Privacy federal learning method based on fine-grained symbol gradient descent optimization

Through the privacy federated learning method combining fine-grained symbol gradient descent optimization and reinforcement learning and game theory strategies, the efficiency and privacy protection of federated learning in large-scale data processing is solved, and an efficient and secure learning process is achieved, adapting to a variety of ways to adapt to Data type and application requirements.

CN119090025APending Publication Date: 2024-12-06CETC DIGITAL INTELLIGENCE TECH (BEIJING) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411052507.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Existing federated learning technologies are inefficient in processing large-scale data and are difficult to adapt to different data types and application needs while ensuring data privacy, especially in fine-grained gradient optimization, communication efficiency and privacy protection.

Method used

Adopt the privacy federated learning method based on fine-grained symbol gradient descent optimization, by initializing and setting model parameters, introducing reinforcement learning and game theory strategies, dynamically adjusting learning rates and parameters, combining local differential privacy technology, optimizing loss functions to ensure model accuracy, A balance between efficiency and privacy protection.

Benefits of technology

It improves the efficiency and accuracy of model training, significantly improves communication efficiency and privacy protection capabilities, is suitable for efficient and secure learning of large-scale data sets, and expands the scope of application of federated learning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090025A_ABST
    Figure CN119090025A_ABST
Patent Text Reader

Abstract

The invention discloses a privacy federated learning method based on fine-grained symbol gradient descent optimization, and by introducing a fine-grained symbol gradient descent optimization method, the privacy federated learning method takes an important step in improving gradient updating efficiency, convergence efficiency and precision in federated learning. Compared with a common coarseness gradient descent method in the prior art, the machine learning convergence speed is increased, and the performance of the final model is improved. This is particularly important for multi-local-node federated learning and large-scale data set processing because it can significantly reduce training time and consumption of computing resources; and meanwhile, the method also shows obvious advantages in the aspect of privacy protection. By introducing an advanced privacy protection mechanism and local differential privacy, the method can maintain effective learning of the model while ensuring data security. The method is of great significance for popularizing federated learning in various application scenes with sensitive current data privacy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of federated learning, and more specifically, particularly relates to a privacy federated learning method based on fine-grained symbolic gradient descent optimization. Background Art

[0002] Federated learning, as a distributed machine learning method that protects data privacy, is rapidly gaining popularity. It allows multiple participants to jointly train models, solving the problems of data silos and privacy leakage, but faces challenges in processing large-scale data and improving model training efficiency. Fine-grained symbolic gradient descent optimization came into being, improving learning efficiency by fine-tuning the gradient descent steps; at the same time, with the development of technology and the improvement of computing power, it has become particularly important to develop efficient federated learning methods suitable for large-scale data sets, which requires the method to not only efficiently process large amounts of data, but also adapt to various data types and application requirements while ensuring data privacy;

[0003] Federated learning improves model performance by training models locally on each participant and then aggregating model updates centrally. This approach effectively reduces data transmission requirements and improves data privacy. The main challenges it faces include uneven data distribution, communication efficiency, security of model updates, and coordination between different participants.

[0004] Traditional gradient descent methods are inefficient when processing large-scale data, especially in federated learning environments. To improve efficiency, a more sophisticated and efficient optimization method is needed. Fine-grained symbolic gradient descent optimization is a potential solution that improves learning efficiency by more precisely adjusting the gradient descent steps.

[0005] In federated learning, data security and privacy protection are core issues; traditional learning methods may expose sensitive data, especially in large-scale data environments; therefore, it is crucial to develop a method that can protect privacy while maintaining learning efficiency;

[0006] With the continuous growth of data volume and the improvement of computing power, it is possible to efficiently process large-scale data. In this context, it is particularly urgent to develop an efficient federated learning method suitable for large-scale data sets. This method not only needs to process a large amount of data, but also needs to provide efficient learning capabilities while ensuring data privacy.

[0007] With the development of technology, the application areas of federated learning continue to expand, involving more industries and scenarios; this requires that federated learning methods must be not only efficient, but also flexible enough to adapt to a variety of different data types and application requirements; at the same time, with the continuous improvement of regulations and standards, how to implement efficient and secure federated learning within the legal framework is also an important technical challenge. Summary of the invention

[0008] The purpose of the present invention is to solve the shortcomings existing in the prior art and propose a privacy federated learning method based on fine-grained symbolic gradient descent optimization. In view of the shortcomings of the existing federated learning technology in fine-grained gradient optimization, communication efficiency and privacy protection, the present invention proposes a privacy federated learning method based on fine-grained symbolic gradient descent optimization, thereby providing a more advanced solution for efficient and secure learning of large-scale data sets.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A privacy federated learning method based on fine-grained symbolic gradient descent optimization includes the following steps:

[0011] S1, initialization and setting, setting the initial parameters of the federated learning environment, including model structure, initial weight, and learning rate;

[0012] S2: Local training and gradient calculation: Participants train models on local data and calculate local model gradients.

[0013] S3. Construction of loss function, definition of loss function L {total} =αL {accuracy} +βL {efficiency} +γL {privacy} Among them, L {accuracy} Indicates the model accuracy loss, L {efficiency} represents the computational efficiency loss, L {privacy} represents the privacy protection loss, and the parameters (α, β, γ) are used to balance the importance of different aspects;

[0014] S4, reinforcement learning optimization, introduce reinforcement learning agent to dynamically adjust the learning rate and key parameters, and adjust the strategy by observing the past training performance and the current loss function value to optimize L {total} ;

[0015] S5. Application of game theory strategies: Analyze the strategic interactions between participants through game theory, calculate the Nash equilibrium point, and ensure the overall stability and efficiency of the system;

[0016] S6, model aggregation and update, update the local models of all participants to the central server for aggregation to update the new global model of this round;

[0017] S7. Privacy protection mechanism: During the model update process, local differential privacy and other technologies are applied to protect the data privacy of participants;

[0018] S8. Solution to the maximum and minimum optimization problem. The entire federated learning process is regarded as a maximum and minimum optimization problem, that is, Use iterative methods and game theory strategies to solve this optimization problem to achieve the best balance between efficiency, accuracy and privacy protection;

[0019] S9. Iterative feedback and adjustment: Based on the performance of the model and the feedback from the participants, the strategy and parameters are continuously adjusted for multiple rounds of iterations until the preset performance standards are reached or the model converges.

[0020] Preferably, in step S1, the loss function L is defined as {total} , integrating model accuracy, computational efficiency, and privacy protection.

[0021] Preferably, in step S2, each gradient component is adjusted specifically to ensure more accurate and efficient gradient updating.

[0022] Preferably, in step S5, a game model is constructed so that the strategy selection of the participants can maximize the utility of the overall system, that is, the maximum minimum strategy.

[0023] Preferably, in step S7, a privacy budget and a noise addition mechanism are designed to define local data sensitivity so as to maximize data privacy while ensuring model performance.

[0024] Preferably, the training model described in step S1 sends the network gradient data to the gradient processing task through the asynchronous communication unit of the model training task, the model training task and the gradient processing task are integrated in a distributed processor, and the network gradient data is obtained by performing gradient calculation on the neural network model;

[0025] Performing data encapsulation processing on the network gradient data based on the gradient processing task to obtain processed network gradient data;

[0026] According to the gradient processing task, the processed network gradient data is subjected to data reduction processing to obtain the back-transmitted network gradient data corresponding to the model training task;

[0027] The neural network model is gradient updated based on the returned network gradient data through the model training task to train the neural network model.

[0028] Preferably, the model training task includes at least one training subtask, each training subtask is set with an execution priority, and the gradient updating of the neural network model based on the backhaul network gradient data by the model training task includes: if there are at least two current tasks to be processed, obtaining the execution priority corresponding to each current task to be processed;

[0029] Determine the target training subtask according to the execution priority of the current task to be processed;

[0030] Based on the target training subtask, the neural network model is gradient updated based on the back-propagated network gradient data through the model training task.

[0031] Preferably, the training subtask includes a gradient update task, and the target-based training subtask performs a gradient update on the neural network model based on the backhaul network gradient data through the model training task, including:

[0032] If the target training subtask is not a gradient update task, the target training subtask is executed, and the target training subtask is determined according to the execution priority of the current task to be processed, and the execution is returned until the target training subtask is a gradient update task;

[0033] If the target training task is a gradient update task, the neural network model is gradient updated based on the returned network gradient data by the distributed training processor.

[0034] Preferably, before the asynchronous communication unit of the model training task performs gradient calculation on the neural network model to obtain network gradient data and sends it to the gradient processing task, the method further comprises: constructing an initial asynchronous communication unit through a linear code compiler;

[0035] Based on the initial asynchronous unit, the asynchronous transmission logic code in the distributed training framework program is called and encapsulated into an asynchronous execution function;

[0036] Based on the asynchronous execution function and the initial asynchronous communication unit, the asynchronous communication unit is obtained, so as to call the asynchronous transmission logic code based on the asynchronous communication unit to send the network gradient data to the gradient processing task; the asynchronous communication unit of the model training task performs gradient calculation on the neural network model to obtain the network gradient data and sends it to the gradient processing task, including: performing gradient calculation on the neural network model based on the model training task to obtain the network gradient data, and the neural network model is deployed in the distributed processor; through the asynchronous execution function in the asynchronous communication unit of the model training task, the asynchronous transmission logic code is called to send the network gradient data to the gradient processing task.

[0037] The step of performing data reduction processing on the processed network gradient data according to the gradient processing task to obtain the return network gradient data corresponding to the model training task includes: sending the processed network gradient data to the gradient transfer function in the distributed training framework program through the gradient processing task;

[0038] The processed network gradient data is gradient-transmitted between the distributed processors through the gradient transfer function to obtain the return network gradient corresponding to each distributed training processor.

[0039] Technical effects and advantages of the present invention: The present invention provides a privacy federated learning method based on fine-grained symbolic gradient descent optimization. The present invention implements fine-grained symbolic gradient descent optimization. By more finely adjusting the symbolic gradient in the gradient descent process, the present invention aims to improve the efficiency and accuracy of model training, especially when processing large-scale data sets; fine-grained optimization helps to converge to the optimal solution faster while reducing the consumption of computing resources.

[0040] Improve communication efficiency. The present invention focuses on improving communication efficiency in federated learning by optimizing the data transmission process, compressing the reward model into symbols, and reducing the amount of data and communication delay for model updates. High communication efficiency is particularly important for distributed learning systems and can significantly improve the efficiency of the overall learning process.

[0041] Enhanced privacy protection capabilities. The present invention is also committed to enhancing privacy protection in federated learning. Through an improved local differential privacy optimization method, the data of participants can be more effectively protected from being leaked. Strengthening the privacy protection mechanism helps to expand the application scope of federated learning, especially in areas with strict privacy requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of federated learning based on symbolic gradient descent of the present invention;

[0043] Figure 2 This is a flowchart of optimization after strengthening learning for efficiency, privacy protection, and efficiency modeling as a joint state in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0045] The present invention provides Figure 1-2 The privacy federated learning method based on fine-grained symbolic gradient descent optimization shown in the present invention aims at the shortcomings of existing federated learning technology in fine-grained gradient optimization, communication efficiency and privacy protection. The present invention proposes a privacy federated learning method based on fine-grained symbolic gradient descent optimization, thereby providing a more advanced solution for efficient and secure learning of large-scale data sets.

[0046] The specific implementation steps of the technical solution are described as follows:

[0047] 1. Initialization and setup: set the initial parameters of the federated learning environment, including model structure, initial weights, and learning rate; define a new loss function L {total} , which integrates multiple aspects of model accuracy, computational efficiency, and privacy protection.

[0048] 2. Local training and gradient calculation: each participant trains the model on local data and calculates the local model gradient Apply fine-grained signed gradient descent, which fine-tunes each gradient component to ensure more accurate and efficient gradient updates.

[0049] 3. Construction of loss function, definition of loss function L {total} =αL {accuracy} +βL { e ffi c i ency } +γL { pr i vacy }

[0050] , where L {accuracy} Indicates the model accuracy loss, L {efficiency} represents the computational efficiency loss, L {privacy} represents the privacy protection loss; the parameters (α, β, γ) are used to balance the importance of different aspects.

[0051] 4. Reinforcement learning optimization, introduce reinforcement learning agents to dynamically adjust the learning rate and other key parameters; the agent adjusts the strategy by observing past training performance and the current loss function value to optimize L { tota l} .

[0052] 5. Application of game theory strategies: Analyze the strategic interactions between participants through game theory to ensure the overall stability and efficiency of the system; build a game model so that each participant’s strategy selection can maximize the utility of the overall system, that is, the maximum-minimum (max-min) strategy.

[0053] 6. Model aggregation and update: Aggregate model updates from all participants to a central server to update the global model; use weighted average or other advanced aggregation techniques to take into account the quality and quantity of data from each participant.

[0054] 7. Privacy protection mechanism: During the model update process, differential privacy and other technologies are applied to protect the data privacy of participants; privacy budget and noise addition mechanism are designed to maximize data privacy while ensuring model performance.

[0055] 8. Solution to the maximum and minimum optimization problem. The entire federated learning process is regarded as a maximum and minimum optimization problem, that is, An iterative approach and game-theoretic strategies are used to solve this optimization problem to achieve the best balance among efficiency, accuracy, and privacy protection.

[0056] 9. Iterative feedback and adjustment: continuously adjust strategies and parameters based on the performance of the model and feedback from participants; perform multiple rounds of iterations until the preset performance standards are reached or the model converges.

[0057] The purpose of the above is to:

[0058] The present invention realizes fine-grained symbolic gradient descent optimization. By more finely adjusting the symbolic gradient in the gradient descent process, the present invention aims to improve the efficiency and accuracy of model training, especially when processing large-scale data sets; fine-grained optimization helps to converge to the optimal solution faster while reducing the consumption of computing resources.

[0059] Improve communication efficiency. The present invention focuses on improving communication efficiency in federated learning by optimizing the data transmission process and reducing the amount of data and communication delay for model updates. High communication efficiency is particularly important for distributed learning systems and can significantly improve the efficiency of the overall learning process.

[0060] Enhanced privacy protection capabilities. The present invention is also committed to enhancing privacy protection in federated learning. Through improved optimization methods, the data of participants can be more effectively protected from being leaked. Strengthening the privacy protection mechanism helps to expand the application scope of federated learning, especially in areas with strict privacy requirements.

[0061] In general, the purpose of the present invention is to address the deficiencies of existing federated learning techniques in terms of fine-grained gradient optimization, communication efficiency, and privacy protection, thereby providing a more advanced solution for efficient and secure learning of large-scale datasets.

[0062] This embodiment also provides a specific method for training the model, which is:

[0063] The network gradient data is sent to the gradient processing task through the asynchronous communication unit of the model training task. The model training task and the gradient processing task are integrated in the distributed processor. The network gradient data is obtained by performing gradient calculation on the neural network model.

[0064] The model training task includes at least one training subtask, each training subtask is set with an execution priority, and the neural network model is gradient updated based on the back-transmitted network gradient data through the model training task, including: if there are at least two current tasks to be processed, obtaining the execution priority corresponding to each current task to be processed;

[0065] Determine the target training subtask according to the execution priority of the current pending task;

[0066] It should be noted that the training subtask includes a gradient update task. Based on the target training subtask, the neural network model is updated based on the back-transmitted network gradient data through the model training task, including:

[0067] If the target training subtask is not a gradient update task, execute the target training subtask and return to execute the target training subtask according to the execution priority of the current pending task until the target training subtask is a gradient update task;

[0068] If the target training task is a gradient update task, the neural network model is gradient updated based on the returned network gradient data through the distributed training processor;

[0069] Based on the target training subtask, the neural network model is updated based on the back-propagated network gradient data through the model training task;

[0070] Based on the gradient processing task, the network gradient data is packaged and processed to obtain the processed network gradient data;

[0071] According to the gradient processing task, the processed network gradient data is processed by data reduction to obtain the back-transmitted network gradient data corresponding to the model training task;

[0072] The neural network model is trained by updating the gradient of the neural network model based on the back-propagated network gradient data through the model training task.

[0073] Before the asynchronous communication unit of the model training task calculates the gradient of the neural network model to obtain the network gradient data and sends it to the gradient processing task, the method further includes: constructing an initial asynchronous communication unit through a linear code compiler;

[0074] Based on the initial asynchronous unit, the asynchronous transmission logic code in the distributed training framework program is called and encapsulated into an asynchronous execution function;

[0075] Based on the asynchronous execution function and the initial asynchronous communication unit, an asynchronous communication unit is obtained, and the asynchronous transmission logic code is called based on the asynchronous communication unit to send the network gradient data to the gradient processing task; through the asynchronous communication unit of the model training task, the gradient of the neural network model is calculated to obtain the network gradient data and send it to the gradient processing task, including: based on the model training task, the neural network model is calculated to obtain the network gradient data, and the neural network model is deployed in the distributed processor; through the asynchronous execution function in the asynchronous communication unit of the model training task, the asynchronous transmission logic code is called to send the network gradient data to the gradient processing task.

[0076] According to the gradient processing task, the processed network gradient data is subjected to data reduction processing to obtain the return network gradient data corresponding to the model training task, including: sending the processed network gradient data to the gradient transfer function in the distributed training framework program through the gradient processing task;

[0077] The processed network gradient data is transferred between distributed processors through the gradient transfer function to obtain the return network gradient corresponding to each distributed training processor.

[0078] With respect to the above-mentioned embodiments, firstly, by introducing a fine-grained symbolic gradient descent optimization method, the present invention has taken an important step in improving the efficiency and accuracy of gradient updates in federated learning. Compared with the coarse-grained gradient descent method commonly used in the prior art, this fine-grained optimization can adjust the gradient more accurately, thereby speeding up the learning speed and improving the performance of the final model. This is particularly important for processing large-scale data sets because it can significantly reduce the training time and consumption of computing resources.

[0079] Secondly, the present invention brings higher flexibility and stability to federated learning by integrating reinforcement learning and game theory strategies. The reinforcement learning algorithm enables the model to dynamically adjust key parameters in the learning process, so as to better adapt to changes and challenges in training data. At the same time, the application of game theory ensures effective collaboration and fairness among participants, which is particularly important in the federated learning environment of multi-party data collaboration.

[0080] Finally, the present invention also shows obvious advantages in terms of privacy protection. By introducing advanced privacy protection mechanisms, such as differential privacy, the present invention can ensure data security while maintaining effective model learning. This is of great significance for promoting federated learning in privacy-sensitive application scenarios.

[0081] In general, compared with the existing technology, the present invention has significant improvements in improving the efficiency and accuracy of federated learning and protecting data privacy, making it an innovative solution that is superior to the existing technology in many aspects.

[0082] The pseudo code of the algorithm of the method proposed in this embodiment is described as follows:

[0083] Input: dataset D, participant set P, learning rate lr, privacy budget epsilon

[0084] Output: Optimized global model M

[0085] initialization:

[0086] For each participant p∈P:

[0087] Initialize the local model Mp

[0088] Initialize the global model M

[0089] Set initial parameters: learning rate lr, privacy budget epsilon

[0090] Set the loss function weights: α, β, γ

[0091] Iteration process:

[0092] While the stop condition is not met do

[0093] For each participant p∈P do

[0094] Train the model Mp on local data

[0095] Computing Gradients

[0096] Apply fine-grained signed gradient descent optimization

[0097] Apply reinforcement learning agent to adjust learning rate lr and parameters

[0098] Applying game theory to strategy optimization

[0099] Add privacy-preserving noise based on epsilon

[0100] The optimized Send to central server

[0101] end for

[0102] On the central server:

[0103] Aggregate gradient updates from all participants

[0104] Update the global model M

[0105] Feedback model performance and loss function L_total to all participants

[0106] Evaluate the performance of the global model M

[0107] Adjust strategies and parameters based on performance feedback

[0108] end while

[0109] Returns the global model M

[0110] End of algorithm

[0111] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A privacy-preserving federated learning method based on fine-grained symbolic gradient descent optimization, It is characterized in that The steps include: S1, initialization and setting, setting the initial parameters of the federated learning environment, including model structure, initial weight, and learning rate; S2: Local training and gradient calculation: Participants train models on local data and calculate local model gradients. S3. Construction of loss function, definition of loss function L {total} =αL {accuracy} +βL {efficiency} +γL {privacy} Among them, L {accuracy} Indicates the model accuracy loss, L {efficiency} represents the computational efficiency loss, L {privacy} represents the privacy protection loss, and the parameters (α, β, γ) are used to balance the importance of different aspects; S4, reinforcement learning optimization, introduce reinforcement learning agent to dynamically adjust the learning rate and key parameters, and adjust the strategy by observing the past training performance and the current loss function value to optimize L {total} ; S5. Application of game theory strategies: Analyze the strategic interactions between participants through game theory, calculate the Nash equilibrium point, and ensure the overall stability and efficiency of the system; S6, model aggregation and update, update the local models of all participants to the central server for aggregation to update the new global model of this round; S7. Privacy protection mechanism: During the model update process, local differential privacy and other technologies are applied to protect the data privacy of participants; S8. Solution to the maximum and minimum optimization problem. The entire federated learning process is regarded as a maximum and minimum optimization problem, that is, Use iterative methods and game theory strategies to solve this optimization problem to achieve the best balance between efficiency, accuracy and privacy protection; S9. Iterative feedback and adjustment: Based on the performance of the model and the feedback from the participants, the strategy and parameters are continuously adjusted for multiple rounds of iterations until the preset performance standards are reached or the model converges.

2. A privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 1, Features: In step S1, define the loss function L {total} , integrating model accuracy, computational efficiency, and privacy protection.

3. A privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 1, Features: In step S2, each gradient component is adjusted specifically to ensure more accurate and efficient gradient update.

4. A privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 1, Features: In step S5, a game model is constructed so that the strategy selection of the participants can maximize the utility of the overall system, that is, the maximum minimum strategy.

5. The privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 1, Features: In step S7, the privacy budget and noise addition mechanism are designed to define local data sensitivity and maximize data privacy while ensuring model performance.

6. A privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 1, Features: The training model described in step S1 sends the network gradient data to the gradient processing task through the asynchronous communication unit of the model training task, the model training task and the gradient processing task are integrated in the distributed processor, and the network gradient data is obtained by performing gradient calculation on the neural network model; Performing data encapsulation processing on the network gradient data based on the gradient processing task to obtain processed network gradient data; According to the gradient processing task, the processed network gradient data is subjected to data reduction processing to obtain the back-transmitted network gradient data corresponding to the model training task; The neural network model is gradient updated based on the returned network gradient data through the model training task to train the neural network model.

7. A privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 6, Features: The model training task includes at least one training subtask, each training subtask is set with an execution priority, and the gradient updating of the neural network model based on the backhaul network gradient data by the model training task includes: if there are at least two current tasks to be processed, obtaining the execution priority corresponding to each current task to be processed; Determine the target training subtask according to the execution priority of the current task to be processed; Based on the target training subtask, the neural network model is gradient updated based on the back-propagated network gradient data through the model training task.

8. A privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 7, Features: The training subtask includes a gradient update task, and the target-based training subtask performs a gradient update on the neural network model based on the backhaul network gradient data through the model training task, including: If the target training subtask is not a gradient update task, the target training subtask is executed, and the target training subtask is determined according to the execution priority of the current task to be processed, and the execution is returned until the target training subtask is a gradient update task; If the target training task is a gradient update task, the neural network model is gradient updated based on the returned network gradient data by the distributed training processor.

9. The privacy federated learning method based on fine-grained symbolic gradient descent optimization according to claim 1, Features: Before the asynchronous communication unit of the model training task calculates the gradient of the neural network model and obtains the network gradient data and sends it to the gradient processing task, the method further comprises: constructing an initial asynchronous communication unit through a linear code compiler; Based on the initial asynchronous unit, the asynchronous transmission logic code in the distributed training framework program is called and encapsulated into an asynchronous execution function; Based on the asynchronous execution function and the initial asynchronous communication unit, the asynchronous communication unit is obtained, so as to call the asynchronous transmission logic code based on the asynchronous communication unit to send the network gradient data to the gradient processing task; the asynchronous communication unit of the model training task performs gradient calculation on the neural network model to obtain the network gradient data and sends it to the gradient processing task, including: performing gradient calculation on the neural network model based on the model training task to obtain the network gradient data, and the neural network model is deployed in the distributed processor; through the asynchronous execution function in the asynchronous communication unit of the model training task, the asynchronous transmission logic code is called to send the network gradient data to the gradient processing task. The step of performing data reduction processing on the processed network gradient data according to the gradient processing task to obtain the return network gradient data corresponding to the model training task includes: sending the processed network gradient data to the gradient transfer function in the distributed training framework program through the gradient processing task; The processed network gradient data is gradient-transmitted between the distributed processors through the gradient transfer function to obtain the return network gradient corresponding to each distributed training processor.