Federal learning communication method and system based on Poisson flow generation model
By introducing the Poisson flow generation model and lottery assumption in federated learning, the problems of non-independent and homogeneous distribution of data and high communication overhead are solved, the model is efficiently trained and stable, and the performance and efficiency of the federated learning system are optimized.
Patent Information
- Application Number
- CN202510556832.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-29
- Publication Date
- 2025-08-15
AI Technical Summary
The existence of non-independent homogeneous distribution of data (Non-IID) problems in federated learning leads to poor generalization performance, high communication overhead and model redundancy problems, affecting the performance and efficiency of the model.
The Poisson flow generation model is introduced in combination with lottery assumptions, and through data augmentation and model pruning, data distribution is optimized and redundant parameters are reduced. The Poisson flow generation model is used for data augmentation and combined with lottery assumption pruning model, optimize communication and computing efficiency.
Effectively alleviate the Non-IID problem, improve the model convergence performance, reduce communication burden and computing cost, and improve the stability and efficiency of the federated learning system.
Smart Images
Figure CN120494127A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information security technology, and in particular to a federated learning communication method and system based on a Poisson flow generation model. Background Art
[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.
[0003] With the increasing demand for data privacy protection, federated learning (FL), as a distributed training method, can fully utilize data distributed across devices to build global models while protecting data privacy. However, federated learning faces numerous challenges during training, the most prominent of which is the non-independent and identically distributed (Non-IID) nature of the data. Due to significant differences in data distribution across devices, this heterogeneity can lead to inconsistent learning results across devices, resulting in model drift, which reduces the convergence speed and accuracy of the global model. In severe cases, the server may even be unable to build an efficient and stable global model.
[0004] To address the non-IID problem, the introduction of generative models has become a hot research topic. Generative models can generate realistic synthetic data by learning the underlying distribution of the data, thereby alleviating the impact of distribution inconsistency. However, traditional generative models have inherent flaws. For example, generative adversarial networks (GANs) are prone to mode collapse and training oscillation, making them unable to generate stable samples. Variational autoencoders (VAEs) often suffer from artifacts and ambiguity in the generated samples during data distribution recovery and loss function calculation. Diffusion models (DMs) also have some significant limitations and shortcomings, particularly in terms of computational efficiency and communication overhead. These shortcomings limit the practical application of generative models in federated learning, and there is an urgent need to develop more efficient and stable generative models.
[0005] Therefore, federated learning faces the following technical problems in practical applications: (1) Non-IID data problem: The data distribution of different clients may be significantly different, which will lead to poor generalization performance of the global model and affect the performance of the model; (2) High communication overhead: Due to the large number of model parameters, frequent communication between the client and the server will occupy a lot of bandwidth; (3) Model redundancy problem: Many parameters in large-scale neural networks may not be necessary for training, which increases computing and storage costs. Summary of the Invention
[0006] To overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a federated learning communication method and system based on the Poisson flow generation model. Combining the Poisson flow generation model with the lottery hypothesis, the dynamic data supplement capability of the Poisson flow generation model and the sparse characteristics of the lottery hypothesis form a good complementary relationship, which not only improves the personalized performance of the model, but also optimizes the communication and computing efficiency of the system, and is suitable for large-scale heterogeneous distributed federated learning environments.
[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides a federated learning communication method based on a Poisson flow generation model, comprising: Constructing a Poisson flow generation model, and performing federated training on the Poisson flow generation model based on a federated learning system to obtain a trained Poisson flow generation model; the federated learning system includes a server and multiple clients; Based on the trained Poisson flow generation model, data enhancement is performed on each client's local dataset to obtain an enhanced local dataset; The server randomly initializes the global model and sends the global model and pruning mask to the corresponding client; the client generates a pruned model based on the corresponding pruning mask, trains the pruned model using the enhanced local dataset and updates the model parameters, and sends the updated model parameters to the server; the server receives the model parameters sent by each client, aggregates them, and iteratively updates the global model to obtain a converged global model.
[0008] A further technical solution is to perform federated training on the Poisson flow generation model, including a forward process and a backward process.
[0009] A further technical solution is that the forward process includes client local operations and server aggregation to update the Poisson field parameters in the Poisson flow generation model; the backward process includes client steps and server steps, and an attention mechanism is introduced to dynamically adjust the direction of data flow.
[0010] In a further technical solution, the local operations of the client in the forward process are specifically as follows: Expand local data into high-dimensional space to obtain expanded local data; Based on the expanded local data, a Poisson field is constructed to obtain the Poisson flow generation model and its Poisson field parameters; Based on the Poisson field parameters, the data flow is simulated to obtain uniformly distributed local data; The Poisson field is optimized according to the evenly distributed local data, and updated Poisson field parameters are obtained and uploaded to the server.
[0011] Further technical solutions, the server aggregation in the forward process is specifically as follows: The server performs weighted aggregation on the Poisson field parameters uploaded by all clients to construct a global Poisson field and obtain the global Poisson field parameters; The global Poisson field parameters are distributed back to each client.
[0012] In a further technical solution, the client steps in the backward process are specifically as follows: Each client generates perturbation samples using the perturbation function; Calculate the normalized gradient for each perturbation sample, and introduce an attention mechanism into the normalized gradient to dynamically adjust the gradient direction and intensity; A loss function for each client is defined, and the Poisson field parameters are updated to minimize the loss function.
[0013] In a further technical solution, the server-side steps in the backward process are specifically as follows: The server receives the Poisson field parameters uploaded by all clients and performs global aggregation to obtain the global Poisson flow generation model parameters; A global loss function is defined, and the global Poisson flow generation model parameters are updated based on the global loss function. An attention mechanism optimization model is introduced to obtain optimized global Poisson flow generation model parameters, thereby obtaining a trained Poisson flow generation model.
[0014] In a second aspect, the present invention provides a federated learning communication system based on a Poisson flow generation model, comprising: A generation model training module is configured to: construct a Poisson flow generation model, and perform federated training on the Poisson flow generation model based on a federated learning system to obtain a trained Poisson flow generation model; the federated learning system includes a server and multiple clients; A data enhancement module is configured to: perform data enhancement on each client's local dataset based on the trained Poisson flow generation model to obtain an enhanced local dataset; The functional model training module is configured as follows: the server randomly initializes the global model and sends the global model and pruning mask to the corresponding client; the client generates a pruned model according to the corresponding pruning mask, uses the enhanced local data set to train the pruned model and update the model parameters, and sends the updated model parameters to the server; the server receives the model parameters sent by each client, aggregates them, and iteratively updates the global model to obtain a converged global model.
[0015] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the federated learning communication method based on the Poisson flow generation model as described in the first aspect.
[0016] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the federated learning communication method based on the Poisson flow generation model as described in the first aspect are implemented.
[0017] One or more of the above technical solutions have the following beneficial effects: This paper introduces a generative model based on Poisson flow into federated learning, effectively alleviating the challenges posed by non-IID data. Furthermore, to further optimize communication efficiency, the Poisson flow-based federated learning communication method can be combined with the Lottery Ticket Hypothesis (LTH). Through pruning, high-performing subnetworks are identified, achieving performance close to that of the original network at a smaller scale, thereby reducing redundant information transmission during model updates. This approach not only improves the convergence performance of federated learning but also provides a new direction for building more efficient and stable federated learning systems.
[0018] In federated learning, combining the Poisson flow generative model with the lottery ticket hypothesis can significantly optimize model sparsity, personalization, and communication efficiency. Specifically, the Poisson flow generative model can be used for client data generation modeling, helping to supplement missing data categories in small-sample clients and generate sparse stream data, thereby improving data diversity and representativeness. This data augmentation approach effectively mitigates the performance degradation caused by non-independent and identically distributed (Non-IID) client data distribution in federated learning, significantly improving training results for clients with insufficient sample sizes.
[0019] The application of the Lottery Ticket Hypothesis in federated learning focuses on model pruning and sparsification. Using the Lottery Ticket Hypothesis, we can identify sparse subnetworks with excellent performance within the global model, significantly reducing the number of model parameters while maintaining model accuracy. This sparsification strategy directly reduces the communication burden of clients uploading model updates and servers distributing model parameters. This pruning method significantly improves communication efficiency and reduces overall system operating costs, especially for bandwidth-constrained edge devices and resource-constrained weak clients. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0021] Figure 1 This is a diagram of the overall framework of federated learning in the embodiment of the present invention; Figure 2Flowchart of the forward process and reverse process of the Poisson flow model in the embodiment of the present invention; Figure 3 This is a flow chart of the federated learning training phase according to an embodiment of the present invention; Figure 4 This is a lottery hypothetical flow chart of an embodiment of the present invention. DETAILED DESCRIPTION
[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.
[0023] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0024] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.
[0025] Example 1 like Figure 1 As shown, this embodiment discloses a federated learning communication method based on a Poisson flow generation model, which includes the following steps: In this embodiment, the federated learning system includes a server and multiple clients. First, N federated learning clients, E local training rounds, and S server rounds are set up. Then, the entire experimental data set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. Then, the training set is used to generate a Non-IID data set according to the Dirichlet distribution, and finally distributed to each client as local data. Among them, 80% of the data is used for model training, 10% is used for model performance verification and hyperparameter tuning, and the other 10% is used for the final test evaluation to comprehensively measure the generalization ability and actual performance of the model. In order to simulate the actual situation of uneven data distribution among the clients in federated learning, the Dirichlet distribution is introduced. To generate a non-independent and identically distributed (Non-IID) data set, where the parameters Control the distribution uniformity of data: The smaller it is, the higher the degree of non-independent and identically distributed data between clients; The larger it is, the closer the data distribution is to independent and identically distributed (IID).
[0026] S1: Construct a Poisson flow generation model, and perform federated training on the Poisson flow generation model based on a federated learning system to obtain a trained Poisson flow generation model; In this embodiment, if Figure 1 As shown in the first stage (Stage 1) in Figure 2, combining the characteristics of federated learning and Poisson flow generation model, the federated training Poisson flow generation model stage can be divided into a forward process and a backward process.
[0027] Forward process: The forward process of the Poisson flow generation model is to construct a mapping relationship between data distribution and high-dimensional uniform distribution by simulating the Poisson field in the physical world. Specifically, the data distribution is regarded as a charge distribution, and then the data is repelled to a uniform distribution (high-dimensional hemisphere) under the action of its own electric field.
[0028] (1) Client local operation Data dimensionality increase: Each client will local data Adding a new dimension , expand the original data To higher dimensional space , where the initial dimension Set to zero, here It is a newly added dimension variable used to map data to a higher-dimensional space. The purpose of increasing the dimension is to provide additional degrees of freedom for the flow of data and simplify the mapping process from the original distribution to the uniform distribution.
[0029] Constructing a Poisson field: Each client constructs a Poisson field based on local data. Constructing a Poisson field is a key step in the Poisson flow generation model. Its core goal is to define the potential energy field (potential field) of the data distribution through the Poisson equation, thereby providing a dynamic basis for subsequent data flow. The definition of the Poisson equation is specifically expressed as:
[0030] in, It is expressed as a Poisson potential field, describing the potential energy of the data points; Indicates the density distribution of local data on the client.
[0031] The gradient of the Poisson field is defined as:
[0032] in, Indicates the flow direction and speed of data points.
[0033] Simulate data flow: Figure 2As shown in the Backward phase, the core purpose of this phase is to map the data from the original complex distribution to a high-dimensional uniform distribution through a potential field-driven dynamic process, providing a physical basis for the Poisson flow generation model to generate data in reverse. The flow process is controlled by an ordinary differential equation (ODE), which can be expressed as:
[0034] in, is the time parameter, which represents the flow process.
[0035] Data Points At each time step Moving along the direction of the electric field, as the flow proceeds, the data gradually spreads from the original distribution to the hemispherical surface of the high-dimensional space, forming a more uniform distribution.
[0036] Poisson field parameters (model parameters of the Poisson flow generation model) update: the client adjusts the local data density function based on the data flow results , and then optimize the Poisson potential field , update the Poisson field parameters.
[0037] (2) Server-side aggregation Upload local parameters: Each client uploads the updated Poisson field parameters (such as gradient ) to the server.
[0038] Constructing a global Poisson field: The server performs weighted aggregation on the Poisson field parameters uploaded by all clients, expressed as:
[0039] in, Represents the client The weight is usually proportional to the amount of data; represents the gradient of the global Poisson field. Represents the client The flow direction and speed of the data points in .
[0040] Distribute global parameters: The aggregated global Poisson field parameters are distributed back to each client to provide guidance for the next round of forward process.
[0041] Backward Process: In a federated learning scenario, each client trains its local data and uploads the local gradients to the server, which aggregates and updates the global Poisson flow generation model. The backward process uses the attention mechanism to dynamically adjust the direction of data flow, specifically in gradient normalization and loss calculations.
[0042] (1) Client steps Each client Data distribution , the goal is to optimize the loss function through back propagation , guide the data to be mapped from uniform distribution to target distribution ,in Represents conditional information, which is represented as the label of the image in the image classification task.
[0043] Data perturbation: Use the perturbation function (perturb) to generate perturbation samples, expressed as:
[0044] in, Indicates the Data on the client, The output is , which includes the perturbed data and the extra dimension.
[0045] Gradient normalization (introducing attention mechanism): for each perturbation sample , the calculation of normalized gradient is specifically defined as:
[0046] in, Represents a large batch of data The calculated normalized Poisson field, Indicates the number of clients, is the gradient of the Poisson field; is a normalization constant to prevent numerical instability.
[0047] The attention mechanism is used in the normalized gradient to dynamically adjust the gradient direction and intensity, which can be defined as:
[0048] in, Represents the attention weight, which is determined by the condition and data Decide.
[0049] Loss calculation: define the loss function for each client The error between the generated distribution and the target distribution can be defined as:
[0050] in, Represents large batches of data, is the cross entropy loss, ensuring that the category of the generated sample is , is the balance coefficient.
[0051] Parameter update: based on learning rate Update the model parameters (i.e., Poisson field parameters), which can be defined as: This formula minimizes the client's loss function and gradually optimizes the performance of the entire model through global server aggregation.
[0052] (2) Server-side steps In federated learning, the server accepts model parameters uploaded by the client. , and then perform global aggregation, and then weighted aggregation into global model parameters.
[0053] Global loss function: This loss function is the weighted average of the loss functions of each client, which can be defined as:
[0054] in, Is the client The weight of represents the global loss function.
[0055] Global gradient update: The server updates the model parameters based on the global loss function: ,in, Represents the parameters of the Poisson flow generation model, usually using the NSCN++ architecture as the model The trunk.
[0056] Attention mechanism optimization: In the Poisson flow generation model, the attention mechanism parameters The optimization is performed through the back-propagation process. Its core goal is to make the Poisson flow generation model dynamically focus on key features (specifically conditional information ), thereby achieving directional generation control of data. The optimized attention mechanism can effectively guide the flow of data from the original distribution to the target distribution, improving the accuracy and controllability of the generated results. Its gradient is expressed as:
[0057] in, Indicates the Poisson flow loss function for each client, Represents the output of the attention function, representing the data point In the conditions The feature weights below.
[0058] S2: Based on the trained Poisson flow generation model, data enhancement is performed on each client's local dataset to obtain an enhanced local dataset; In this embodiment, if Figure 1 As shown in the second stage (Stage 2), it is assumed that the data label set of each client is , whose missing tags are ,in, is the set of all tags for each client, Indicates the The above variables are calculated when training the Poisson flow generation model in the first phase, and then generated for the client by the Poisson flow generation model in the second phase. The data is input, that is, the label of the category missing from the client, and then the Poisson flow generation model generates the data corresponding to the label, and finally the enhanced local data set is obtained .
[0059] like Figure 1 As shown in the figure, the purpose of the Poisson flow generation model is to balance the data of federated learning clients in order to alleviate the problem of model performance degradation in federated learning training caused by data heterogeneity.
[0060] S3: The server initializes the global model and sends the global model and pruning mask to the corresponding client; the client generates a pruned model based on the pruning mask, uses the enhanced local dataset to train the pruned model and updates the model parameters, and sends the updated model parameters to the server; the server receives the model parameters sent by each client, aggregates them, and updates the global model to obtain a converged global model.
[0061] In this embodiment, if Figure 1 As shown in the third stage (Stage 3), the application of Lottery Hypothesis in federated learning is to improve communication efficiency and performance by pruning the model, while also taking into account the training of personalized models, such as Figure 3 As shown in Figure 2, the pruning model can not only reduce redundant parameters in the model but also find a balance between local and global performance. Therefore, applying the lottery hypothesis in federated learning can solve the Non-IID problem of federated learning.
[0062] like Figure 4 As shown, the training steps are as follows: Step 1: Initialize the global model on the server The server randomly initializes a complete global model (i.e. classification model) (Use CNN network, such as ResNet-18) as the basic model for classification tasks, initialize the global uniform mask, and the initial mask is all 1, indicating that all parameters are retained; Step 2: Select a client subset The server in each round During communication, a portion of clients are randomly selected Participate in training; Step 3: Download the model and prune The server will be the global model and pruning mask Send to the client , each client generates a local model based on its own pruning mask, which can be defined as: , where denotes element-wise multiplication, mask Used to select which parameters are retained. Note that the pruning mask calculation here is generated using an element-by-element exponential function, and its mathematical expression is:
[0063] in, is a classification model; is a threshold value, which is usually determined by the target pruning rate; for example, if the goal is to prune off 20% of the minimum weights, then Can be set so that 20% of the parameters meet , It is an indicator function that returns 1 when the condition in the brackets is met, otherwise it returns 0. The result is a mask tensor with the same dimension as the model parameter, 1 means that the parameter is retained, and 0 means that the parameter is pruned. Specifically, the importance estimate of the model parameter is calculated first, and when this value exceeds the threshold It will be retained, otherwise it will be set to 0.
[0064] Step 4: Client pruning and retraining Client uses the validation dataset Test the performance of the current pruned model. If the pruning rate Failure to achieve goals , and verify the accuracy If the accuracy threshold is met, the mask is regenerated And reset the model parameters. After the model is successfully pruned, set the initial parameters of the pruned model to the parameters of the original model, and then use the local training set Pruning Model Train and update the parameters to .
[0065] The model accuracy is tested on a validation dataset. If the pruning rate reaches a preset value and the accuracy exceeds a threshold, the final pruned model is obtained. The relationship between the local model and the pruned model is as follows: through iterative parameter pruning, a pruned model with accuracy close to that of the original model (which has a large number of parameters) is obtained, thereby reducing the amount of data transmitted by federated learning.
[0066] Step 5: Model upload and server aggregation The client will update the model parameters Send to the server, the server receives the pruned model Perform aggregation and update the global model, which is defined as: .
[0067] Step 6: Repeat the iteration until convergence Repeat the above steps until the target number of communication rounds is reached or the performance of the global model converges.
[0068] The Poisson Flow Generative Model (PFGM), with its physics-inspired approach to modeling probability flows, offers a new approach to efficient data generation. PFGM relates the data generation process to the Poisson equation and, by defining the evolutionary path of the probability flow in high-dimensional space, generates realistic and stable samples. Compared to traditional generative models, PFGM offers significant advantages in computational efficiency, generated sample quality, and model stability. PFGM also supports inference using deterministic ordinary differential equations (ODEs), reducing communication overhead while maintaining high generation quality.
[0069] Federated learning is a distributed machine learning framework that enables multi-party collaborative training while ensuring data privacy and security. However, in practical applications, federated learning faces the following technical challenges: (1) Data non-independent and identically distributed (Non-IID) problem: Data distributions of different clients may differ significantly, which leads to poor generalization performance of the global model and affects model performance; (2) High communication overhead: Due to the large number of model parameters, frequent communication between the client and the server consumes a large amount of bandwidth; (3) Model redundancy problem: Many parameters in large-scale neural networks may not be necessary for training, which increases computational and storage costs. To address the above issues, the present invention uses Poisson flow to generate model synthetic data to alleviate the data non-independent and identically distributed (Non-IID) problem; secondly, it uses the lottery hypothesis to perform model pruning to reduce communication overhead and computational cost.
[0070] In summary, the first aspect of this invention uses a Poisson flow generation model to enhance non-IID data, balancing data categories across clients. This category balance benefits by accelerating model convergence during federated learning training, reducing the number of communication rounds, and improving overall efficiency. Secondly, the client utilizes the lottery hypothesis: a "winning network" is selected from the classification model, meaning a model with comparable performance to the original network but with fewer parameters. This reduces the communication volume per round of federated learning.
[0071] Example 2 This embodiment discloses a federated learning communication system based on a Poisson flow generation model, including: A generation model training module is configured to: construct a Poisson flow generation model, and perform federated training on the Poisson flow generation model based on a federated learning system to obtain a trained Poisson flow generation model; the federated learning system includes a server and multiple clients; A data enhancement module is configured to: perform data enhancement on each client's local dataset based on the trained Poisson flow generation model to obtain an enhanced local dataset; The functional model training module is configured as follows: the server randomly initializes the global model and sends the global model and pruning mask to the corresponding client; the client generates a pruned model according to the corresponding pruning mask, uses the enhanced local data set to train the pruned model and update the model parameters, and sends the updated model parameters to the server; the server receives the model parameters sent by each client, aggregates them, and iteratively updates the global model to obtain a converged global model.
[0072] Example 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of embodiment 1 when executing the program.
[0073] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, which performs the steps of the method of embodiment 1 when executed by a processor.
[0074] The steps involved in the apparatuses of Examples 3 and 4 above correspond to those of Method Example 1. For detailed implementation, please refer to the relevant description of Example 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any of the methods of the present invention.
[0075] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.
[0076] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.
[0077] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.
Claims
1. A federated learning communication method based on a Poisson flow generation model, characterized in that: include: Constructing a Poisson flow generation model, and performing federated training on the Poisson flow generation model based on a federated learning system to obtain a trained Poisson flow generation model; The federated learning system includes a server and multiple clients; Based on the trained Poisson flow generation model, data enhancement is performed on each client's local dataset to obtain an enhanced local dataset; The server randomly initializes the global model and sends the global model and pruning mask to the corresponding client; the client generates a pruned model based on the corresponding pruning mask, trains the pruned model using the enhanced local dataset and updates the model parameters, and sends the updated model parameters to the server; the server receives the model parameters sent by each client, aggregates them, and iteratively updates the global model to obtain a converged global model.
2. The federated learning communication method based on the Poisson flow generation model according to claim 1, characterized in that: The federated training of the Poisson flow generation model includes a forward process and a backward process.
3. The federated learning communication method based on the Poisson flow generation model according to claim 2, characterized in that: The forward process includes client local operations and server aggregation, and updates the Poisson field parameters in the Poisson flow generation model; the backward process includes client steps and server steps, and introduces an attention mechanism to dynamically adjust the direction of data flow.
4. The federated learning communication method based on the Poisson flow generation model according to claim 3, characterized in that: The local operations of the client in the forward process are specifically as follows: Expand local data into high-dimensional space to obtain expanded local data; Based on the expanded local data, a Poisson field is constructed to obtain the Poisson flow generation model and its Poisson field parameters; Based on the Poisson field parameters, the data flow is simulated to obtain uniformly distributed local data; The Poisson field is optimized according to the evenly distributed local data, and updated Poisson field parameters are obtained and uploaded to the server.
5. The federated learning communication method based on the Poisson flow generation model according to claim 3, characterized in that: The server-side aggregation in the forward process is specifically as follows: The server performs weighted aggregation on the Poisson field parameters uploaded by all clients to construct a global Poisson field and obtain the global Poisson field parameters; The global Poisson field parameters are distributed back to each client.
6. The federated learning communication method based on the Poisson flow generation model according to claim 3, characterized in that: The client steps in the backward process are specifically as follows: Each client generates perturbation samples using the perturbation function; Calculate the normalized gradient for each perturbation sample, and introduce an attention mechanism into the normalized gradient to dynamically adjust the gradient direction and intensity; A loss function for each client is defined, and the Poisson field parameters are updated to minimize the loss function.
7. The federated learning communication method based on the Poisson flow generation model according to claim 3, characterized in that: The server-side steps in the backward process are specifically as follows: The server receives the Poisson field parameters uploaded by all clients and performs global aggregation to obtain the global Poisson flow generation model parameters; A global loss function is defined, and the global Poisson flow generation model parameters are updated based on the global loss function. An attention mechanism optimization model is introduced to obtain optimized global Poisson flow generation model parameters, thereby obtaining a trained Poisson flow generation model.
8. A federated learning communication system based on a Poisson flow generation model, characterized in that: include: A generation model training module is configured to: construct a Poisson flow generation model, and perform federated training on the Poisson flow generation model based on a federated learning system to obtain a trained Poisson flow generation model; the federated learning system includes a server and multiple clients; A data enhancement module is configured to: perform data enhancement on each client's local dataset based on the trained Poisson flow generation model to obtain an enhanced local dataset; The functional model training module is configured as follows: the server randomly initializes the global model and sends the global model and pruning mask to the corresponding client; the client generates a pruned model according to the corresponding pruning mask, uses the enhanced local data set to train the pruned model and update the model parameters, and sends the updated model parameters to the server; the server receives the model parameters sent by each client, aggregates them, and iteratively updates the global model to obtain a converged global model.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the federated learning communication method based on the Poisson flow generation model as described in any one of claims 1 to 7 are implemented.
10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps in the federated learning communication method based on the Poisson flow generation model according to any one of claims 1 to 7 are implemented.