Cloud switching data management method and system combined with federal learning
By using federated learning methods to manage data between cloud processing centers and edge nodes, the problem of low collaboration efficiency between the cloud and edge nodes is solved, achieving efficient data integration and secure sharing, and improving the robustness and adaptability of data management.
Patent Information
- Application Number
- CN202510984473.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-17
- Publication Date
- 2025-10-31
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing technologies suffer from low collaboration efficiency between the cloud and edge nodes, low collaboration efficiency due to data dispersion, and difficulties in data sharing and security among multiple edge nodes.
The federated learning approach is adopted to obtain learning tasks from the cloud processing center and multiple edge nodes, perform shared feature extraction, establish shared nodes, train task models and generate adversarial examples, generate lightweight edge models and adversarial example data, perform aggregation robustness testing and updates, generate aggregated models and distribute them to edge nodes.
It improves the collaboration efficiency between the cloud and edge nodes, enhances the integration efficiency and security of data management, and adapts to the needs of complex segmentation scenarios.
Smart Images
Figure CN120881066A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data management technology, and more specifically to a cloud-based data management method and system that incorporates federated learning. Background Technology
[0002] Cloud-cutting data refers to the data on the Cloud-Cutting Online Steel Plate Cutting Sharing Platform (hereinafter referred to as the Cloud-Cutting Platform). It covers various information during the steel plate cutting process, including but not limited to steel plate material, thickness, dimensions, cutting patterns, and cutting parameters. This data is efficiently collected, organized, and analyzed on the platform, providing strong support for the accuracy and efficiency of steel plate cutting and offering a scientific basis for enterprise production decisions. However, existing technologies still have some prominent problems in data management and collaborative processing, such as low collaboration efficiency due to data fragmentation and the difficulty of data sharing and security assurance among multiple edge nodes. Furthermore, how to efficiently integrate, dynamically update, and robustly test cloud-cutting data to adapt to complex cutting scenarios and further improve data management efficiency remains a pressing technical challenge. Summary of the Invention
[0003] This application provides a cloud-based data management method and system that combines federated learning, which solves the technical problem of low collaboration efficiency between cloud and edge nodes in existing technologies.
[0004] In view of the above problems, this application provides a cloud-based data management method and system that combines federated learning.
[0005] The first aspect of this application provides a cloud-based data management method incorporating federated learning, the method comprising:
[0006] The process involves: acquiring a cloud processing center and multiple edge nodes; acquiring multiple learning tasks corresponding to the edge nodes; extracting shared features from the multiple learning tasks to establish M groups of shared nodes with shared features, where M is an integer greater than or equal to 1; training task models and generating adversarial examples based on the local datasets of each edge node in the M groups of shared nodes using federated learning, resulting in M groups of lightweight edge models and M groups of adversarial example data; sending the M groups of lightweight edge models and the M groups of adversarial example data to the cloud processing center, performing aggregation robustness testing on the lightweight edge models using the M groups of adversarial example data, and updating the aggregation based on the robustness test results to generate M aggregated models; and distributing the M aggregated models to the corresponding edge nodes in the M groups of shared nodes.
[0007] A second aspect of this application provides a cloud-based data management system incorporating federated learning, the system comprising:
[0008] First resource acquisition module: acquires the cloud processing center and multiple edge nodes; Second resource acquisition module: acquires multiple learning tasks corresponding to the multiple edge nodes; Feature extraction module: extracts shared features from the multiple learning tasks to establish M groups of shared nodes with shared features, where M is an integer greater than or equal to 1; Model training module: based on federated learning, trains task models and generates adversarial examples based on the local datasets of each edge node in the M groups of shared nodes, obtaining M groups of lightweight edge models and M groups of adversarial example data; Testing module: sends the M groups of lightweight edge models and the M groups of adversarial example data to the cloud processing center, performs aggregation robustness testing on the lightweight edge models using the M groups of adversarial example data, and performs aggregation updates based on the robustness test results to generate M aggregated models; Model distribution module: distributes the M aggregated models to the corresponding edge nodes in the M groups of shared nodes.
[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:
[0010] First, the cloud processing center and multiple edge nodes are acquired. Next, multiple learning tasks corresponding to the edge nodes are acquired. Further, shared features are extracted from the multiple learning tasks, establishing M groups of shared nodes with shared features, where M is an integer greater than or equal to 1. Next, based on federated learning, task models are trained and adversarial examples are generated using the local datasets of each edge node in the M groups of shared nodes, resulting in M groups of lightweight edge models and M groups of adversarial example data. Then, the M groups of lightweight edge models and M groups of adversarial example data are sent to the cloud processing center. The aggregation robustness test of the lightweight edge models is performed using the M groups of adversarial example data. Based on the robustness test results, aggregation and updates are performed to generate M aggregated models. Finally, the M aggregated models are distributed to the corresponding edge nodes in the M groups of shared nodes. This solves the technical problem of low collaboration efficiency between the cloud and edge nodes in existing technologies, achieving the technical effect of improving the collaborative efficiency of federated learning tasks. Attached Figure Description
[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 A schematic diagram of the cloud-based data management method incorporating federated learning is provided for embodiments of this application.
[0013] Figure 2This is a schematic diagram of the cloud-based data management system architecture that incorporates federated learning, provided as an embodiment of this application.
[0014] Explanation of reference numerals in the attached diagram: First resource acquisition module 11, Second resource acquisition module 12, Feature extraction module 13, Model training module 14, Testing module 15, Model distribution module 16. Detailed Implementation
[0015] This application addresses the technical problem of low collaboration efficiency between cloud and edge nodes in existing technologies by providing a cloud-based data management method and system that combines federated learning.
[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0017] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.
[0018] Example 1, as Figure 1 As shown, this application provides a cloud-based data management method that incorporates federated learning, wherein the method includes:
[0019] Acquire cloud processing center and multiple edge nodes.
[0020] The system identifies and connects to the cloud processing center, the core node of the entire system. The cloud processing center possesses powerful computing capabilities and massive data storage capacity, primarily responsible for centralized data processing, storage, distribution, and model updates and optimization. By receiving data and task information from edge nodes, the cloud processing center performs global coordination and optimization operations, including data integration, analysis, and providing intelligent service support to edge nodes. Simultaneously, the system also identifies and connects to multiple edge nodes, typically deployed near the data source, such as sensor terminals, industrial equipment, or user terminals, possessing certain computing and storage capabilities. Edge nodes can collect and process local data in real-time or near real-time, and preprocess or perform preliminary analysis based on task requirements to alleviate the data transmission pressure on the cloud processing center.
[0021] Obtain multiple learning tasks corresponding to the multiple edge nodes.
[0022] Each edge node will have different learning tasks depending on its environment, device type, or business needs. The learning task refers to the computation or data analysis goal that needs to be completed, such as classification or prediction, and the specific task depends on the data content and application scenario of the edge node.
[0023] Shared features are extracted for the multiple learning tasks to establish M groups of shared nodes with shared features, where M is an integer greater than or equal to 1.
[0024] After acquiring multiple learning tasks, the system performs detailed shared feature extraction on these tasks to identify common features or correlations between them. Specifically, firstly, data analysis is performed on the learning task of each edge node, including the data types, data structures, input-output relationships, and target characteristics involved in the task. Then, feature extraction algorithms are used to extract and compare key features of these learning tasks to identify shared features between different tasks, such as the same data dimensions, similar computational logic, or common analysis objectives. Based on the extracted shared features, the system groups edge nodes with similar characteristics or strong correlations, establishing M groups of shared nodes, where M is an integer greater than or equal to 1. For example, for edge nodes involved in steel plate cutting tasks, the system may find that some nodes have highly similar data features (such as steel plate material, thickness, or cutting parameters), thus classifying these nodes into the same group of shared nodes. Through shared feature extraction, repeated analysis and processing of repetitive features can be avoided, improving the overall efficiency of the system. The grouped shared nodes can collaborate and share resources and information during task execution, forming a more efficient collaboration mechanism. Shared feature extraction lays the foundation for subsequent steps such as model training and task allocation optimization.
[0025] Based on federated learning, task model training and adversarial sample generation are performed using the local datasets of each edge node in the M groups of shared nodes, resulting in M groups of lightweight edge models and M groups of adversarial sample data.
[0026] Based on the federated learning method, a distributed task model is trained on the local datasets of each edge node in the M groups of shared nodes, and adversarial sample data is generated during the training process to improve the security and robustness of the model.
[0027] An initial model structure and parameter configuration are predefined in the cloud processing center and distributed to each edge node in the M shared nodes. Each edge node loads the initial model for local training based on its local dataset and task requirements. Each edge node trains the model locally using its own dataset; the training process is entirely within the edge node, and model parameter updates are based on its independent dataset to ensure data privacy is not leaked to the cloud. During local training, the system employs adversarial example generation algorithms (such as FGSM or PGD) to generate adversarial examples for the training model. These adversarial examples simulate possible attacks or anomalous inputs to test the model's robustness under malicious interference. Each edge node trained on local data receives a set of lightweight local models, called edge lightweight models. These models are small in size, designed specifically for edge computing environments, and suitable for running on resource-constrained edge nodes. After completing local training, each edge node uploads the generated edge lightweight models and adversarial example data to the cloud processing center; ultimately, M sets of edge lightweight models and M sets of adversarial example data are obtained.
[0028] Furthermore, based on federated learning, task model training and adversarial example generation are performed using the local datasets of each edge node in the M groups of shared nodes, resulting in M groups of lightweight edge models and M groups of adversarial example data, including:
[0029] Extract the first set of shared nodes from the M sets of shared nodes and train the corresponding first set of lightweight edge models; collect test error sample data for each shared node in the first set of shared nodes; constrained by a preset shared data volume threshold, calculate the loss gradient for each lightweight edge model corresponding to each shared node based on the test error sample data of each node, add perturbation along the loss gradient direction, and generate each adversarial sample dataset; add each adversarial sample dataset to the first set of adversarial sample data, and so on, to generate the M sets of adversarial sample data.
[0030] Specifically, a first set of shared nodes is extracted from the M sets of shared nodes. For each edge node in this set, the initial model is trained in the local environment using its local dataset, generating the first set of lightweight edge models. Test error sample data generated during the testing process is collected from each node in the first set of shared nodes to reflect the local adaptability and weaknesses of the model. Based on these test error sample data, the loss gradient is calculated for each lightweight model, and perturbations are added along the direction of the loss gradient to generate adversarial sample datasets for each node. When generating adversarial samples, a preset shared data volume threshold is used as a constraint to ensure that the scale of the adversarial samples is moderate and effective. The adversarial sample datasets generated by each node are integrated into the first set of adversarial sample data, and the above process is repeated for the remaining M-1 sets of shared nodes, performing lightweight model training, error sample collection, loss gradient calculation, and adversarial sample generation in sequence, until the processing of the M sets of shared nodes is completed, generating M sets of lightweight edge models and M sets of adversarial sample data.
[0031] Furthermore, extracting the first group of shared nodes from the M groups of shared nodes and training the corresponding first group of lightweight edge models includes:
[0032] Obtain the first shared task of the first group of shared nodes; perform model initialization configuration in the cloud processing center based on the first shared task to generate an initialization model; download the initialization model to the first group of shared nodes, train it using the local dataset of each node in the first group of shared nodes, and generate the first group of lightweight edge models.
[0033] Specifically, the first shared task corresponding to the first group of shared nodes is obtained. The first shared task reflects the commonalities and requirements of the tasks of each node in the first group of shared nodes, such as data features, target parameters, or analysis logic. Based on the first shared task, the model is initialized and configured in the cloud processing center. The cloud processing center presets the model structure, parameter range, and initial weights according to the requirements of the shared task, and generates an initial model suitable for the task scenario of the first group of shared nodes. The generated initial model is distributed and downloaded to each edge node in the first group of shared nodes. After receiving the initial model, each edge node uses its local dataset to further train the model, completes the model parameter optimization for the local data environment, and finally generates the first group of lightweight edge models.
[0034] The M sets of lightweight edge models and the M sets of adversarial sample data are sent to the cloud processing center. The aggregation robustness test of the lightweight edge models is performed using the M sets of adversarial sample data. Based on the robustness test results, the aggregation is updated to generate M aggregate models.
[0035] After sending M sets of lightweight edge models and M sets of adversarial example data to the cloud processing center, the center centrally processes the received data and models. Specifically, by inputting the adversarial example data into each lightweight edge model, the model's performance under malicious interference or abnormal input is evaluated, including metrics such as accuracy, stability, and anti-interference capability. Based on the results of the aggregation robustness test, the cloud processing center optimizes and updates the model parameters. By fusing the model parameters from different edge nodes, M more global and robust aggregated models are generated. These aggregated models not only integrate the learning results of multiple edge nodes but also significantly improve the model's adaptability and anti-interference capability in complex scenarios. Finally, the M aggregated models will be distributed to the corresponding edge nodes in subsequent steps to provide optimized model support for the execution of actual tasks.
[0036] Furthermore, the M sets of lightweight edge models are sent to the cloud processing center, and the aggregation robustness test of the lightweight edge models is performed using the M sets of adversarial sample data. Based on the robustness test results, the aggregation is updated to generate M aggregated models, including:
[0037] In the cloud processing center, the first group of edge lightweight models in the M groups of edge lightweight models are aggregated according to the initial aggregation weights to generate a primary aggregated model. The first group of adversarial sample data corresponding to the first group of edge lightweight models is retrieved from the M groups of adversarial sample data, and federated adversarial training is performed on the primary aggregated model to generate a primary adversarial robustness score set. Based on the primary adversarial robustness score set, abnormal edge nodes in the adversarial training are located, and the aggregation weights and adversarial sample ratios are adjusted for these abnormal edge nodes to generate updated aggregation weights and updated adversarial sample ratios. The primary aggregated model is distributed to each node in the first group of shared nodes for local training again, and after updating the adversarial samples based on the updated adversarial sample ratios, aggregation and federated adversarial training continue in the cloud processing center until a first aggregated model is obtained whose adversarial robustness score meets the preset convergence requirements. The first aggregated model is then added to the M aggregated models.
[0038] In the cloud processing center, the first group of lightweight edge models in the M groups of edge models undergoes weighted fusion of model parameters according to the initial aggregation weights to generate an initial aggregated model, laying the foundation for subsequent robustness testing and optimization. From the M groups of adversarial sample data, the first group of adversarial sample data corresponding to the first group of lightweight edge models is retrieved, and this data is used to perform federated adversarial training on the initial aggregated model, simulating an adversarial attack environment. By testing the model's performance under adversarial examples, an adversarial robustness score set is generated, containing the robustness evaluation results for each edge node. Based on this adversarial robustness score set, edge nodes that exhibit abnormal performance during federated adversarial training (i.e., nodes with significantly lower robustness than expected) are identified. For these edge nodes with abnormal adversarial training, their aggregation weights are dynamically adjusted (reducing the weights of abnormal nodes and increasing the weights of reliable nodes), and their corresponding adversarial sample proportions are updated (increasing the adversarial sample proportions to strengthen robustness training), generating an updated aggregation weight and an updated adversarial sample proportion. The initial aggregated model is distributed to each edge node in the first group of shared nodes, allowing each node to retrain locally using its local dataset. Simultaneously, adversarial sample data is updated according to the proportion of adversarial samples updated in one iteration, and model parameter aggregation and federated adversarial training are performed again in the cloud processing center. This aggregation and federated adversarial training process is iterated repeatedly until the adversarial robustness score meets the preset convergence requirements, generating the final aggregated model (i.e., the first aggregated model) corresponding to the first group of shared nodes. This model exhibits high robustness and adaptability under adversarial examples. The generated first aggregated model is added to M aggregated models, and the above process is repeated sequentially for the other M-1 groups of lightweight edge models until the aggregated model optimization for all shared nodes is completed.
[0039] Furthermore, from the M sets of adversarial sample data, the first set of adversarial sample data corresponding to the first set of edge lightweight models is retrieved, and federated adversarial training is performed on the aggregated model to generate an adversarial robust score set, including:
[0040] The first set of adversarial sample data is divided according to the corresponding nodes to determine the adversarial sample data corresponding to each node in the first set of shared nodes; the adversarial sample data is tested through the first aggregation model, the test results are recorded, and the first adversarial robustness score set is generated.
[0041] Specifically, the adversarial sample data in the first set of shared nodes is divided according to the correspondence between the nodes, ensuring that each node obtains a subset of adversarial sample data that matches its task requirements and data characteristics. For example, if a node is responsible for processing steel plate cutting tasks of a specific material, its corresponding adversarial sample data will include anomalous or perturbation data related to the characteristics of that material. The adversarial sample data subset of each node is tested using an aggregated model, and the test results of each adversarial sample are recorded, including prediction accuracy, error distribution, and the model's response to adversarial perturbations. Based on these test results, a robustness score is calculated for each shared node, reflecting the model's ability to resist interference on the adversarial sample corresponding to that node. The test scores of all shared nodes are then aggregated to form a set of adversarial robustness scores.
[0042] Furthermore, the preset convergence requirement is that the adversarial robustness scores of each aggregation model on the adversarial sample data of each corresponding group of shared nodes all meet the preset score threshold.
[0043] During federated adversarial training of the aggregated model on each group of shared nodes, the system continuously evaluates the model's robustness. Through multiple iterations, it optimizes the aggregation weights and the proportion of adversarial examples, gradually improving the model's robustness. Once the model's adversarial robustness score reaches a preset threshold on adversarial example data across all shared nodes, the training of the aggregated model is considered to have converged, possessing sufficient robustness and adaptability, and can be officially deployed as the final aggregated model to the corresponding shared nodes. By setting this convergence requirement, the system ensures that the generated aggregated model has high anti-interference capability and stability in complex scenarios, thus providing more reliable support for task execution.
[0044] Furthermore, based on the aforementioned adversarial robustness score set, abnormal edge nodes in adversarial training are located, and aggregate weights and adversarial sample ratios are adjusted for these abnormal edge nodes to generate updated aggregate weights and updated adversarial sample ratios, including:
[0045] Based on the adversarial robustness score set, edge nodes whose adversarial robustness scores do not meet the preset score threshold are selected and designated as the adversarial training abnormal edge nodes; an aggregation weight adjustment strategy and an adversarial sample ratio adjustment strategy are obtained, wherein the aggregation weight adjustment strategy is to reduce the aggregation weight of the adversarial training abnormal edge nodes by a preset step size, and the adversarial sample ratio adjustment strategy is to increase the adversarial sample ratio of the adversarial training abnormal edge nodes to other nodes by a preset compensation; based on the aggregation weight adjustment strategy and the adversarial sample ratio adjustment strategy, the aggregation weight and adversarial sample ratio of the adversarial training abnormal edge nodes are adjusted to generate the first updated aggregation weight and the first updated adversarial sample ratio.
[0046] From a set of adversarial robustness scores, edge nodes whose robustness scores do not meet a preset threshold are selected and marked as adversarial training anomalous edge nodes. These nodes represent those that perform poorly in adversarial example testing and lack robustness, and are the key targets for model optimization. Next, aggregation weight adjustment strategies and adversarial example ratio adjustment strategies are obtained. The aggregation weight adjustment strategy reduces the aggregation weight of anomalous nodes by a preset step size to reduce their negative impact on the overall aggregated model. The adversarial example ratio adjustment strategy increases the adversarial example ratio of anomalous nodes through a preset compensation mechanism, giving them more adversarial example training resources compared to other nodes, thus focusing on improving their robustness performance. Subsequently, based on the above adjustment strategies, the aggregation weight reduction operation and the adversarial example ratio increase operation are performed for each anomalous node, generating an updated aggregation weight and an updated adversarial example ratio in real time. These adjustment results will be used in subsequent model aggregation and federated adversarial training to further enhance the robustness of anomalous nodes and optimize the overall adversarial robustness performance of the model.
[0047] The M aggregation models are distributed to the corresponding edge nodes in the M groups of shared nodes.
[0048] Based on the shared node grouping information corresponding to each aggregate model, the allocation relationship between the M aggregate models and the M groups of shared nodes is confirmed. Subsequently, the cloud processing center packages and encrypts each generated aggregate model to ensure the security and integrity of the model data during transmission. Next, the cloud processing center distributes each aggregate model to each edge node in the corresponding shared node group via the network, while recording the transmission status to ensure all nodes successfully receive the data. Upon receiving the corresponding aggregate model, the edge node decrypts and loads the model locally, deploying it as the working model for the current node to execute subsequent local tasks or data processing. Through this process, the optimized aggregate models generated in the cloud can be quickly and efficiently synchronized to each shared node group, ensuring that each edge node has the latest and more robust aggregate models, providing reliable support for the execution of distributed system tasks.
[0049] In summary, the embodiments of this application have at least the following technical effects:
[0050] First, the cloud processing center and multiple edge nodes are acquired. Next, multiple learning tasks corresponding to the edge nodes are acquired. Further, shared features are extracted from the multiple learning tasks, establishing M groups of shared nodes with shared features, where M is an integer greater than or equal to 1. Next, based on federated learning, task models are trained and adversarial examples are generated using the local datasets of each edge node in the M groups of shared nodes, resulting in M groups of lightweight edge models and M groups of adversarial example data. Then, the M groups of lightweight edge models and M groups of adversarial example data are sent to the cloud processing center. The aggregation robustness test of the lightweight edge models is performed using the M groups of adversarial example data. Based on the robustness test results, aggregation and updates are performed to generate M aggregated models. Finally, the M aggregated models are distributed to the corresponding edge nodes in the M groups of shared nodes. This solves the technical problem of low collaboration efficiency between the cloud and edge nodes in existing technologies, achieving the technical effect of improving the collaborative efficiency of federated learning tasks.
[0051] Example 2, based on the same inventive concept as the cloud-based data management method incorporating federated learning in the aforementioned examples, such as... Figure 2 As shown, this application provides a cloud-based data management system that incorporates federated learning, wherein the system includes:
[0052] First resource acquisition module 11: Acquires the cloud processing center and multiple edge nodes; Second resource acquisition module 12: Acquires multiple learning tasks corresponding to the multiple edge nodes; Feature extraction module 13: Extracts shared features from the multiple learning tasks to establish M groups of shared nodes with shared features, where M is an integer greater than or equal to 1; Model training module 14: Based on federated learning, trains task models and generates adversarial examples based on the local datasets of each edge node in the M groups of shared nodes to obtain M groups of lightweight edge models and M groups of adversarial example data; Testing module 15: Sends the M groups of lightweight edge models and the M groups of adversarial example data to the cloud processing center, performs aggregation robustness testing on the lightweight edge models using the M groups of adversarial example data, and performs aggregation updates based on the robustness test results to generate M aggregate models; Model distribution module 16: Distributes the M aggregate models to the corresponding edge nodes in the M groups of shared nodes.
[0053] Furthermore, the model training module 14 is used to perform the following methods:
[0054] Extract the first set of shared nodes from the M sets of shared nodes and train the corresponding first set of lightweight edge models; collect test error sample data for each shared node in the first set of shared nodes; constrained by a preset shared data volume threshold, calculate the loss gradient for each lightweight edge model corresponding to each shared node based on the test error sample data of each node, add perturbation along the loss gradient direction, and generate each adversarial sample dataset; add each adversarial sample dataset to the first set of adversarial sample data, and so on, to generate the M sets of adversarial sample data.
[0055] Furthermore, the model training module 14 is used to perform the following methods:
[0056] Obtain the first shared task of the first group of shared nodes; perform model initialization configuration in the cloud processing center based on the first shared task to generate an initialization model; download the initialization model to the first group of shared nodes, train it using the local dataset of each node in the first group of shared nodes, and generate the first group of lightweight edge models.
[0057] Furthermore, the test module 15 is used to perform the following methods:
[0058] In the cloud processing center, the first group of edge lightweight models in the M groups of edge lightweight models are aggregated according to the initial aggregation weights to generate a primary aggregated model. The first group of adversarial sample data corresponding to the first group of edge lightweight models is retrieved from the M groups of adversarial sample data, and federated adversarial training is performed on the primary aggregated model to generate a primary adversarial robustness score set. Based on the primary adversarial robustness score set, abnormal edge nodes in the adversarial training are located, and the aggregation weights and adversarial sample ratios are adjusted for these abnormal edge nodes to generate updated aggregation weights and updated adversarial sample ratios. The primary aggregated model is distributed to each node in the first group of shared nodes for local training again, and after updating the adversarial samples based on the updated adversarial sample ratios, aggregation and federated adversarial training continue in the cloud processing center until a first aggregated model is obtained whose adversarial robustness score meets the preset convergence requirements. The first aggregated model is then added to the M aggregated models.
[0059] Furthermore, the test module 15 is used to perform the following methods:
[0060] The first set of adversarial sample data is divided according to the corresponding nodes to determine the adversarial sample data corresponding to each node in the first set of shared nodes; the adversarial sample data is tested through the first aggregation model, the test results are recorded, and the first adversarial robustness score set is generated.
[0061] Furthermore, the test module 15 is used to perform the following methods:
[0062] The preset convergence requirement is that the adversarial robustness scores of each aggregation model on the adversarial sample data of each corresponding group of shared nodes all meet the preset score threshold.
[0063] Furthermore, the test module 15 is used to perform the following methods:
[0064] Based on the adversarial robustness score set, edge nodes whose adversarial robustness scores do not meet the preset score threshold are selected and designated as the adversarial training abnormal edge nodes; an aggregation weight adjustment strategy and an adversarial sample ratio adjustment strategy are obtained, wherein the aggregation weight adjustment strategy is to reduce the aggregation weight of the adversarial training abnormal edge nodes by a preset step size, and the adversarial sample ratio adjustment strategy is to increase the adversarial sample ratio of the adversarial training abnormal edge nodes to other nodes by a preset compensation; based on the aggregation weight adjustment strategy and the adversarial sample ratio adjustment strategy, the aggregation weight and adversarial sample ratio of the adversarial training abnormal edge nodes are adjusted to generate the first updated aggregation weight and the first updated adversarial sample ratio.
[0065] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.
[0066] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0067] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.
Claims
1. A cloud-based data management method combining federated learning, characterized in that: The method includes: Acquire cloud processing centers and multiple edge nodes; Obtain multiple learning tasks corresponding to the multiple edge nodes; Shared features are extracted for the multiple learning tasks to establish M groups of shared nodes with shared features, where M is an integer greater than or equal to 1; Based on federated learning, task model training and adversarial sample generation are performed using the local datasets of each edge node in the M groups of shared nodes, resulting in M groups of lightweight edge models and M groups of adversarial sample data. The M sets of lightweight edge models and the M sets of adversarial sample data are sent to the cloud processing center. The aggregation robustness test of the lightweight edge models is performed using the M sets of adversarial sample data. Based on the robustness test results, the aggregation is updated to generate M aggregate models. The M aggregation models are distributed to the corresponding edge nodes in the M groups of shared nodes.
2. The cloud-based data management method combining federated learning as described in claim 1, characterized in that, Based on federated learning, task model training and adversarial example generation are performed using the local datasets of each edge node in the M groups of shared nodes, resulting in M groups of lightweight edge models and M groups of adversarial example data, including: Extract the first group of shared nodes from the M groups of shared nodes, and train the corresponding first group of lightweight edge models; Each shared node in the first group of shared nodes collects test error sample data for each node. With a preset shared data volume threshold as a constraint, loss gradient calculation is performed on each edge lightweight model corresponding to each shared node based on the test error sample data of each node, and perturbation is added along the loss gradient direction to generate each adversarial sample dataset. Each of the adversarial sample datasets is added to the first set of adversarial sample data, and so on, to generate the M sets of adversarial sample data.
3. The cloud-based data management method combining federated learning as described in claim 2, characterized in that, Extract the first group of shared nodes from the M groups of shared nodes, and train the corresponding first group of lightweight edge models, including: Obtain the first shared task from the first group of shared nodes; Based on the first shared task, the model initialization configuration is performed in the cloud processing center to generate an initialization model; The initialization model is downloaded to the first group of shared nodes, and trained using the local datasets of each node in the first group of shared nodes to generate the first group of lightweight edge models.
4. The cloud-based data management method combining federated learning as described in claim 1, characterized in that, The M sets of lightweight edge models are sent to the cloud processing center. An aggregation robustness test of the lightweight edge models is performed using the M sets of adversarial sample data. Based on the robustness test results, the models are aggregated and updated to generate M aggregated models, including: In the cloud processing center, the first group of edge lightweight models in the M groups of edge lightweight models are aggregated according to the initial aggregation weight to generate a primary aggregated model; From the M sets of adversarial sample data, retrieve the first set of adversarial sample data corresponding to the first set of edge lightweight models, perform federated adversarial training on the first aggregate model, and generate an adversarial robust score set. Based on the adversarial robust score set, abnormal edge nodes of adversarial training are located, and the aggregation weight and adversarial sample ratio are adjusted for the abnormal edge nodes of adversarial training to generate an updated aggregation weight and an updated adversarial sample ratio. The first aggregation model is distributed to each node in the first group of shared nodes for local training again. After the adversarial sample is updated based on the first update adversarial sample ratio, aggregation and federated adversarial training are continued in the cloud processing center until the first aggregation model with adversarial robustness score meets the preset convergence requirements is obtained. Add the first aggregation model into the M aggregation models.
5. The cloud-based data management method combining federated learning as described in claim 4, characterized in that, From the M sets of adversarial sample data, retrieve the first set of adversarial sample data corresponding to the first set of edge lightweight models, perform federated adversarial training on the aggregated model, and generate an adversarial robust score set, including: The first group of adversarial sample data is divided according to the corresponding nodes, and the adversarial sample data corresponding to each node in the first group of shared nodes is determined. The adversarial sample data are tested using the first-order aggregation model, and the test results are recorded to generate the first-order adversarial robustness score set.
6. The cloud-based data management method combining federated learning as described in claim 5, characterized in that, The preset convergence requirement is that the adversarial robustness scores of each aggregation model on the adversarial sample data of each corresponding group of shared nodes all meet the preset score threshold.
7. The cloud-based data management method combining federated learning as described in claim 6, characterized in that, Based on the aforementioned adversarial robustness score set, abnormal edge nodes in adversarial training are located. For these abnormal edge nodes, aggregate weights and adversarial sample ratios are adjusted to generate updated aggregate weights and updated adversarial sample ratios, including: Based on the set of adversarial robust scores, edge nodes whose adversarial robust scores do not meet the preset score threshold are selected and designated as the adversarial training abnormal edge nodes; The aggregation weight adjustment strategy and the adversarial sample ratio adjustment strategy are obtained. The aggregation weight adjustment strategy is to reduce the aggregation weight of the adversarial training abnormal edge node by a preset step size. The adversarial sample ratio adjustment strategy is to increase the adversarial sample ratio of the adversarial training abnormal edge node to other nodes by a preset compensation. Based on the aggregation weight adjustment strategy and the adversarial sample ratio adjustment strategy, the aggregation weight and adversarial sample ratio are adjusted for the adversarial training abnormal edge nodes to generate the first-updated aggregation weight and the first-updated adversarial sample ratio.
8. A cloud-based data management system incorporating federated learning, characterized in that: The system is used to implement the cloud-based data management method incorporating federated learning as described in any one of claims 1-7, the system comprising: First resource acquisition module: Acquires cloud processing center and multiple edge nodes; Second resource acquisition module: acquires multiple learning tasks corresponding to the multiple edge nodes; Feature extraction module: Performs shared feature extraction on the multiple learning tasks and establishes M groups of shared nodes with shared features, where M is an integer greater than or equal to 1; Model training module: Based on federated learning, task model training and adversarial sample generation are performed using the local datasets of each edge node in the M groups of shared nodes, resulting in M groups of lightweight edge models and M groups of adversarial sample data; Testing module: Sends the M sets of lightweight edge models and the M sets of adversarial sample data to the cloud processing center, performs aggregation robustness testing on the lightweight edge models using the M sets of adversarial sample data, and performs aggregation updates based on the robustness test results to generate M aggregate models; Model distribution module: Distributes the M aggregated models to the corresponding edge nodes in the M groups of shared nodes.