A federated learning model training method and system based on a container orchestration platform
By employing a container orchestration platform and differential privacy technology in federated learning, an isolated operating environment and dynamic resource management are constructed, solving the problems of privacy leakage and resource management efficiency in multi-client environments, and achieving efficient and secure model training.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA TOWER CO LTD
- Filing Date
- 2026-01-08
- Publication Date
- 2026-06-02
AI Technical Summary
Existing federated learning solutions fail to effectively combine containerized deployment and privacy protection in multi-customer environments, resulting in high privacy leakage risks, low resource management efficiency, and a lack of automated scheduling and scalability.
A container orchestration platform is used to create an isolated containerized runtime environment for each client. Differential privacy enhancement technology is combined to prune model gradients and add noise. Resource quotas and role-based access control are used to isolate tenants, dynamically select participants, and leverage the automated resource scheduling and coordination capabilities of the container orchestration platform.
It enables efficient, stable, and easily manageable large-scale collaborative model training in multi-client environments, improving privacy and resource management efficiency, and ensuring privacy protection for model parameter updates and the system's elastic scalability.
Smart Images
Figure CN122133846A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of federated learning, and specifically relates to a method for training federated learning models based on a container orchestration platform. Background Technology
[0002] With the deepening application of artificial intelligence technology in edge computing and the Internet of Things, massive amounts of data are constantly being generated on terminal devices. The traditional centralized machine learning paradigm requires these scattered raw data to be aggregated to a central server for unified processing and model training. This not only brings huge communication bandwidth pressure, but also causes serious privacy leakage risks due to the direct transmission of sensitive user data.
[0003] To achieve collaborative intelligence while protecting data privacy, federated learning has emerged as an innovative distributed machine learning framework. Under this framework, each participant trains the model locally using its own data and only uploads model parameter updates, rather than the original data, to the central server. The server then aggregates these updates to build a global model, thus theoretically achieving model movement without data movement. In addition, to address the complexity of deployment, scheduling, and resource management in distributed training tasks, container orchestration technologies, such as Kubernetes, have been introduced. These technologies can automatically manage multi-node clusters and provide elastic resource scheduling, fault recovery, and multi-client isolation mechanisms based on namespaces for computing tasks.
[0004] However, although federated learning avoids the direct transmission of raw data, research shows that attackers can still infer sensitive information in the training data by analyzing shared model gradients or parameters, posing a privacy risk. Common solutions often fail to deeply integrate with containerized deployment environments to form a systematic privacy protection system. At the same time, although Kubernetes can improve the efficiency of resource management, existing distributed training solutions based on this platform often focus on task deployment and resource isolation itself, failing to fully consider the complex coordination requirements of federated learning, such as multiple rounds of iteration, dynamic selection of participants, and balancing efficiency and privacy. They are inadequate in the organic combination of adaptive management of the training process, privacy enhancement technologies, and multi-client resource scheduling.
[0005] Therefore, how to build a federated learning system that can simultaneously guarantee strong privacy and security, support efficient automated scheduling and expansion, and be suitable for multi-client environments has become an urgent technical problem to be solved. Summary of the Invention
[0006] To address the aforementioned issues, this application provides a multi-tenant federated learning method based on a container orchestration platform, which has the advantages of achieving efficient, stable, and easily manageable large-scale collaborative model training in a multi-client environment.
[0007] A multi-tenant federated learning method based on a container orchestration platform, applied to a central server, includes the following: Based on a container orchestration platform, isolated containerized runtime environments are created for each client participating in federated learning; In each training round of the model iteration, a subset of clients is dynamically selected from each client as the participants in the current round, and the global model is distributed to each client in the client subset. Send federated learning tasks to clients and receive locally updated models from each client in a subset of clients. The locally updated models are obtained by each client in an isolated running environment after performing local training and privacy enhancement processing on the global model based on its local private dataset. Aggregate the received local updated models to generate an updated global model; Among them, the container orchestration platform provides cross-node automated resource scheduling and multi-tenant coordination support for the execution of methods.
[0008] Furthermore, based on the container orchestration platform, mutually isolated containerized runtime environments are created for each client, including: Each client is assigned an independent namespace, and resource and permission isolation between tenants is achieved through resource quotas and role-based access control.
[0009] This application provides a multi-tenant federated learning method based on a container orchestration platform, applied to a client, the method including: The federated learning client container runs in a mutually isolated containerized runtime environment created by a central server through a container orchestration platform. Receive the global model to be trained from the central server; The federated learning task sent by the central server runs in a containerized runtime environment, and the global model is trained locally based on the local private dataset. Privacy enhancement processing is performed on the locally updated model obtained from training; The locally updated model, after undergoing privacy enhancement processing, is uploaded to the central server to participate in the global model update.
[0010] Furthermore, based on the container orchestration platform, mutually isolated containerized runtime environments are created for each client, including: Each client is given an independent container instance, each configured with independent computing resources, network namespaces and storage volumes, and a pre-built federated learning client runtime environment is deployed within it.
[0011] Furthermore, the global model is trained locally based on a local private dataset, including: Receive training tasks from the central server and start container instances for the training tasks; Mount the directory storing the local dataset to the container instance; Extract sample data from the mounted dataset; Starting with the received global model parameters, multiple rounds of stochastic gradient descent training are performed using sample data to obtain the local model gradient.
[0012] Furthermore, privacy enhancement processing is performed on the locally updated model obtained through training, including: Prune the gradients of the local model obtained from local training; Add Gaussian noise to the clipped gradient to obtain the noisy gradient; The local model parameters are updated based on the noise-added gradient and the preset learning rate, thus obtaining the locally updated model.
[0013] This application also provides a federated learning system based on a container orchestration platform, the system including a central server deployed on the container orchestration platform and several clients; The central server, deployed on the management node of the container orchestration platform, coordinates global federated learning tasks, including: The tenant environment management module is used to create isolated runtime environments for each client participating in federated learning through a container orchestration platform; The client selection module is used to dynamically select a subset of clients from all clients as participants in each training round. The task distribution module is used to distribute the global model to each client among the participants; The model aggregation module is used to receive locally updated models uploaded by each participating party's client, and perform aggregation operations on the received local updated models to generate an updated global model. The client is deployed on the worker nodes of the container orchestration platform and runs in a mutually isolated runtime environment created by the tenant environment management module. Each client includes: The local training module is used to train the global model received from the central server locally based on a local private dataset. The privacy enhancement module is used to perform privacy enhancement processing on the locally trained model to obtain a locally updated model. The communication module is used to receive the global model from the central server and upload the locally updated model to the central server.
[0014] This application also provides an electronic device, which includes at least one processor and at least one memory, the memory being data-connected to the processor, wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform any of the methods described above.
[0015] This application also provides a computer-storeable medium storing computer instructions, which, when executed by a processor, specifically perform the steps of any of the methods described above.
[0016] This application also provides a computer program product, including computer instructions, which, when executed by a processor, specifically perform the steps in any of the methods described above.
[0017] Compared with the prior art, this application has the following advantages: This invention addresses the potential privacy risks associated with model parameter updates in federated learning, which could leak raw data. It mandates the introduction of a differential privacy enhancement mechanism into the client-side local training process, pruning and adding noise to model gradients. This ensures, from the data source, that even if parameter updates are intercepted or analyzed, no sensitive information can be effectively deduced. Simultaneously, the defined namespace, access control, and resource quota mechanisms based on a container orchestration platform provide strict logical and physical isolation for each client, preventing unauthorized cross-client data access at the system architecture level. This constructs a dual privacy protection system combining algorithm-level noise perturbation and system-level resource isolation, significantly improving the privacy and security level of federated learning in real-world multi-client scenarios.
[0018] Secondly, addressing the shortcomings of existing solutions in automated scheduling, scalability, and integration with the complex processes of federated learning, this invention uses a container orchestration platform as its core infrastructure. Leveraging its inherent automated resource scheduling and task coordination capabilities, it efficiently achieves unified management of massive heterogeneous client nodes, dynamic distribution of training tasks, and elastic scaling. Simultaneously, the server's dynamic selection of client subsets and coordination of training rounds, along with the design of the task scheduling module, enable the system to flexibly adapt to the multi-round iterations and dynamic participant selection requirements unique to federated learning. This deep integration gives the federated learning training process the same deployment elasticity, automated operation and maintenance, and horizontal scaling capabilities as cloud-native applications, thereby achieving efficient, stable, and easily manageable large-scale collaborative model training in multi-client environments.
[0019] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 A flowchart according to an embodiment of this application is shown; Figure 2 A structural block diagram according to an embodiment of this application is shown; Figure 3 A flowchart illustrating the local training and privacy enhancement process according to an embodiment of this application is shown. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0023] This invention relates to the field of distributed machine learning technology, and more specifically to a federated learning method, system, and device for privacy enhancement based on a container orchestration platform in a multi-tenant scenario.
[0024] With the rapid development of edge computing and the Internet of Things (IoT), data on terminal devices is exploding. Traditional centralized machine learning paradigms require the aggregation of scattered raw data to a central server for processing, which brings huge communication overhead and serious privacy risks. Federated learning, as a distributed machine learning framework, allows participants to share only model updates rather than raw data, thereby achieving collaborative modeling while protecting data privacy. However, existing federated learning solutions face significant challenges in practical deployment. First, as the number of participants (tenants, clients) increases, system deployment, scaling, and task management become complex and inefficient, lacking automated and flexible resource coordination mechanisms. Second, although raw data is not directly transmitted, shared model updates may still leak sensitive information through reverse analysis. Existing privacy protection schemes are often independent of the training framework and fail to deeply integrate with the underlying resource management system to form a systematic privacy protection system.
[0025] To address the aforementioned problems, this invention proposes a multi-tenant federated learning scheme based on a container orchestration platform. The core inventive concept of this application lies in the deep engineering integration of the automated resource scheduling and multi-tenant isolation capabilities of the container orchestration platform with the distributed training paradigm of federated learning and localized differential privacy technology, constructing an integrated, secure, and efficient learning platform. Specifically, the method includes steps executed collaboratively by a central server and multiple clients.
[0026] To make the technical solution of the present invention clearer, the specific embodiments are described step by step with reference to the accompanying drawings. Those skilled in the art should understand that these specific step descriptions are intended to provide full disclosure rather than limitation.
[0027] S1. System initialization and creation of a multi-tenant isolation environment.
[0028] Please see Figure 1 The central server, based on a container orchestration platform, allocates an independent namespace to each client (tenant) participating in federated learning. It also achieves logical isolation of computing resources, network, and storage through resource quotas and role-based access control policies. Simultaneously, the central server initializes the parameters of the global machine learning model to be trained. .
[0029] This step provides a secure, independent, and resource-controlled sandbox for subsequent training tasks, which is the foundation for achieving secure collaboration and efficient resource management in a multi-tenant environment.
[0030] S2. Start the federated learning iterative loop.
[0031] The system sets a total number of training rounds T and begins an iterative process from t=1 to T. Each iteration aims to utilize local data from a portion of the clients to collectively improve the global model.
[0032] S3. Dynamically select participating clients.
[0033] At the start of the t-th iteration, the central server dynamically selects a subset of clients of size K from all clients using a preset sampling strategy, preferably through random and uniform sampling. As a participant in this round of training, this partial participation mechanism significantly reduces the communication and computational overhead of a single round, and is a key design feature that supports large-scale client expansion.
[0034] S4. Distribute the global model and training tasks.
[0035] The central server will display the current global model parameters. The training configuration, such as the learning rate and the number of local iterations, is encapsulated into containerized tasks and precisely distributed to subsets through the scheduler of the container orchestration platform. Within the isolated namespace corresponding to each client, we ensure automated, highly available, and elastic task deployment.
[0036] S5, client-side local training and privacy enhancement processing.
[0037] Please see Figure 3 This step is executed in parallel by all selected clients in their respective isolated environments, and specifically includes the following sub-steps: S5-1: The client node pulls the federated learning client image, starts the container instance, and mounts its local private dataset storage directory into the container to achieve secure data access.
[0038] S5-2. Perform local iterative training. Starting from the received global model ω_t, the client repeats the following steps E times.
[0039] S5-2-1. Sample extraction: Randomly extract a small batch of samples from the mounted dataset.
[0040] S5-2-2, Calculate the gradient: Based on the current local model parameters and this batch of samples, calculate the gradient of the loss function. g.
[0041] S5-2-3, Gradient clipping: Performs a clipping operation on the gradient, restricting its L2 norm to within a preset threshold C, thus obtaining the gradient.
[0042] S5-2-4. Add Gaussian noise: Add random noise sampled from the distribution to the clipped gradient to obtain a noise-reduced gradient.
[0043] S5-2-5, Update local parameters. Update the local model parameters using the noisy gradient θ and the learning rate η.
[0044] Furthermore, step S5-2 above can be represented by the following algorithm: Input: Gradient clipping threshold Local dataset Differential privacy noise amplitude Learning rate Local iteration count loss function Initial local parameters ; Output: worker node The local model parameters obtained from this round of iterative training ; 1. do: 2. ; 3.
[0045] 4.
[0046] 5. ; 6.
[0047] Among them, represents the gradient clipping operation to ensure that the gradient norm does not exceed the threshold C, represents the updated model parameters obtained after the local training of client k is completed, represents the secure gradient after clipping and adding noise, represents a Gaussian distribution with a mean of 0 and a variance of σ², which is used to generate differential privacy noise, represents the gradient of the loss function calculated by client k on the sample x.
[0048] S5-3. Generate the local updated model. After completing E local iterations, the client obtains the final local updated model parameters ω_{t+1}^k for this round. This model has incorporated local data knowledge, and its update trajectory is protected by privacy due to noise injection.
[0049] S6. Upload the local updated model.
[0050] Each participating client uploads the processed local updated model to the central server through a secure channel.
[0051] S7. Aggregate the global model. The central server collects the model updates uploaded by all participating clients in this round and uses an aggregation algorithm such as the weighted average algorithm for fusion to generate a new round of global model.
[0052] S8. Iterative judgment.
[0053] The central server determines whether the current iteration round t has reached the preset total round T. If t < T, then set t = t + 1, update the global model to ω{t+1}, and return to step S3 to start the next round of iteration; if t = T, then execute step S9.
[0054] S9. Output the final model.
[0055] The federated learning training is completed, and the central server outputs the final global model ω_T. This model has incorporated the characteristics and laws of multi-party data while protecting the data privacy of each participating party.
[0056] Furthermore, the above steps S1-S9 can be represented as the following algorithm: Input: the number of worker tenant nodes , the number of iterative communication rounds of federated learning; Output: Global model parameters ; Initialization: Initialize global parameters ; 1. for t do; 2. The number of random selections is worker tenant node batch ; 3. Traverse worker tenant node batches Each worker node in do; 4.
[0058] 6. Parameter Aggregation
[0059] 7.
[0060] Where K represents the number of clients randomly selected in each round of training.
[0061] Through the above steps, this invention fully realizes automated task scheduling, multi-tenant resource isolation, and differential privacy protection embedded locally in the training process based on a container orchestration platform. The entire process systematically solves the core challenges of traditional federated learning in terms of scalability, management complexity, and privacy security.
[0062] See Figure 1 The system architecture of this invention includes a central server deployed on a container orchestration platform (such as a Kubernetes cluster) and several clients. The central server is deployed on a management node, and its core is the integration of multiple functional modules. The tenant environment management module is responsible for calling the container orchestration platform interface, creating an independent namespace for each client, and achieving dual isolation of physical resources and logical permissions through resource quotas and role-based access control policies. This ensures the independence and security of task execution for different tenants and avoids cross-tenant data leakage and resource conflicts.
[0063] The client selection module is responsible for dynamically selecting a subset from all clients in each training round according to a preset strategy (such as random uniform sampling). This partial participation mechanism significantly improves training efficiency in large-scale scenarios. The task distribution module is responsible for encapsulating the current global model and scheduling it to the selected clients. The model aggregation module is responsible for receiving and aggregating updates uploaded by clients. Clients are deployed on worker nodes and run in the aforementioned isolated environment. Each client contains a local training module, a privacy enhancement processing module, and a communication module, which work together to complete local computation and privacy protection tasks.
[0064] Combination Figure 2 The workflow shown involves the central server's tenant environment management module allocating an independent namespace to each registered client through a container orchestration platform, and configuring corresponding resource quotas and access permissions. This completes the construction of an isolated environment, providing a containerized and isolated execution sandbox for subsequent secure and parallel training tasks, greatly simplifying the management of complex environments. Subsequently, the system enters an iterative training loop. In each iteration, the client selection module dynamically selects a subset of clients to participate in the current training round. The task distribution module then uses the scheduling capabilities of the container orchestration platform to efficiently and reliably distribute the latest global model and training task parameters to each selected client. This process fully leverages the automated deployment and elastic scaling advantages of the container orchestration platform. Key processes on the client side are as follows: Figure 3 As shown, after receiving the task, the client first starts a container instance with a pre-configured federated learning framework in its own namespace and mounts the local data storage directory into the container, thus enabling data access that is available but not visible.
[0065] The local training module then extracts samples from the mounted dataset, takes the received global model as a starting point, performs multiple rounds of stochastic gradient descent training, and calculates the gradient of the model parameters.
[0066] The privacy enhancement processing module performs two steps on the calculated original gradient. First, gradient clipping is performed, which limits the L2 norm of the gradient using a preset threshold to prevent individual data samples from having too much influence on the model. This step is a key prerequisite for controlling the magnitude of subsequent noise addition. Then, random noise that conforms to a Gaussian distribution is added to the clipped gradient.
[0067] This combined cropping and noise-adding operation mathematically satisfies the definition of differential privacy, ensuring that even if an attacker obtains the processed gradient, it is difficult to infer information about any individual training sample, thus embedding strong privacy guarantees before the model update leaves the client. Based on this noisy, secure gradient and a preset learning rate, the client updates its local model parameters. After a preset number of local iterations, the final locally updated model is generated and uploaded back to the central server via the communication module.
[0068] After collecting updates from all participating clients in the current round, the model aggregation module of the central server aggregates these updates using algorithms such as federated averaging to generate a new generation of global models with superior performance. The system repeats this iterative cycle until a preset number of training rounds or model convergence criteria are reached. Through this integrated design, this invention not only achieves scalable and automated management of training tasks via a container platform, but also provides an end-to-end privacy protection solution for multi-tenant federated learning, from system architecture to algorithm level, through a local differential privacy mechanism built into the training process.
[0069] In one specific implementation, the central server runs on the control plane of a Kubernetes cluster. When three tenants (clients A, B, and C) participate, the tenant environment management module creates three namespaces: ns-a, ns-b, and ns-c, and binds resource quota objects that restrict CPU and memory usage, as well as role-based access control policies that only allow access to their respective namespaces.
[0070] In the first round of training, the client selection module randomly selected clients A and C as participants. The task distribution module then distributed the global model... The image was packaged into a mirror and deployed to ns-a and ns-c via K8sJob, with the local directories of each client mounted.
[0071] In ns-a, the client's container starts, reads samples from the local directory, begins stochastic gradient descent training, calculates the gradient g, and the gradient clipping module restricts its norm to C, obtaining g'. Then, noise n sampled from the distribution N(0,σ²) is added to obtain the safe gradient θ = g' + n. The local training module updates the model using θ and the learning rate η, resulting in the local update. .
[0072] Client C executes a similar process to obtain .
[0073] Subsequently, and The uploaded data is aggregated and calculated by the model aggregation module. =( + ) / 2, completing this iteration.
[0074] This example fully demonstrates the complete closed loop from isolated environment creation, dynamic selection, distribution, local privacy training to aggregation.
[0075] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. When executed, the program can implement the embodiments of the methods described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk (HDD), or solid-state drive (SSD), etc.; the program may include, for example,... Figures 1 to 3 The steps described.
[0076] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
[0077] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A multi-tenant federated learning method based on a container orchestration platform, applied to a central server, characterized in that... The method includes: Based on a container orchestration platform, isolated containerized runtime environments are created for each client participating in federated learning; In each training round of the global model iteration, a subset of clients is dynamically selected from each client as the participants in the current round, and the global model is distributed to each client in the subset of clients. Send federated learning tasks to clients and receive local updated models from each client in the client subset, wherein the local updated models are obtained by each client in the isolated running environment after performing local training and privacy enhancement processing on the global model based on its local private dataset; Aggregate the received local updated models to generate an updated global model; The container orchestration platform provides cross-node automated resource scheduling and multi-tenant coordination support for the execution of the method.
2. The method according to claim 1, characterized in that, The container orchestration platform creates isolated containerized runtime environments for each client, including: Each client is assigned an independent namespace, and resource and permission isolation between tenants is achieved through resource quotas and role-based access control.
3. A multi-tenant federated learning method based on a container orchestration platform, applied to a client, characterized in that... The method includes: The federated learning client container runs in a mutually isolated containerized runtime environment created by a central server through a container orchestration platform. Receive the global model to be trained from the central server; The federated learning task sent by the central server is run in a containerized runtime environment, and the global model is trained locally based on the local private dataset. Privacy enhancement processing is performed on the locally updated model obtained from training; The locally updated model, after undergoing privacy enhancement processing, is uploaded to the central server to participate in the global model update.
4. The method according to claim 2, characterized in that, The container orchestration platform creates isolated containerized runtime environments for each client, including: Each client is given an independent container instance, each configured with independent computing resources, network namespaces and storage volumes, and a pre-built federated learning client runtime environment is deployed within it.
5. The method according to any one of claims 1-4, characterized in that, The global model is trained locally based on a local private dataset, including: Receive training tasks from the central server and start container instances for the training tasks; Mount the directory storing the local dataset to the container instance; Extract sample data from the mounted dataset; Using the received global model parameters as the initial point, the local model gradient is obtained by performing multiple rounds of stochastic gradient descent training with the sample data.
6. The method according to any one of claims 1-4, characterized in that, Privacy enhancements to the locally updated model obtained through training include: Prune the gradients of the local model obtained from local training; Add Gaussian noise to the clipped gradient to obtain the noisy gradient; Based on the noise-added gradient and the preset learning rate, the local model parameters are updated to obtain the locally updated model.
7. A federated learning system based on a container orchestration platform, characterized in that, The system includes a central server deployed on a container orchestration platform and several clients; The central server is deployed on the management node of the container orchestration platform and is used to coordinate global federated learning tasks, including... The tenant environment management module is used to create mutually isolated runtime environments for each client participating in federated learning through the container orchestration platform. The client selection module is used to dynamically select a subset of clients from all clients as participants in each training round. The task distribution module is used to distribute the global model to each client among the participating parties; The model aggregation module is used to receive locally updated models uploaded by each participating party's client, and perform aggregation operations on the received local updated models to generate an updated global model. The client is deployed on a worker node of the container orchestration platform and runs in an isolated runtime environment created by the tenant environment management module. Each client includes: The local training module is used to train the global model received from the central server locally based on a local private dataset. The privacy enhancement module is used to perform privacy enhancement processing on the locally trained model to obtain a locally updated model. A communication module is used to receive the global model from the central server and upload the locally updated model to the central server.
8. An electronic device, characterized in that, The electronic device includes at least one processor and at least one memory, the memory being data-connected to the processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6.
9. A computer-storable medium, characterized in that, The storable medium stores computer instructions, which, when executed by a processor, specifically perform the steps of the method as described in any one of claims 1-6.
10. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they specifically perform the steps in the method as described in any one of claims 1-6.