronous federated learning tasks are needed to mitigate the heterogeneity of the execution environment and prevent wasting considerable time and computing power due to long waiting times. For use in the model marketplace, only inference tasks need to be established.

[0057] 3) Algorithm Pool: The algorithm pool contains platform-preset algorithms and user-defined algorithms, including federated and non-federated algorithms. Federated algorithms include horizontal federated learning algorithms (FedAvg), vertical federated learning algorithms, user-preset personalized federated learning algorithms, and various safe aggregation methods. Non-federated algorithms include model optimization methods such as stochastic gradient descent.

[0058] 2. The MaaS service layer includes model management, model application, and inference.

[0059] Model as a Service (MaaS) is an important way for federated platforms to generate revenue and make applications. Each federated or non-federated user co-builds a model as an independent product, which is then displayed on the MaaS layer and provides interfaces for initiating training or inference tasks.

[0060] On the platform side, model management requires storing each user's model, including its structure, the deep learning framework used, training parameters (such as pre-trained models, batch size, number of iterations, and learning rate, allowing users to fine-tune models more quickly), and input data format. A distributed storage approach is used to ensure important model assets are not lost. When a model is listed on the marketplace, a basic inference cost is calculated based on the model's parameters and required computing power as a reference. Security measures are implemented according to the model's public and private access levels to prevent parameter leakage and potential loss of user assets.

[0061] On the user side, model management includes model uploading, downloading, using, billing, and a model marketplace. Uploading, downloading, using, and billing all take place in the model marketplace, providing users with flexible and diverse model management methods. The platform's pre-built models include decision trees, random forests, linear regression, and convolutional neural networks.

[0062] On the user side, the pre-trained models provided by the platform can quickly start model training using services such as model increment and model transfer, reducing the process of exploring initial model parameters, reducing task iteration rounds, and accelerating model training efficiency. Meanwhile, to ensure full utilization of the model, intelligent model routing based on opportunistic incremental learning can also be used to optimize model flow management during incremental learning.

[0063] In terms of model promotion on the user side, the platform can provide corresponding model inference APIs, enabling users to perform model inference using their devices and facilitating the promotion of their own businesses.

[0064] In the model marketplace, users build and upload models through federated or non-federated tasks in the task pool, while platforms provide related trained models. Users and platforms can provide model services through the model marketplace, generate profits from models, realize the value of models, and empower models.

[0065] 3. DMA-MaaS linkage.

[0066] DMA-MaaS synergy refers to the organic combination and linkage of various components within the architecture of the data pool, task pool, algorithm pool, and model marketplace. The entire DMA-MaaS framework revolves around tasks. When creating a task, users need to select the corresponding data, algorithm, and model. Tasks can be initiated and created directly from any module in the algorithm pool, data pool, or model marketplace. At this point, only the missing two items need to be selected. Simultaneously, the platform will recommend relevant options based on the frequency of user matching and selection, facilitating quick initiation and search for users.

[0067] When users select and match algorithms, data, and models, the platform will provide default parameter configuration files based on the algorithm type (federated algorithm, secure algorithm) and data type (image multi-classification, multi-label classification, image segmentation, text, speech, etc.) to facilitate users in adjusting and modifying the corresponding parameters.

[0068] To make the above features and effects of the present invention clearer and easier to understand, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings.

[0069] In this invention, users train the model locally, but the trained model participates in global (federated) model aggregation to ensure user data security. Following the user's operation flow, this invention is divided into the following stages: Figure 3 As shown.

[0070] 1. Data preparation stage:

[0071] When users use the architecture platform designed based on this invention, the stage they are in is the data preparation stage, which is as follows:

[0072] 1) Data Import: Click the "Create Data" button on the interactive interface. The system will then prompt you with the necessary preparations and steps for the upcoming data upload operation. After verifying the data is correct, set data permissions (public or private), fill in the data description, and upload the data to the cluster data management interface provided by the underlying framework through the interactive interface. Users can then upload data to the underlying driver's K3S cluster via the data upload service provided by the federation platform.

[0073] 2) Data Pool Processing: The system checks the relevant attributes of the user-submitted data to ensure that the data format uploaded by all parties is consistent, thus ensuring the normal progress of subsequent model training. After the check is passed, the user-submitted data is added to the data pool. If it is public data, other users can also see the data in the public data pool. Subsequently, a data quality assessment procedure is executed on the user side, and corresponding incentive mechanisms are applied for appropriate processing.

[0074] The data quality assessment process on the user side includes identifying areas where user data may contain excessive duplicate content, redundancy, or a lot of fabricated or noisy data, or the data volume may be very small, thus resulting in low data quality. Subsequent incentive mechanisms are then implemented to encourage users who provide high-quality data samples.

[0075] 2. Task Initiation Phase:

[0076] After completing the data preparation phase, the user enters the task initiation phase, which proceeds as follows:

[0077] 1) Task Initiation: Users click "Initiate Task" on the interactive interface. The system will then prompt the user with the upcoming task discovery process and the necessary preparations and steps. Users fill in the task name, select task visibility (public, private), task type (federated, non-federated), the required algorithm and its parameters, the required model and its parameters, the number of participants, the required data format, and complete the task initiation by filling in the task description.

[0078] 2) Task pool processing: When the system receives a task initiated by a user, it adds the task to the public task pool based on the task's visibility. Other users can then choose to add the task.

[0079] 3. Task preparation phase:

[0080] After the user initiates the task, the task preparation phase begins, which proceeds as follows:

[0081] 1) Users can view their initiated tasks by clicking "My Tasks," select tasks to join from the public task pool, or view and join tasks using invitation codes from others. Each task includes the task name, task visibility, task flow type, task description, task-related parameters, task data, task model, task algorithm, the readiness status of each participant and their equipment, task status, and task operations. Tasks can be categorized as public or private based on visibility, and as federated or non-federated based on the task flow.

[0082] 2) After the user clicks to join the task, they will be asked to select data. If the user has already completed the data preparation stage, they can directly select the data they need. Otherwise, they need to go through the data preparation stage again to select the data.

[0083] 3) After the user completes the joining task and grants the relevant permissions, the joining task is complete.

[0084] 4) After a user joins a task, the system begins to check whether the user's device is functioning properly, including data preparation, network environment, and connection status. It also performs an execution environment assessment and uses the corresponding federated task mode, such as asynchronous federated task or personalized federated task.

[0085] 4. Task Implementation and Completion Phase:

[0086] Once all participants have completed the task preparation phase, the task can begin. The process for this phase is as follows:

[0087] 1) Each user, through centralized management, uses the algorithms and models required for the task. After incremental adjustments to the corresponding models, they train the models using their own data. The trained model parameters and results are then returned to the platform via the cluster for parameter aggregation until the model reaches the required performance.

[0088] 2) During the task, each user can see the training progress and status. If the model does not reach the required performance after one round of training, the model flow-based incremental model method will be used to continue optimization training.

[0089] 3) After all training tasks are completed, the task status changes to "completed" and the model with the task completed is provided.

[0090] 5. Model Empowerment Phase:

[0091] After completing the task, users can proceed to the model empowerment phase as needed. The process for this phase is as follows:

[0092] 1) Users can select the model after the task is completed, fill in the relevant model description, and list it on the model marketplace, including information on the participants, data attributes, quantity, etc.

[0093] 2) Other users can view the model and its description in the model marketplace, purchase and download the model as needed, or use the online inference service.

[0094] 3) Users who have uploaded the model can earn revenue through the purchase and use of other users and settle accounts, thus completing the model empowerment.

[0095] 6. DMA-MaaS linkage phase:

[0096] Tasks can be quickly initiated using the DMA-MaaS linkage method, based on existing data, tasks, algorithms, and models on the platform.

[0097] 1) Initiating from the data pool: After seeing the public data they want in the data pool, users can initiate tasks on the data page. At this time, they also need to select the corresponding algorithm and model.

[0098] 2) Initiating from the algorithm pool: Users can see the algorithms they want to use for training in the algorithm pool and initiate tasks on the algorithm page. At this time, they also need to select the corresponding data and model.

[0099] 3) Initiating from the Model Marketplace: After finding a desired model in the model marketplace, users can initiate a task on the model page, choosing between an inference task or a federated task as needed. In an inference task, users simply select their data to perform model inference and obtain the result. In a federated task, users also need to select the data and algorithm to use to initiate the task.

[0100] Specific application case: Hospital A wants to build an ophthalmic fundus image-assisted diagnostic model, but because the data requires expert annotation, it only has a limited amount of annotated data, perhaps less than 1,000 images, insufficient to build a robust diagnostic model on its own. Alternatively, the hospital may have a small sample size of certain diagnosed cases, potentially leading to overfitting to a particular disease. Hospital A learns that Hospitals B and C have similar needs and data. All three parties can upload their data to their local clusters via a federated platform. Within the federated platform system, Hospital A clicks on its uploaded data, creating an ophthalmic auxiliary diagnostic model building task through DMA linkage. Simultaneously, based on platform recommendations, it quickly locates and selects the desired model and sends task invitations to Hospitals B and C. Hospitals B and C find the ophthalmic auxiliary diagnostic model building task initiated by Hospital A in the task pool, upload their data, and the federated platform checks it according to Hospital A's data requirements. After preparing their respective equipment, the task begins upon Hospital A's confirmation.

[0101] Once the training task is completed, hospitals A, B, and C can all see the model obtained after the task in My Models, and can choose whether to make it public. If it is made public, other users can see the model in the model pool.

[0102] Hospital D happens to have a relevant need, but because it lacks its own data, it did not participate in the task. It can wait until the model is made public, find it in the model pool, and leverage the platform's MaaS (Model-as-a-Service) model empowerment feature to acquire the model for a fee and initiate an inference task. By using its own data for inference, it can obtain auxiliary diagnostic results. Hospitals A, B, and C can also receive revenue based on their respective levels of participation through the model's use.

[0103] If Hospital A wants to promote its work and utilize its computing power, it can use the model promotion service provided by the platform's MaaS function to bind its computing power. The platform will then provide a corresponding model inference service interface, allowing Hospital A to build its own inference platform service and provide model inference service functions to external users.

[0104] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0105] This invention also proposes a model training system based on the DMA-MaaS federated learning platform, which includes:

[0106] The data preparation module is used to upload training data to the federated learning platform and set whether the training data is public. The federated learning platform checks the training data and adds it to the data pool.

[0107] The task initiation module is used to upload tasks to the federated learning platform. The task content includes the task name, whether the task is public, whether the task type is federated learning, the algorithm and parameters required for the task, the model and parameters required for the task, and the number of participants in the task. The federated learning platform adds public tasks to the public task pool for other users to choose from.

[0108] The task preparation module is used to determine whether the selected task type is federated learning by selecting a task initiated by itself or by selecting a task from the public task pool. If it is, the federated learning module is executed; otherwise, the non-federated learning module is executed.

[0109] This federated learning module is used to perform federated learning locally on the user device where the user terminal is located, based on the training data, and return the learned model parameters and results to the federated learning platform for parameter aggregation until the aggregated model reaches the required performance.

[0110] This unfederated learning module enables the federated learning platform to perform unfederated learning in the cloud based on the training data, thereby obtaining a model that meets the required performance.

[0111] The model training system based on the DMA-MaaS federated learning platform includes an algorithm pool containing federated learning algorithms and non-federated algorithms. The federated learning algorithms include horizontal federated learning algorithms, vertical federated learning algorithms, user-preset federated learning algorithms, and secure aggregation methods. The non-federated learning algorithms include a stochastic gradient descent model optimization algorithm.

[0112] The model training system based on the DMA-MaaS federated learning platform includes a model pool containing various preset models. The model management of the federated learning platform includes model uploading, model downloading, model usage, model billing, model marketplace, and model promotion. Model uploading, downloading, usage, billing, and promotion are all conducted in the model marketplace, providing users with diverse model management options.

[0113] Within the existing data, tasks, algorithms, and models on this federated learning platform, tasks can be initiated using the following methods. The platform will also recommend appropriate parameter configurations based on user frequency:

[0114] Initiating from the data pool: After users see the desired public data in the data pool, they can initiate a task on the data page. At this point, they also need to select the corresponding algorithm and model, or choose an algorithm and model recommended by the platform; or

[0115] Initiating from the algorithm pool: Users can see the algorithms they want to train in the algorithm pool and initiate the task on the algorithm page. At this time, they also need to select the corresponding data and model, or select data and models recommended by the platform; or

[0116] Starting from the model marketplace: After seeing the model they want to use in the model marketplace, users can initiate a task on the model page, either an inference task or a federated task, as needed. In an inference task, users only need to select the data they want to use to perform model inference and obtain the inference result. In a federated task, users also need to select the data and algorithm to use to initiate the task, or select the data and algorithm recommended by the platform.

[0117] The model training system based on the DMA-MaaS federated learning platform, wherein the model is an image classification model, the training data is images with labeled categories, and the task is an image classification task;

[0118] It also includes a model execution module, which inputs the image to be identified into the image classification model to perform the image classification task and obtain the category to which the image to be identified belongs.

[0119] like Figure 4 As shown, the present invention also proposes a storage medium for storing a program that executes a model training method based on any of the DMA-MaaS federated learning platforms as described in claims 1 to 4.

[0120] The present invention also proposes a client for a model training system based on any one of the DMA-MaaS federated learning platforms as described in claims 5 to 8.

Claims

1. A model training method based on a DMA-MaaS federated learning platform, characterized in that, Comprise: Data preparation step, user end uploads training data to the federated learning platform, and sets whether the training data is public, the federated learning platform checks and adds the training data to the data pool; Task initiation step, the user end uploads the task to the federated learning platform, the task content includes task name, whether the task is public, whether the task type is federated learning, task required algorithm and parameter, task required model and parameter and task participant number; The federated learning platform adds the public task to the public task pool for other user ends to select; Task preparation step, the user end selects the task type of the selected task by selecting the task initiated by itself or selecting the task in the public task pool, if yes, execute federated learning step, otherwise execute non-federated learning step; The federated learning step, based on the training data, the federated learning is executed locally in the user device of the user end, the model parameters and results learned are returned to the federated learning platform for parameter aggregation, until the aggregated model reaches the required performance; The non-federated learning step, the federated learning platform executes non-federated learning in the cloud based on the training data, and obtains a model meeting the required performance.

2. The model training method based on the DMA-MaaS federated learning platform according to claim 1, wherein, The federated learning platform includes an algorithm pool, which contains federated learning algorithms and non-federated algorithms, the federated learning algorithms include horizontal federated learning algorithms, vertical federated learning algorithms, user preset federated learning and secure aggregation methods; The non-federated learning algorithm includes a stochastic gradient descent model optimization algorithm.

3. The model training method based on the DMA-MaaS federated learning platform according to claim 1, wherein, The federated learning platform includes a model pool, which includes a variety of preset models; The model management of the federated learning platform includes model listing, model downloading, model using, model charging, model market and model promotion; Model listing, downloading, using, charging and promotion are carried out on the model market, providing various model management for user ends; Among the existing data, tasks, algorithms and models in the federated learning platform, the task is initiated by the following method, and the platform will recommend the corresponding parameter configuration according to the user usage frequency: From the data pool: after the user sees the public data he wants in the data pool, he can initiate the task in the data page, at this time he also needs to select the corresponding algorithm and model, or select the algorithm and model recommended by the platform; Or From the algorithm pool: the user sees the algorithm he wants to apply for training in the algorithm pool, and can initiate the task in the algorithm page, at this time he also needs to select the corresponding data and model, or select the data and model recommended by the platform; Or From the model market: after the user sees the model he wants to use in the model market, he can initiate the task in the model page according to the need to initiate inference task or federated task; In inference task, only the data used by himself is needed to carry out model inference to get inference result; In federated task, the data and algorithm to be used are also needed to select the task, or the data and algorithm recommended by the platform are selected. 4.The method of claim 1, wherein, The model is an image classification model, the training data is an image with labeled categories, and the task is an image classification task; The model execution step is further included, and the image classification model is input with the image to be recognized to perform the image classification task, so as to obtain the category to which the image to be recognized belongs. 5.A model training system based on a DMA-MaaS federated learning platform, characterized in that, The method comprises the following steps: The data preparation module is configured to upload training data to a federated learning platform and set whether the training data is public, and the federated learning platform checks the training data and adds the training data to a data pool; The task initiation module is configured to upload a task to the federated learning platform, and the task content comprises a task name, whether the task is public, whether the task type is federated learning, a required algorithm and parameters of the task, a required model and parameters of the task, and a number of task participants; the federated learning platform adds a public task to a public task pool for selection by other user terminals; The task preparation module is configured to select a task initiated by the user terminal or a task in the public task pool, determine whether the selected task type is federated learning, and perform a federated learning module if the selected task type is federated learning, or perform a non-federated learning module if the selected task type is not federated learning; The federated learning module is configured to perform federated learning locally based on the training data on a user device of the user terminal, return learned model parameters and results to the federated learning platform for parameter aggregation, and perform the federated learning until the aggregated model reaches a required performance; The non-federated learning module is configured to enable the federated learning platform to perform non-federated learning in the cloud based on the training data, and obtain a model meeting the required performance.

6. The model training system based on the DMA-MaaS federated learning platform according to claim 5, wherein, The federated learning platform comprises an algorithm pool, and the algorithm pool comprises federated learning algorithms and non-federated algorithms; the federated learning algorithms comprise horizontal federated learning algorithms, vertical federated learning algorithms, user preset federated learning, and a secure aggregation method; the non-federated learning algorithms comprise a stochastic gradient descent model optimization algorithm.

7. The model training system based on the DMA-MaaS federated learning platform according to claim 5, wherein, The federated learning platform comprises a model pool, and the model pool comprises a plurality of preset models; model management of the federated learning platform comprises model uploading, model downloading, model use, model charging, a model market, and model promotion; model uploading, downloading, use, charging, and promotion are performed on the model market, and the model market provides diverse model management for user terminals; In the federated learning platform, the following methods are used to initiate a task based on existing data, tasks, algorithms, and models, and the platform recommends corresponding parameter configurations according to user usage frequency: A user initiates a task in a data page after seeing desired public data in a data pool, and needs to select corresponding algorithms and models, or select algorithms and models recommended by the platform; or A user initiates a task in an algorithm page after seeing desired algorithms for training in an algorithm pool, and needs to select corresponding data and models, or select data and algorithms recommended by the platform; or A user initiates a task in a model page after seeing desired models in a model market, and needs to initiate an inference task or a federated task according to needs; in the inference task, the user only needs to select data used by the user to perform model inference to obtain an inference result; in the federated task, the user needs to select data and algorithms to be used to initiate a task, or select data and algorithms recommended by the platform.

8. The model training system based on the DMA-MaaS federated learning platform of claim 5, wherein, The model is an image classification model, the training data is an image with a labeled category, and the task is an image classification task. The model execution module is further included, and the image classification model is input with an image to be identified to perform the image classification task, so as to obtain a category to which the image to be identified belongs.

9. A storage medium for storing a program for executing any one of the model training methods based on the DMA-MaaS federated learning platform according to claims 1 to 4.

10. A client for any one of the model training systems based on the DMA-MaaS federated learning platform according to claims 5 to 8.

Images