Federal learning edge device training time delay optimization system based on privacy protection

Abstract

The invention discloses a federated learning edge device training time delay optimization system based on privacy protection, and belongs to the field of federated learning and privacy protection, the system comprises an edge device layer, a central server layer and a coordination layer, the edge device layer, the central server layer and the coordination layer are connected through electric signals, the edge device layer collects and preprocesses data and encrypts and transmits the data, and the central server layer is connected with the coordination layer. Carrying out local training and privacy protection; the coordination layer sets task complexity, time delay and privacy overhead weight, adjusts a strategy according to a comprehensive objective function, optimizes the time delay, and connects the two layers; and the central server layer decrypts, trains and aggregates the data, updates the global model and feeds back. The method comprises the following steps: initializing parameters by a system, and carrying out iterative training such as equipment selection, local updating, strategy judgment and global parameter calculation by using a comprehensive trade-off dynamic sampling federation training algorithm until a preset round is reached to obtain an optimal model. The problems of privacy protection and time delay are solved, and the training efficiency and the data security are improved.

Description

Technical Field

[0001] The present invention relates to the fields of federated learning and privacy protection, and particularly relates to a system for optimizing the training latency of edge devices in federated learning based on privacy protection. Background Art

[0002] In today's digital age, data, as an important asset, has attracted much attention regarding its privacy and security issues. Many industries possess a large amount of sensitive data. Once this data is leaked, it will cause serious losses to users. A large number of edge devices generate a vast amount of data. The traditional centralized data processing method requires uploading a large amount of data to the cloud server for analysis and processing, which exposes the data to a high risk of leakage and causes pressure on network bandwidth, resulting in a relatively high latency. Therefore, federated learning has emerged.

[0003] In federated learning, privacy protection is of utmost importance, but excessive privacy protection measures may affect the performance of the model. Therefore, how to improve the performance of the model as much as possible while ensuring data privacy is an urgent problem to be solved. For this reason, a system for optimizing the training latency of edge devices in federated learning based on privacy protection is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a system for optimizing the training latency of edge devices in federated learning based on privacy protection to solve the problems raised in the above background art.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: A system for optimizing the training latency of edge devices in federated learning based on privacy protection includes an edge device layer, a central server layer, and a coordination layer. The edge device layer, the central server layer, and the coordination layer are connected by electrical signals; The edge device layer collects data through edge devices, preprocesses the data, and encrypts and transmits it; The coordination layer sets the task complexity weight, latency factor weight, and privacy protection overhead weight according to the device status and privacy requirements, adjusts the training strategy according to the task complexity weight, latency factor weight, and privacy protection overhead weight, and connects the edge devices and the central server; The central server layer decrypts the data of the edge layer, trains the decrypted data of the edge layer, aggregates it, and then updates the global model, and feeds back the trained model.

[0006] Further, in the edge device layer, the detailed process of implementing encrypted transmission includes: The edge device adjusts the format of the processed original data stored locally, converts it into vector data suitable for the homomorphic encryption scheme, so that the data meets the requirements of a specific encryption mode.

[0007] Select the additive homomorphic encryption mode and perform a linear combination operation on the processed original data that has been converted into vector data. In this way, the encrypted data is obtained. The advantage of the additive homomorphic encryption mode is that after performing operations such as linear combination on the encrypted data, the decryption result is equivalent to the result of performing the corresponding operation on the plaintext, which provides convenience for subsequent operations such as aggregation on the encrypted data by the central server, while ensuring data privacy.

[0008] After completing the encryption operation, the encrypted data is transmitted to the central server. During the transmission process, the data remains encrypted at all times, ensuring that even if the data is stolen in the data transmission link, it is difficult for the thief to obtain the true information of the original data, thus ensuring the security of the data during transmission.

[0009] Furthermore, in the central server layer, the method for training the global model includes the following steps: Step 1: Set the number of terminal devices participating in the system, the ratio of terminal devices selected in each round, the system training learning rate, the number of local training times of each terminal device in each round of iteration, the local mini-batch size updated by each terminal device, the number of iteration rounds required for the overall system training, initialize the global model parameters , and the threshold of the local sample size of each device; according to the system requirements or the priority of the task, set the weights of task complexity, latency factor, and privacy protection overhead; Step 2: At the beginning of each round of iteration, select devices from the terminal devices according to the ratio of terminal devices selected in each round to form the set of devices selected in this round; Step 3: Call the device update function, use the current global model parameters and the local sample size threshold as inputs to update the local model parameters of the device, and obtain the updated local model parameters, where the local model parameters refer to the parameters involved in the model training by the edge device according to the data collected and processed by itself during the federated learning process; Step 4: After each device completes the update of the local model parameters, calculate the device-level device measurement quantity. The device measurement quantity includes the task complexity measurement quantity, the privacy protection overhead measurement quantity, and the latency factor measurement quantity. Aggregate the device-level measurement quantities of all participating devices to obtain the global measurement quantity. Use the set weights of task complexity, latency factor, and privacy protection overhead to calculate the value of the comprehensive objective function. According to the calculated value of the comprehensive objective function, determine whether it is necessary to adjust the subsequent training strategy; Step 5: Calculate the updated global model parameters: The central server calculates the updated global model parameters through the global model parameter calculation formula based on the local updated model parameters uploaded by all participating devices and the sample data volume of each device; Step 6: If the current iteration round is less than the preset total number of iteration rounds, return to Step 2 to start the next round of iterative training. If the current iteration round is equal to the preset total number of iteration rounds, obtain the optimal final global model. The above Steps 1 to 6 can be referred to as the comprehensive trade-off dynamic sampling federated training algorithm.

[0010] A further improvement of the technical solution of the present invention lies in that: in the above Step 1, the process of setting relevant weights according to system requirements or task priorities includes: Firstly, deeply understand the requirements applied by the system. Secondly, collect the performance index data of the system during past operations, analyze the resource usage of edge devices, and their occupancy under different task complexities, time delays, and privacy protection measures. Through these data, the associations and mutual influences between complexity, time delay, and privacy protection can be evaluated.

[0011] Based on the analysis of system requirements and historical data, determine the key indicators that can measure system performance. For task complexity, indicators such as model training time and device resource utilization rate can be adopted; for the time delay factor, data transmission time and model update interval indicators can be used; for privacy protection overhead, indicators such as the computational cost of encryption algorithms and the impact degree of privacy protection mechanisms on model accuracy can be used.

[0012] Attempt to quantify the relationships between these indicators and the overall system performance. Through statistical methods such as regression analysis and correlation analysis, determine the specific impact degrees of changes in task complexity, time delay, and privacy protection overhead on the key performance indicators of the system.

[0013] Decompose the system objective into multiple levels, construct a hierarchical structure model, including an objective layer, a criterion layer, and a scheme layer. The objective layer is the optimal system performance, the criterion layer is task complexity, time delay factor, and privacy protection overhead, and the scheme layer is specific training strategies and privacy protection methods. By pairwise comparing the relative importance of each factor, construct a judgment matrix, calculate the weight vectors of task complexity, time delay factor, and privacy protection overhead, and conduct a consistency test to obtain the weights of task complexity, time delay factor, and privacy protection overhead. A further improvement of the technical solution of the present invention lies in that: in the above Step 4, the process of calculating the comprehensive objective function and adjusting subsequent training strategies includes: The calculation of the comprehensive objective function is:

[0014] where is for calculating the comprehensive objective function, is the measure of task complexity, is the measure of the time delay factor, is the measure of privacy protection overhead. is the weight of task complexity, is the weight of the time delay factor, is the weight of the privacy protection overhead; The calculation formula for the task complexity measurement quantity is: , , where is the task complexity measurement quantity, is the task complexity measurement quantity of device k, is the CPU frequency of the device, is the number of CPU cycles required to train a single sample of data, is the amount of local sample data of the device, is the number of local training times per round of iteration; The calculation formula for the time delay factor measurement quantity is: , where is the time delay factor measurement quantity, is the local update training time of the device, is the communication time between the device and the server; The calculation formula for the privacy protection overhead measurement quantity is:

[0015]

[0016] where is the privacy protection overhead measurement quantity, is the time delay factor measurement quantity of device k, is the complexity of the encryption algorithm, is the amount of encrypted data, is the computational amount of additional calculation steps introduced by the privacy protection mechanism, is the number of communications, is the privacy protection overhead for each communication.

[0017] Adjust the subsequent training strategy according to the value of the calculated comprehensive objective function. If the value of the comprehensive objective function indicates that there are major problems in the current system in terms of time delay, privacy protection overhead, or task complexity, then dynamically adjust the subsequent training strategy according to the main factors that lead to the unsatisfactory value of the comprehensive objective function; If the value of the calculated comprehensive objective function is significantly higher than the value of the comprehensive objective function in the past normal training rounds, and it is found through calculation that the proportion of the time delay factor measurement quantity in the value of the comprehensive objective function exceeds the pre-set time delay influence threshold + 5%, then it is determined that the current system is mainly affected by the time delay factor, and the dynamic sampling strategy is adjusted to preferentially select devices with short training time and high communication efficiency to participate in the subsequent iterative training; If the value of the calculated comprehensive objective function is significantly higher than the value of the comprehensive objective function in previous normal training rounds, and it is found through calculation that the proportion of the privacy protection overhead measurement in the value of the comprehensive objective function exceeds the preset time delay impact threshold by +5%, it is determined that the current system is mainly affected by the privacy protection overhead factor, and the privacy protection strategy is adjusted to reduce the overhead while ensuring a certain privacy level; If the value of the calculated comprehensive objective function is significantly higher than the value of the comprehensive objective function in previous normal training rounds, and it is found through calculation that the proportion of the task complexity measurement in the value of the comprehensive objective function exceeds the preset time delay impact threshold by +5%, it is determined that the current system is mainly affected by the task complexity factor, and the device computing resource allocation is optimized to balance the task load of each device.

[0018] A further improvement of the technical solution of the present invention lies in that: in step five, the global model parameter calculation process includes: First, initialize the system, including: setting the number of terminal devices participating in the system , the ratio of terminal devices selected to participate in each round , the system training learning rate , the number of local training times of each terminal device in each iteration , the local mini-batch size updated by each terminal device , the number of iterations required for the overall system training , setting the weight of task complexity according to system requirements or task priorities , the weight of the time delay factor , the weight of the privacy protection overhead , initializing the global model parameters , and the threshold of the local sample size of each device ( ); Then, perform round-by-round iteration. For the following operations are performed: Select devices, where is the number of devices participating in training in this iteration, and belongs to the set , forming the set of selected devices in the th round ; Parallelly, for each execute the call to the device update function , obtain the updated model parameters of device , calculate the relevant measurement of this device, summarize the measurements of all participating devices, and calculate the comprehensive objective function; Adjust the strategy according to the comprehensive objective function, according to the calculated comprehensive objective function The value and the contributions of each part of the measurement quantity to it are used to dynamically adjust the subsequent training strategy: If is mainly affected by that is, the latency is too high, the dynamic sampling strategy can be optimized in subsequent iterations, and devices with short training time are preferentially selected to participate in training; If is mainly affected by , the privacy protection strategy can be considered to be adjusted to reduce the overhead while ensuring privacy; If is mainly affected by , the allocation of device computing resources or the data processing flow can be tried to be optimized; Finally, model aggregation is performed to calculate the updated global model parameters: , where , n is the sum of the local sample data volumes of all participating devices, , is the total number of iterations required for the overall system training, represents the updated global model parameters at the (t + 1)-th iteration, is the total number of terminal devices participating in the system, which is a fixed value, is the -th device, is the size of the available local sample data on the -th device, A further improvement of the technical solution of the present invention lies in: The calculation process of the device update function DeviceUpdate includes: Input the device index , the current model parameters and the threshold of the device local sample volume size ; Initialize , the available local sample data set on device is ; If : Initialize , select an initial subset from the local sample data set of device , for to , perform local training iterations: Calculate the gradient , where is the current model parameter, is the sample subset is the gradient operator; Update the model parameters: , using the learning rate Update the model parameters and select new samples: Select a subset of samples such that the size of the new sample subset is the same as that of the previous round ; Calculate the sample variance: Calculate the variance of the current sample subset, and then judge the condition: If the condition of the formula is not satisfied, that is, the gradient variance of the current sample subset and other conditions do not meet the requirements, then increase the sample size: According to the formula increase the size of the sample subset , where is the calculation formula of the convex empirical loss objective function , is the prediction function for gradient descent calculation using the parameter , is the convex loss function; Otherwise , if the local sample size of the device is less than or equal to the threshold, for to , perform local training iterations, and then update the model parameters: , directly use the overall sample resources, and perform local gradient descent to update the model parameters based on the FedAvg algorithm; Finally, return the updated model parameters of the device .

[0019] Furthermore, in the central server, the model feedback process includes: After the central server completes the global model update, through the classification performance comprehensive evaluation formula, evaluate the effectiveness and accuracy of the updated global model from the aspects of accuracy and recall rate. According to the evaluation results, if the model performance reaches or exceeds the preset performance threshold, then feedback the updated global model to the edge device; The classification performance comprehensive evaluation formula is: , where F1 is the performance threshold. In some cases, the individual accuracy or recall rate may not fully reflect the performance of the model, while the F1 value can balance these two metrics and provide a more comprehensive evaluation. When one of the accuracy and recall rate is high and the other is low, the F1 value can more objectively reflect the performance of the model in the classification task.

[0020] Due to the adoption of the above technical solutions, the technical progress achieved by the present invention compared with the prior art is: 1. The present invention takes privacy protection measures at the edge device layer, introduces an encryption algorithm when processing the original device data, and runs through the whole process of data transmission and model training. By setting the privacy protection overhead weight, privacy protection and other factors are comprehensively considered in the comprehensive objective function, and the privacy protection strategy can be dynamically optimized according to the actual situation. When the privacy protection overhead is too large and affects the performance, the encryption algorithm is adjusted in time or the data processing flow is optimized, so as to ensure data privacy and reduce the negative impact on model training; 2. The coordination layer of the present invention can adjust the training strategy according to the device status and privacy requirements, and preferentially select devices with high training and communication efficiency during the device selection stage. At the same time, considering factors such as the local training and communication time of the device, the delay factor measurement is calculated. When it is found that the delay problem is serious, the dynamic sampling strategy is adjusted in time to effectively reduce the waiting time and communication overhead, reduce the training delay, and improve the system training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments recorded in the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings.

[0022] Figure 1 is a schematic flow chart of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] Embodiment, as Figure 1 described, the federated learning edge device training delay optimization system based on privacy protection includes, in the system architecture, covering the edge device layer, the central server layer, and the coordination layer. Each layer is interconnected through electrical signals and collaborates to complete the model training task.

[0025] The edge device layer undertakes the key responsibilities of data processing and local model training. At this layer, various types of data are first collected, and then preprocessing operations are performed to lay a foundation for subsequent model training. Then, the model training task is executed locally, and gradient calculation and encrypted transmission are completed.

[0026] The central server layer is mainly responsible for overall planning of the update and management of the global model. It receives the model update information uploaded from each edge device, aggregates and processes this information, thereby completing the update of the global model. After the update is completed, the new model parameters are sent to each edge device. During the whole process, data privacy protection is of utmost importance. For this reason, the central server layer uses homomorphic encryption technology to encrypt the update gradients transmitted from each edge device. In this way, even if the central server receives encrypted gradient information, it can still perform effective aggregation operations without revealing any privacy, ensuring the accuracy and security of model updates.

[0027] The coordination layer plays an important role in balancing and coordinating in the whole system. Based on information such as training progress, real-time status of each device, and network latency, it dynamically adjusts the privacy protection strategy and latency optimization strategy, aiming to achieve the best balance between privacy protection and latency. Through precise strategy formulation and flexible adjustment, it ensures that while the system operates efficiently, it maximally guarantees data privacy and security.

[0028] Furthermore, the working principle of the present invention is illustrated by the following embodiments: There are multiple intelligent environment monitoring devices distributed in different households set as edge devices. They jointly train a model for predicting the comfort level of the indoor environment. This model can predict whether the indoor environment is comfortable based on data such as temperature, humidity, and air quality, provide environmental adjustment suggestions for users, while protecting the privacy of users' household data and optimizing the training latency.

[0029] According to the requirements of the smart home environment monitoring system, the number of intelligent environment monitoring devices participating in the training is set to 50, and 20% of the devices are selected to participate in the training in each round, that is, 10 devices; the system training learning rate is set to 0.01; the number of local training iterations for each device in each round is 3 times; the local mini-batch size is 8; the overall system training needs to iterate 50 rounds.

[0030] The global model parameters are obtained through random initialization. The local sample size threshold is set according to the storage and computing capabilities of the devices. For example, the threshold for devices with higher configuration is set to 100, and that for ordinary devices is set to 60. Considering the requirements of the smart home scenario for privacy protection and real-time performance, the historical performance index data of the system is collected, and the judgment matrix is constructed using the analytic hierarchy process. After calculation and verification, the task complexity weight is set to 0.3, the latency factor weight is set to 0.35, and the privacy protection overhead weight is set to 0.35.

[0031] The intelligent environmental monitoring device collects data such as indoor temperature, humidity, and air quality through temperature and humidity sensors and air quality sensors. After collection, the data is cleaned to remove outliers, denoised using a filtering algorithm, and the data format is unified. At the same time, the original data of the device itself, such as the CPU frequency, the number of CPU cycles required to train a single sample of data, and the amount of local sample data, is recorded, and after Gaussian filtering and normalization processing, it is stored in the local storage medium.

[0032] Taking a high - configuration intelligent environmental monitoring device as an example, if its local sample data volume is 120, which is greater than the threshold of 100. When it is selected to participate in training in each round, the device update function is called to initialize a sample subset. 50 samples are randomly selected from the local sample dataset, and 3 local training iterations are performed. The gradient is calculated in each iteration, the model parameters are updated according to the learning rate, the new sample subset is selected to keep the size unchanged, and the sample variance is calculated. If the gradient variance of the sample subset does not meet the preset conditions, the sample size is increased according to the formula. During the training process, differential privacy technology is used to add appropriate noise to the model gradient to protect the privacy of the training data.

[0033] At the beginning of each round of iteration, 10 devices are randomly selected from 50 intelligent environmental monitoring devices at a ratio of 20% to form the set of devices selected in this round. The coordination layer monitors the training progress, device status, and network latency in real - time. The device status includes CPU usage and power. When the network latency of a certain device is high and affects the training, if the comprehensive objective function shows that the latency factor has a great impact, the device with good network conditions and short training time is preferentially selected in the next iteration. During the training process, according to the measurement values calculated after each device completes the update of the local model parameters, the global measurement value is summarized to calculate the comprehensive objective function. If the comprehensive objective function shows that the privacy protection overhead is too large, the coordination layer notifies the edge device to adjust the privacy protection strategy, such as optimizing the parameters of the encryption algorithm.

[0034] After the edge device completes the local model training and calculates the gradient, through the homomorphic encryption algorithm, the generated public key is used to encrypt the gradient. The encrypted gradient data is transmitted to the central server. The central server receives the encrypted gradient data and directly performs an aggregation operation on it using the homomorphic encryption property. The central server calculates the updated global model parameters through the global model calculation formula based on the local updated model parameters uploaded by all participating devices and the sample data volume of each device. After the update is completed, the new global model parameters are sent to each edge device.

[0035] After each round of global model update is completed, check whether the current iteration round is less than 50 rounds. If it is less, return to the device selection step to start the next round of iterative training; if it is equal to 50 rounds, it is considered that the model training is completed, and the optimal final global model parameters are obtained.

[0036] Finally, after the central server completes the global model update, it calculates the F1 value through the comprehensive classification performance evaluation formula, and evaluates the effectiveness and accuracy of the updated global model in terms of accuracy and recall rate. Assume that the preset performance threshold is that the F1 value reaches 0.7. If the model performance reaches or exceeds this threshold, the updated global model is fed back to the edge device. If the threshold is not reached, the training strategy can be adjusted according to the specific situation, such as increasing the number of training rounds, adjusting weights, etc., and retraining and evaluating again.

[0037] As described above, it is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claimed rights.

Claims

1. A privacy - protected federated learning edge device training latency optimization system, comprising: An edge device layer, a central server layer, and a coordination layer. The edge device layer, the central server layer, and the coordination layer are connected by electrical signals. It is characterized in that The edge device layer collects data through edge devices, preprocesses the data, and encrypts and transmits it. The coordination layer sets the task complexity weight, the delay factor weight, and the privacy protection overhead weight according to the device status and privacy requirements, adjusts the training strategy according to the task complexity weight, the delay factor weight, and the privacy protection overhead weight, and connects the edge devices and the central server. The central server layer decrypts the edge layer data, trains the decrypted edge layer data and aggregates it, then updates the global model, and feeds back the trained model.

2. The latency optimization system for edge device training in federated learning based on privacy protection according to claim 1, wherein In the edge device layer, the process of collecting and preprocessing the own original data of the edge device includes: Establish a connection with the detector equipped on the edge device to obtain the own original data of the edge device. The own original data includes: the CPU frequency of the edge device, the number of CPU cycles required to train a single sample data, the amount of local sample data of the edge device, the number of local training times per round of iteration, the local update training time of the edge device, the communication time between the edge device and the central server, the central server waiting time, the computational amount of the encryption algorithm, the amount of encrypted data, the computational amount of the additional calculation steps introduced by the privacy protection mechanism, the number of communications, and the privacy protection overhead index for each communication. Perform data outlier processing on the collected own original data of the edge device, remove the outliers in the data that exceed the range of the mean ± 3 times the standard deviation, perform a normalization operation on the data, and scale the normalized data to the interval [0, 1]. Store the normalized own original data in the local storage medium of the edge device.

3. The latency optimization system for edge device training in federated learning based on privacy protection according to claim 2, wherein, The process of encrypting and transmitting the normalized own original data includes: The edge device adjusts the format of the normalized own original data, converts the normalized own original data into vector data suitable for the homomorphic encryption scheme, selects the additive homomorphic encryption mode, performs a linear combination on the encrypted vector data to obtain the encrypted data, and transmits it to the central server.

4. The latency optimization system for edge device training in federated learning based on privacy protection according to claim 3, wherein, In the coordination layer, the process of setting the coefficients of the task complexity weight, the delay factor weight, and the privacy protection overhead weight, and calculating the comprehensive objective function value by using the coefficients of the task complexity weight, the delay factor weight, and the privacy protection overhead weight through the comprehensive objective function includes: Clarify the application scenario of the federated learning edge device training delay optimization system based on privacy protection, and combine the performance index data of the collection system during past operations to set the coefficients of the task complexity weight, the delay factor weight, and the privacy protection overhead weight. The comprehensive objective function is: ； Among them To calculate the comprehensive objective function Is a measure of task complexity Is a measure of the delay factor Is a measure of the privacy protection cost Is the weight of task complexity Is the weight of the delay factor Is the weight of the privacy protection cost 5. The latency optimization system for edge device training in federated learning based on privacy protection according to claim 4, characterized in that , The calculation formula for the task complexity measurement quantity is as follows: , where is the task complexity measurement quantity, is the task complexity measurement quantity of device k, is the CPU frequency of the device, is the number of CPU cycles required to train a single sample data, is the amount of local sample data of the device, is the number of local training times per round of iteration; The calculation formula for the delay factor measurement quantity is as follows: , where is the delay factor measurement quantity, is the local update training time of the device, is the communication time between the device and the server; The calculation formula for the privacy protection overhead measurement quantity is as follows: , , where is the privacy protection overhead measurement quantity, is the delay factor measurement quantity of device k, is the complexity of the encryption algorithm, is the amount of encrypted data, is the amount of computation for the additional computation steps introduced by the privacy protection mechanism, is the number of communications, is the privacy protection overhead for each communication.

6. The latency optimization system for edge device training in federated learning based on privacy protection according to claim 5, wherein The method of adjusting the subsequent training strategy according to the calculated comprehensive objective function value: If the value of the calculated comprehensive objective function is significantly higher than the value of the comprehensive objective function in previous normal training rounds, and it is found through calculation that the proportion of the delay factor measurement in the value of the comprehensive objective function exceeds the pre-set delay impact threshold by +5%, it is determined that the current system is mainly affected by the delay factor, and the dynamic sampling strategy is adjusted to preferentially select devices with short training time and high communication efficiency to participate in subsequent iterative training; If the value of the calculated comprehensive objective function is significantly higher than the value of the comprehensive objective function in previous normal training rounds, and it is found through calculation that the proportion of the privacy protection overhead measurement in the value of the comprehensive objective function exceeds the pre-set delay impact threshold by +5%, it is determined that the current system is mainly affected by the privacy protection overhead factor, and the privacy protection strategy is adjusted to reduce the overhead while ensuring a certain privacy level; If the value of the calculated comprehensive objective function is significantly higher than the value of the comprehensive objective function in previous normal training rounds, and it is found through calculation that the proportion of the task complexity measurement in the value of the comprehensive objective function exceeds the pre-set delay impact threshold by +5%, it is determined that the current system is mainly affected by the task complexity factor, and the device computing resource allocation is optimized to balance the task load of each device.

7. The optimized system for training latency of edge devices in federated learning based on privacy protection according to claim 6, characterized in that, In the central server layer, the method for training the global model includes the following steps: Step 1: Set the number of terminal devices involved in the system, the ratio of terminal devices selected to participate in each round, the system training learning rate, the number of local training times of each terminal device in each round of iteration, the local mini-batch size updated by each terminal device, the number of iteration rounds required for the overall system training, and initialize the global model parameters , and the threshold of the local sample size of each device; Set the weights of task complexity, delay factor, and privacy protection overhead according to system requirements or task priorities; Step 2: Device selection. At the beginning of each iteration, devices are selected from the terminal devices according to the ratio of the terminal devices participating in each round to form the set of devices selected in this round; Step 3: Call the device update function. Using the current global model parameters and the local sample volume threshold as inputs, update the local model parameters of the device to obtain the updated local model parameters, where the local model parameters refer to the parameters involved in the model training by the edge device according to the data collected and processed by itself during the federated learning process; Step 4: After each device completes the update of the local model parameters, calculate the device-level device measurement. The device measurement includes the task complexity measurement, the privacy protection overhead measurement, and the delay factor measurement. Aggregate the device-level measurements of all participating devices to obtain the global measurement. Use the set weights of the task complexity, the delay factor, and the privacy protection overhead to calculate the value of the comprehensive objective function. According to the calculated value of the comprehensive objective function, determine whether it is necessary to adjust the subsequent training strategy; Step 5: Calculate the updated global model parameters: The central server calculates the updated global model parameters through the global model parameter calculation formula based on the local updated model parameters uploaded by all participating devices and the sample data volume of each device; Step 6: If the current iteration round is less than the pre-set total iteration rounds, return to Step 2 to start the next round of iterative training. If the current iteration round is equal to the pre-set total iteration rounds, obtain the optimal final global model.

8. The system for optimizing the training latency of edge devices in federated learning based on privacy protection according to claim 7, wherein The calculation formula for the global model parameters is: , among which ; where n is the sum of the local sample data volumes of all participating devices, , is the number of iterations required for the overall system training, represents the updated global model parameters at the (t + 1)-th iteration, is the total number of terminal devices participating in the system, which is a fixed value, is the size of the available local sample data on the -th device, and represents the updated local model parameters of the -th device at the (t + 1)-th iteration; The device update function is , , where is the threshold of the local sample size of the device, is the updated local model parameter of the th device at the (t + 1)-th iteration, is the updated global model parameter at the t-th iteration.

9. The latency optimization system for edge device training in federated learning based on privacy protection according to claim 8, wherein, The model feedback process of the central server layer is as follows: After the central server completes the global model update, it evaluates the effectiveness and accuracy of the updated global model in terms of accuracy and recall through the classification performance comprehensive evaluation formula. According to the evaluation results, if the model performance reaches or exceeds the preset performance threshold, the updated global model is fed back to the edge device; The comprehensive evaluation formula for the classification performance is as follows: , where F1 is the performance threshold.

Images