Federal learning global model training method and system based on active learning and model compression
By adopting active learning and model compression techniques in federated learning, optimizing sample selection and model transmission, the problems of data redundancy and high communication costs in traditional federated learning methods are solved, and efficient global model training and model updates are achieved.
Patent Information
- Application Number
- CN202510151270.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-11
- Publication Date
- 2025-06-17
AI Technical Summary
The traditional federated learning global model training method has problems such as data redundancy, high communication costs, and low model update efficiency, especially when the model is large or the number of local nodes is large.
The federated learning global model training method based on active learning and model compression is adopted. Through modules such as data preprocessing, active learning sample selection, model compression and global model aggregation, sample selection and model transmission are optimized to improve training efficiency and model performance.
Through active learning, selecting high-value samples for training, reducing interference from low-value samples and improving training efficiency; model compression reduces the size of model parameters and transmission costs, and reduces communication costs; system integration and linkage control modules ensure smooth communication and efficient collaboration between various parts of the system.
Smart Images
Figure CN120163262A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of model training, and particularly to a method and system for training a global model of federated learning based on active learning and model compression. Background Art
[0002] Active learning is a machine learning method that allows a learning system to train by selectively querying the most informative samples, rather than passively accepting all available data. This method aims to achieve as high a model performance as possible with the least amount of labeled data. Through strategies such as uncertainty sampling and diversity sampling, active learning can preferentially select those samples that are most valuable for model training, thereby improving training efficiency and model accuracy. Model compression refers to the technology of reducing the size or computational complexity of a model without significantly sacrificing the model performance. This includes methods such as pruning, quantization, and low-rank decomposition. Model compression helps to reduce the cost of model storage and transmission, making the model easier to deploy and run on resource-constrained devices. Federated learning is a distributed machine learning method that allows multiple local nodes to jointly train a global model while keeping the data local. Each local node uses its local data to train the model and then sends the training results to a central server for aggregation to generate a new global model. This method not only protects data privacy but also utilizes distributed data resources to improve model performance.
[0003] In traditional methods for training a global model of federated learning, local nodes use all local data for training and directly upload the trained model parameters to the central server for aggregation. Although this method is simple and effective, it has problems such as data redundancy, high communication cost, and low model update efficiency. Local nodes may contain a large number of low-value samples that contribute little to model training, resulting in waste of training resources. Transmitting complete model parameter updates may occupy a large amount of bandwidth, especially when the model is large or the number of local nodes is large. Since all local nodes participate in training, the central server needs to process a large number of model parameter updates, resulting in a slow aggregation process.
[0004] Based on this, the present invention provides a method and system for training a global model of federated learning based on active learning and model compression to solve the above-mentioned technical problems. Summary of the Invention
[0005] The purpose of the present invention is to provide a method and system for training a global model of federated learning based on active learning and model compression to solve the problems raised in the above background art.
[0006] To achieve the above purpose, the present invention provides the following technical solutions:
[0007] The present invention proposes a federated learning global model training system based on active learning and model compression, including a data preprocessing module, a local model training module, a model compression and transmission module, a global model aggregation module, a global model update module, an active learning strategy adjustment module, a model compression algorithm optimization module, and a system integration and linkage module;
[0008] The data preprocessing module is used for active learning sample selection;
[0009] The local model training module is used for model initialization and local training execution;
[0010] The model compression and transmission module is used for model compression and model transmission;
[0011] The global model aggregation module is used for receiving model parameters and aggregating model parameters;
[0012] The global model update module is used for model update decision-making and model update execution;
[0013] The active learning strategy adjustment module is used for learning strategy evaluation and learning strategy adjustment;
[0014] The model compression algorithm optimization module is used for compression effect evaluation and compression algorithm optimization;
[0015] The system integration and linkage module is used for system integration and system linkage control.
[0016] Preferably, the data preprocessing module further includes a data collection and distribution unit and an active learning sample selection unit;
[0017] The data collection and distribution unit is composed of a distributed network, and is used for collecting raw data from each client, and distributing the data to the corresponding local model training nodes according to the federated learning strategy, performing preliminary screening and formatting of the data, and ensuring the consistency and availability of the data;
[0018] The active learning sample selection unit is composed of an algorithm strategy, and is used for selecting the most valuable samples for model training from the local data, scoring and screening the samples through uncertainty sampling and diversity sampling strategies, and marking the high-value samples as those that need to be uploaded to the global model for further training.
[0019] Preferably, the local model training module further includes a model initialization unit and a local training execution unit;
[0020] The model initialization unit is composed of receiving global model parameters, and is used for initializing the model locally, ensuring the consistency of the local model and the global model, performing parameter setting and initialization of the model, and preparing for local training;
[0021] The local training execution unit is composed of local computing resources and is used to execute the training process of the local model. It performs iterative training using selected local data, conducts forward propagation, loss calculation, backpropagation, and parameter update of the model to improve the accuracy of the local model.
[0022] Preferably, the model compression and transmission module further includes a model compression unit and a model transmission unit;
[0023] The model compression unit is composed of a compression algorithm and is used to compress the locally trained model, reduce the number or precision of model parameters, lower the transmission cost, and perform pruning, quantization, or low-rank decomposition operations on the model to generate a compressed model;
[0024] The model transmission unit is composed of a network communication protocol and is used to securely and efficiently transmit the compressed local model parameters to the global model server, perform data encryption, transmission, and decryption to ensure the security and integrity of the model parameters during transmission.
[0025] Preferably, the global model aggregation module further includes a model parameter receiving unit and a model parameter aggregation unit;
[0026] The model parameter receiving unit is composed of a network interface and is used to receive the compressed model parameters from each local node, perform storage and preprocessing, and perform data decoding and verification to ensure the correctness and integrity of the received model parameters;
[0027] The model parameter aggregation unit is composed of an aggregation algorithm and is used to aggregate the parameters of each received local model to generate new global model parameters, perform weighted average, federated average, or more complex aggregation strategies on the parameters to improve the generalization ability of the global model.
[0028] Preferably, the global model update module further includes a model update decision-making unit and a model update execution unit;
[0029] The model update decision-making unit is composed of an evaluation algorithm and is used to evaluate the performance improvement of the global model after aggregation, decide whether to update the model, evaluate the accuracy, loss function value, or other metrics of the model, and generate an update decision;
[0030] The model update execution unit is composed of a model management system and is used to execute the update operation of the global model, apply the new model parameters to the global model, perform parameter replacement, version management, and logging of the model to ensure the traceability and controllability of the global model update process.
[0031] Preferably, the active learning strategy adjustment module further includes a learning strategy evaluation unit and a learning strategy adjustment unit;
[0032] The learning strategy evaluation unit is composed of performance monitoring, used to evaluate the effect of the current active learning strategy, including the quality of sample selection and the efficiency of model training, calculate performance indicators of the strategy, conduct comparative analysis and generate feedback, providing a basis for strategy adjustment;
[0033] The learning strategy adjustment unit is composed of algorithm optimization, used to adjust the active learning strategy according to the evaluation results, optimize the sample selection algorithm or parameter settings, conduct parameter adjustment, algorithm improvement or strategy switching of the strategy, and improve the adaptability and effectiveness of the active learning strategy.
[0034] Preferably, the model compression algorithm optimization module further includes a compression effect evaluation unit and a compression algorithm optimization unit;
[0035] The compression effect evaluation unit is composed of compression ratio and model performance monitoring, used to evaluate the effect of the current model compression algorithm, including the compression ratio and the accuracy of the compressed model, conduct quantitative analysis, comparative experiments and feedback generation of the compression effect, providing a basis for algorithm optimization;
[0036] The compression algorithm optimization unit is composed of algorithm research and experiments, used to optimize the model compression algorithm according to the evaluation results, improve the compression ratio and the performance of the compressed model, conduct algorithm improvement, exploration of new algorithms or parameter adjustment, and improve the efficiency and effect of the model compression algorithm.
[0037] Preferably, the system integration and linkage module further includes a system integration unit and a system linkage control unit;
[0038] The system integration unit is composed of modular design, used to integrate each module into a unified federated learning system, ensure the compatibility of interfaces between modules and data circulation, conduct system architecture design, module integration and test verification, and ensure the integrity and stability of the system;
[0039] The system linkage control unit is composed of control logic and process management, used to coordinate the work processes and data interactions of each module, realize the automated and linked operation of the system, conduct process control, data scheduling and exception handling, and ensure the smooth operation and high-efficiency collaboration of the system.
[0040] Based on the above system, the present invention also proposes a global model training method for federated learning based on active learning and model compression, including the following steps:
[0041] S1. The system starts, begins to collect raw data from each client, and distributes the data to the corresponding local training nodes according to the federated learning strategy, and conducts preliminary data screening and formatting;
[0042] S2. Apply the sample selection strategy to select the most valuable samples for model training from the local data. Score and screen the samples through a specific strategy, and mark the high-value samples for uploading.
[0043] S3. Receive the global model parameters, initialize the model locally to ensure consistency with the global model, and use the selected local data for iterative training, including forward propagation, loss calculation, backward propagation, and parameter update.
[0044] S4. Compress the locally trained model to reduce the number of model parameters or precision to lower the transmission cost.
[0045] S5. Securely and efficiently transmit the compressed model parameters to the global model server, and perform data encryption during the process.
[0046] S6. Receive the compressed model parameters from each local node, perform storage and preprocessing, including data decoding and verification, aggregate the received model parameters, and generate new global model parameters.
[0047] S7. Evaluate the performance improvement of the global model after aggregation, decide whether to update the model. If it is decided to update, apply the new model parameters to the global model.
[0048] S8. Evaluate the effectiveness of the current sample selection strategy, including sample quality and training efficiency. According to the evaluation results, adjust the strategy to optimize sample selection.
[0049] S9. Evaluate the effectiveness of the current model compression algorithm, including compression ratio and the accuracy of the compressed model. According to the evaluation results, improve or adjust the algorithm to enhance the compression efficiency and effect.
[0050] S10. Ensure the interface compatibility and data flow between different parts of the system, coordinate the work process and data interaction, realize the automation and linkage operation of the system, and be used for process control, data scheduling, and exception handling to ensure the smooth operation and efficient cooperation of the system.
[0051] Compared with the prior art, the beneficial effects of the present invention are:
[0052] The present invention selects high-value samples for training through active learning, reducing the interference of low-value samples on the training process and improving the training efficiency. The model compression technology reduces the size of model parameters and the transmission cost, enabling efficient federated learning in resource-constrained environments. The system integration and linkage control module ensures smooth communication and efficient cooperation among various parts of the system, enabling the system to be easily extended to more local nodes. The active learning strategy can accurately select the samples most valuable for model training, improving data utilization. The model compression technology significantly reduces the number of model parameters to be transmitted, reducing the communication cost. Through the global model aggregation and update decision module, the system can dynamically adjust and optimize the global model to maintain high performance in a changing data environment. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 FIG. is a topology diagram of a global model training system for federated learning based on active learning and model compression according to the present invention;
[0054] Figure 2 FIG. is a method for training a global model of federated learning based on active learning and model compression according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0055] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. 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.
[0056] Embodiment 1
[0057] Please refer to Figure 1 , the present invention proposes a global model training system for federated learning based on active learning and model compression. This system includes a data preprocessing module, a local model training module, a model compression and transmission module, a global model aggregation module, a global model update module, an active learning strategy adjustment module, a model compression algorithm optimization module, and a system integration and linkage module;
[0058] Among them, it should be noted that the data preprocessing module is used for active learning sample selection, the local model training module is used for model initialization and local training execution, the model compression and transmission module is used for model compression and model transmission, the global model aggregation module is used for model parameter reception and model parameter aggregation, the global model update module is used for model update decision and model update execution, the active learning strategy adjustment module is used for learning strategy evaluation and learning strategy adjustment, the model compression algorithm optimization module is used for compression effect evaluation and compression algorithm optimization, and the system integration and linkage module is used for system integration and system linkage control.
[0059] It should also be noted in this embodiment that the data preprocessing module further includes a data collection and distribution unit and an active learning sample selection unit;
[0060] Furthermore, the data collection and distribution unit is composed of a distributed network, which is used to collect raw data from each client and distribute the data to the corresponding local model training nodes according to the strategy of federated learning, perform preliminary screening and formatting of the data, and ensure the consistency and availability of the data;
[0061] Furthermore, the active learning sample selection unit is composed of algorithmic strategies, which is used to select the most valuable samples for model training from the local data, perform scoring and screening of the samples through uncertainty sampling and diversity sampling strategies, and mark the high-value samples as those that need to be uploaded to the global model for further training.
[0062] It should also be noted in this embodiment that the local model training module further includes a model initialization unit and a local training execution unit;
[0063] Furthermore, the model initialization unit is composed of receiving global model parameters, which is used to initialize the model locally, ensure the consistency between the local model and the global model, perform parameter setting and initialization of the model, and prepare for local training;
[0064] Furthermore, the local training execution unit is composed of local computing resources, which is used to execute the training process of the local model, perform iterative training using the selected local data, perform forward propagation, loss calculation, backpropagation and parameter update of the model, and improve the accuracy of the local model.
[0065] It should also be noted in this embodiment that the model compression and transmission module further includes a model compression unit and a model transmission unit;
[0066] Furthermore, the model compression unit is composed of compression algorithms, which is used to compress the locally trained model, reduce the number or precision of model parameters, reduce the transmission cost, perform pruning, quantization or low-rank decomposition operations on the model, and generate the compressed model;
[0067] Furthermore, the model transmission unit is composed of network communication protocols, which is used to securely and efficiently transmit the compressed local model parameters to the global model server, perform encryption, transmission and decryption of the data, and ensure the security and integrity of the model parameters during transmission.
[0068] It should also be noted in this embodiment that the global model aggregation module further includes a model parameter receiving unit and a model parameter aggregation unit;
[0069] Further, the model parameter receiving unit is composed of a network interface and is used to receive the compressed model parameters from each local node, store and preprocess them, decode and verify the data to ensure the correctness and integrity of the received model parameters;
[0070] Further, the model parameter aggregation unit is composed of an aggregation algorithm and is used to aggregate the parameters of each received local model to generate new global model parameters, perform weighted averaging, federated averaging or more complex aggregation strategies of the parameters to improve the generalization ability of the global model.
[0071] It should also be noted in this embodiment that the global model update module further includes a model update decision-making unit and a model update execution unit;
[0072] Further, the model update decision-making unit is composed of an evaluation algorithm and is used to evaluate the performance improvement of the global model after aggregation, decide whether to update the model, evaluate the accuracy, loss function value or other metrics of the model, and generate an update decision;
[0073] Further, the model update execution unit is composed of a model management system and is used to execute the update operation of the global model, apply the new model parameters to the global model, perform parameter replacement, version management and log recording of the model to ensure the traceability and controllability of the global model update process.
[0074] It should also be noted in this embodiment that the active learning strategy adjustment module further includes a learning strategy evaluation unit and a learning strategy adjustment unit;
[0075] Further, the learning strategy evaluation unit is composed of performance monitoring and is used to evaluate the effect of the current active learning strategy, including the quality of sample selection and the efficiency of model training, calculate performance metrics of the strategy, conduct comparative analysis and generate feedback to provide a basis for strategy adjustment;
[0076] Further, the learning strategy adjustment unit is composed of algorithm optimization and is used to adjust the active learning strategy according to the evaluation results, optimize the sample selection algorithm or parameter settings, perform parameter adjustment, algorithm improvement or strategy switching of the strategy to improve the adaptability and effectiveness of the active learning strategy.
[0077] It should also be noted in this embodiment that the model compression algorithm optimization module further includes a compression effect evaluation unit and a compression algorithm optimization unit;
[0078] Further, the compression effect evaluation unit is composed of a compression ratio and model performance monitoring and is used to evaluate the effect of the current model compression algorithm, including the compression ratio and the accuracy of the compressed model, conduct quantitative analysis of the compression effect, comparative experiments and generate feedback to provide a basis for algorithm optimization;
[0079] Furthermore, the compression algorithm optimization unit is composed of algorithm research and experiments, and is used to optimize the model compression algorithm according to the evaluation results, improve the compression ratio and the performance of the compressed model, improve the algorithm, explore new algorithms or adjust parameters, so as to improve the efficiency and effect of the model compression algorithm.
[0080] It should also be noted in this embodiment that the system integration and linkage module further includes a system integration unit and a system linkage control unit;
[0081] Furthermore, the system integration unit is composed of modular design and is used to integrate each module into a unified federated learning system, ensure the interface compatibility and data flow between modules, conduct system architecture design, module integration and test verification, and ensure the integrity and stability of the system;
[0082] Furthermore, the system linkage control unit is composed of control logic and process management, and is used to coordinate the work processes and data interactions of each module, realize the automation and linkage operation of the system, conduct process control, data scheduling and exception handling, and ensure the smooth operation and efficient collaboration of the system.
[0083] Embodiment 2
[0084] Please refer to Figure 2 , in practical applications, based on the above-mentioned federated learning global model training method of the system, specifically, it includes the following steps:
[0085] S1. System startup, data collection and distribution
[0086] System initialization: Start the federated learning system and load necessary configuration files and system parameters;
[0087] Client data collection: Collect raw data from each client to ensure the diversity and representativeness of the data;
[0088] Data distribution strategy: According to the federated learning strategy, distribute the data to the corresponding local training nodes, considering data privacy and security;
[0089] Initial data screening: Conduct initial screening on the data distributed to the local nodes to remove invalid or abnormal data;
[0090] Data formatting: Format the screened data into the format required for model training to ensure data consistency;
[0091] S2. Application of sample selection strategy
[0092] Strategy initialization: Load sample selection strategies, including uncertainty sampling and diversity sampling;
[0093] Sample Scoring: Score the local data according to the strategy to evaluate the value of each sample for model training;
[0094] Sample Screening: According to the scoring results, screen out high-value samples for subsequent model training;
[0095] Sample Marking: Mark the screened high-value samples as those that need to be uploaded to the global model for further training;
[0096] Sample Storage: Store the marked samples locally and prepare to upload them to the global model server;
[0097] S3. Local Model Initialization and Training
[0098] Receive Global Model Parameters: Receive the latest model parameters from the global model server for local model initialization;
[0099] Model Initialization: Initialize the model locally according to the received global model parameters to ensure consistency with the global model;
[0100] Data Preparation: Use the selected local data for necessary preprocessing, including data augmentation and normalization;
[0101] Iterative Training: Execute the training process of the local model, including forward propagation, loss calculation, backpropagation, and parameter update;
[0102] Training Monitoring: Monitor the loss function values and accuracy metrics during the training process to ensure the effectiveness of model training;
[0103] S4. Model Compression
[0104] Compression Algorithm Selection: Select a suitable compression algorithm according to the model size and transmission cost requirements, including pruning, quantization, or low-rank factorization;
[0105] Model Parameter Compression: Apply the selected compression algorithm to compress the locally trained model to reduce the number or precision of model parameters;
[0106] Compression Effect Evaluation: Evaluate the size of the compressed model and the transmission cost to ensure that the compression effect meets the requirements;
[0107] Compressed Model Storage: Store the compressed model locally and prepare to upload it to the global model server;
[0108] S5. Model Parameter Transmission
[0109] Data Encryption: Encrypt the compressed model parameters to ensure security during transmission;
[0110] Network Transmission: Transmit the encrypted model parameters to the global model server through a secure and efficient network communication protocol;
[0111] Transmission Monitoring: Monitor network latency and packet loss rate metrics during the transmission process to ensure transmission stability;
[0112] Transmission Confirmation: The global model server receives and confirms the receipt of the model parameters to ensure transmission integrity;
[0113] S6. Aggregation of Global Model Parameters
[0114] Parameter Receiving: The global model server receives the compressed model parameters from each local node;
[0115] Parameter Storage: Store the received model parameters in preparation for preprocessing and aggregation;
[0116] Parameter Decoding: Decode the received encrypted model parameters to restore the original model parameters;
[0117] Parameter Verification: Verify the correctness and integrity of the decoded model parameters to ensure the accuracy of aggregation;
[0118] Parameter Aggregation: Aggregate the verified model parameters to generate new global model parameters;
[0119] S7. Global Model Update Decision and Execution
[0120] Performance Evaluation: Evaluate the performance improvement of the global model after aggregation, including accuracy and loss function value metrics;
[0121] Update Decision: Decide whether to update the model based on the performance evaluation results. If the performance improvement is significant, decide to update the global model;
[0122] Model Update: Apply the new global model parameters to the global model to replace the old model parameters;
[0123] Version Management: Manage the version of the global model, record the model update history and version information;
[0124] Logging: Record the process, results and relevant information of the model update for subsequent analysis and traceability;
[0125] S8. Optimization of Sample Selection Strategy
[0126] Strategy Effect Evaluation: Evaluate the effect of the current sample selection strategy, including sample quality and training efficiency;
[0127] Problem Analysis: Analyze the problems and deficiencies of the strategy, including sample selection bias and low training efficiency;
[0128] Strategy adjustment: Based on the evaluation results and analyzed problems, adjust the strategy parameters or algorithms to optimize the sample selection process;
[0129] Strategy testing: Test the adjusted strategy to verify its effectiveness and improvement points;
[0130] Strategy deployment: Deploy the strategy that passes the test into the system for the subsequent sample selection process;
[0131] S9. Optimization of Model Compression Algorithm
[0132] Algorithm effectiveness evaluation: Evaluate the effectiveness of the current model compression algorithm, including the compression ratio and the accuracy of the compressed model;
[0133] Problem analysis: Analyze the problems and deficiencies of the algorithm, including low compression ratio and decreased accuracy of the compressed model;
[0134] Algorithm improvement: Based on the evaluation results and analyzed problems, improve the algorithm or adjust the parameters to enhance the compression efficiency and effectiveness;
[0135] Algorithm testing: Test the improved algorithm to verify its effectiveness and improvement points;
[0136] Algorithm deployment: Deploy the algorithm that passes the test into the system for the subsequent model compression process;
[0137] S10. System Integration and Coordinated Operation
[0138] Interface compatibility: Ensure the interface compatibility between different parts of the system to achieve smooth data flow and interaction;
[0139] Workflow coordination: Coordinate the workflows of different parts of the system to ensure the orderly progress and efficient collaboration of each step;
[0140] Data scheduling: Reasonably schedule data according to the system requirements and resource conditions to ensure the timeliness and availability of data;
[0141] Exception handling: Monitor the abnormal situations during the system operation, handle and recover them in a timely manner to ensure the stability of the system;
[0142] Automated operation: Implement the automated operation of the system to reduce manual intervention and improve the efficiency and accuracy of the system.
[0143] In the description of this specification, the descriptions referring to the terms "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in a suitable manner in any one or more embodiments or examples.
[0144] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.
Claims
1. A federated learning global model training system based on active learning and model compression, characterized by: It includes data preprocessing module, local model training module, model compression transmission module, global model aggregation module, global model update module, active learning strategy adjustment module, model compression algorithm optimization module and system integration linkage module; The data preprocessing module is used for active learning sample selection; The local model training module is used for model initialization and local training execution; The model compression and transmission module is used for model compression and model transmission; The global model aggregation module is used for receiving and aggregating model parameters; The global model updating module is used for model updating decision and model updating execution; The active learning strategy adjustment module is used for learning strategy evaluation and learning strategy adjustment; The model compression algorithm optimization module is used for compression effect evaluation and compression algorithm optimization; The system integration and linkage module is used for system integration and system linkage control.
2. The global model training system for federated learning based on active learning and model compression according to claim 1, characterized in that: The data preprocessing module also includes a data collection and distribution unit and an active learning sample selection unit; The data collection and distribution unit is used to collect raw data from each client and distribute the data to the corresponding local model training node according to the federated learning strategy; The active learning sample selection unit is used to select the most valuable samples for model training in local data.
3. The global model training system for federated learning based on active learning and model compression according to claim 2, characterized in that: The local model training module also includes a model initialization unit and a local training execution unit; The model initialization unit is used to initialize the model locally and perform parameter setting and initialization of the model; The local training execution unit is used to execute the training process of the local model, use the selected local data for iterative training, and perform forward propagation, loss calculation, back propagation and parameter update of the model.
4. The global model training system for federated learning based on active learning and model compression according to claim 3, characterized in that: The model compression and transmission module also includes a model compression unit and a model transmission unit; The model compression unit is used to compress the locally trained model; The model transmission unit is used to securely transmit the compressed local model parameters to the global model server and perform data encryption, transmission and decryption.
5. The global model training system for federated learning based on active learning and model compression according to claim 4, characterized in that: The global model aggregation module also includes a model parameter receiving unit and a model parameter aggregation unit; The model parameter receiving unit is used to receive the compression model parameters from each local node, store and pre-process, decode and verify; The model parameter aggregation unit is used to aggregate the received parameters of each local model to generate new global model parameters.
6. The global model training system for federated learning based on active learning and model compression according to claim 5, characterized in that: The global model updating module also includes a model updating decision unit and a model updating execution unit; The model update decision unit is used to evaluate the performance improvement of the global model after aggregation and decide whether to update the model; The model update execution unit is used to execute the update operation of the global model and apply the new model parameters to the global model.
7. The global model training system for federated learning based on active learning and model compression according to claim 6, characterized in that: The active learning strategy adjustment module also includes a learning strategy evaluation unit and a learning strategy adjustment unit; The learning strategy evaluation unit is used to evaluate the effect of the current active learning strategy, including the quality of sample selection and the efficiency of model training, and to calculate the performance indicators, conduct comparative analysis and generate feedback for the strategy; The learning strategy adjustment unit is used to adjust the active learning strategy according to the evaluation result, and optimize the sample selection algorithm or parameter setting.
8. The global model training system for federated learning based on active learning and model compression according to claim 7, characterized in that: The model compression algorithm optimization module also includes a compression effect evaluation unit and a compression algorithm optimization unit; The compression effect evaluation unit is used to evaluate the effect of the current model compression algorithm, including the compression ratio and the accuracy of the compressed model, and to perform quantitative analysis, comparative experiments and feedback generation of the compression effect; The compression algorithm optimization unit is constructed through algorithm research and experiments, and is used to optimize the model compression algorithm according to the evaluation results.
9. The global model training system for federated learning based on active learning and model compression according to claim 8, characterized in that: The system integration linkage module also includes a system integration unit and a system linkage control unit; The system integration unit is used to integrate various modules into a unified federated learning system to perform system architecture design, module integration and test verification; The system linkage control unit is used to coordinate the workflow and data interaction of each module, realize the automation and linkage operation of the system, and perform process control, data scheduling and exception processing.
10. A global model training method for federated learning based on active learning and model compression, applied to a global model training system for federated learning based on active learning and model compression as claimed in any one of claims 1 to 9, characterized in that: The following steps are involved: S1. The system starts to collect raw data from various clients and distributes the data to the corresponding local training nodes according to the federated learning strategy for preliminary data screening and formatting. S2. Apply sample selection strategies to select the most valuable samples for model training in local data, score and filter samples through specific strategies, and mark high-value samples for uploading; S3. Receive global model parameters, initialize the model locally, and use the selected local data for iterative training, including forward propagation, loss calculation, back propagation, and parameter update; S4. compress the locally trained model to reduce the number or accuracy of model parameters to reduce transmission costs; S5. The compressed model parameters are transmitted to the global model server securely and efficiently, and data encryption is performed during the process; S6. Receive the compressed model parameters from each local node, store and preprocess them, including data decoding and verification, aggregate the received model parameters, and generate new global model parameters; S7. Evaluate the performance improvement of the global model after aggregation and decide whether to update the model. If so, apply the new model parameters to the global model. S8. Evaluate the effectiveness of the current sample selection strategy, including sample quality and training efficiency, and adjust the strategy to optimize sample selection based on the evaluation results; S9. Evaluate the effect of the current model compression algorithm, including the compression ratio and the accuracy of the compressed model, and improve or adjust the algorithm to improve the compression efficiency and effect based on the evaluation results; S10. Ensure interface compatibility and data flow between various parts of the system, coordinate workflow and data interaction, realize system automation and linkage operation, and be used for process control, data scheduling and exception handling to ensure smooth system operation and efficient collaboration.