Federal learning and differential privacy fusion data security processing method based on smart education
By employing edge nodes for local data preprocessing and differential privacy protection through gradient importance noise addition, combined with dynamic privacy budgeting and federated multi-task learning, the problems of secure sharing of educational data and model performance degradation are solved, achieving efficient and secure collaborative data analysis.
Patent Information
- Application Number
- CN202511834800.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-17
AI Technical Summary
In smart education, centralized processing of educational data poses a single point of leakage risk, data cannot be effectively shared, and ordinary federated learning is not optimized for the heterogeneity of educational data, resulting in degraded model performance and a lack of privacy hardening, which poses a risk of privacy inference.
Local data preprocessing and feature extraction are performed at edge nodes. Differential privacy protection is achieved by dynamically adding noise based on gradient importance. A dynamic privacy budget allocation strategy is adopted, and model training and parameter aggregation are performed in conjunction with a federated multi-task learning framework. Global model generation and deployment are carried out using a central server.
It achieves improved model prediction accuracy, reduced cross-institutional collaboration costs, shortened model iteration cycles, and met the diverse needs of complex smart education scenarios while providing strict data privacy guarantees.
Smart Images
Figure CN121682885A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information technology, and in particular to a data security processing method based on federated learning and differential privacy fusion for smart education. Background Technology
[0002] With the deepening development of educational informatization, smart education has become a core direction of modern education. However, when using sensitive information such as student learning data and teacher teaching data for personalized recommendations and learning outcome prediction, serious data security and privacy challenges are faced.
[0003] Specifically, this manifests in several ways: educational data contains a large amount of sensitive personal information. Under the traditional centralized processing model, data is aggregated to a central server, posing a single point of leakage risk. Once a leakage occurs, it will endanger the rights and interests of tens of thousands of students; data between different schools and educational platforms is fragmented and cannot be effectively shared and collaborated on, limiting the cross-regional allocation of high-quality educational resources and the construction of personalized educational models; although ordinary federated learning achieves localized data training, it has not been optimized for the heterogeneity of educational data, resulting in a significant decline in model performance, and lacks targeted privacy reinforcement, still posing a risk of privacy inference.
[0004] Therefore, it is necessary to provide a data security processing method based on federated learning and differential privacy fusion for smart education to solve the above problems. Summary of the Invention
[0005] The purpose of this invention is to provide a data security processing method based on federated learning and differential privacy fusion for smart education, so as to solve the problems existing in the prior art.
[0006] To achieve the above objectives, this invention provides a data security processing method based on federated learning and differential privacy fusion for smart education, comprising the following steps: preprocessing and feature extraction of local educational data through edge nodes of various educational institutions, and training local models based on the processed data to generate local model parameters; dynamically adding noise to different parameters based on the gradient importance of the local model parameters to achieve differential privacy protection, and then uploading the noisy parameters to a central server; the central server aggregating the parameters uploaded by each edge node to generate global model parameters, and distributing them to each edge node to update the local model; performing multiple rounds of iterative training, and adopting a dynamic privacy budget allocation strategy during the iteration process; and deploying the global model that has reached convergence to a smart education application scenario.
[0007] Furthermore, the gradient importance is evaluated in any of the following ways: sensitivity evaluation method: quantifying its importance based on the change in model accuracy before and after adding noise to a specific feature; variance evaluation method: calculating feature importance based on the variance information of the model weights.
[0008] Furthermore, the noise added dynamically to different parameters follows a Laplace distribution, and its standard deviation σ is calculated as follows: σ = α / (1 + β × importance), where α and β are hyperparameters, and importance is the calculated gradient importance value.
[0009] Furthermore, the dynamic privacy budget allocation strategy is specifically as follows: a higher privacy budget is allocated in the early stages of training, and a lower privacy budget is allocated in the later stages of training, with the entire training process satisfying the Renyi differential privacy criterion of ε≤2.
[0010] Furthermore, the feature extraction includes at least one of learning duration features, classroom participation features, and performance trend features.
[0011] Furthermore, the central server uses a weighted federated average algorithm to aggregate parameters, and the aggregation weight is proportional to the amount of data corresponding to each edge node.
[0012] Furthermore, the method employs a federated multi-task learning framework, where the central server maintains shared parameters and task-specific parameters, and dynamically adjusts the aggregation weights of each task based on its data volume or performance.
[0013] Furthermore, the global model trained by the method achieves an accuracy of no less than 89% in predicting student learning outcomes.
[0014] Furthermore, the smart education application scenarios include personalized learning recommendations, learning effect prediction and early warning, and multi-task collaborative analysis in smart classrooms.
[0015] On the other hand, the present invention also provides a federated learning and differential privacy fusion data security processing system for smart education, including multiple edge nodes configured in various educational institutions for performing local data processing, model training and adaptive privacy processing; a central server for aggregating parameters from each edge node, generating and distributing global model parameters, and coordinating multiple rounds of iteration; and an application service module for deploying the trained global model to smart education scenarios and providing services.
[0016] This invention provides a data security processing method based on federated learning and differential privacy fusion for smart education, which has the following beneficial effects: 1. By introducing an adaptive differential privacy mechanism based on gradient importance, smaller noise is added to important features and larger noise is added to minor features, which improves the model's prediction accuracy while providing strict data privacy guarantees.
[0017] 2. By leveraging the distributed nature of federated learning, centralized data transmission and storage are avoided. Combined with optimized communication and aggregation strategies, cross-institutional collaboration costs are reduced, model update cycles are shortened, and the efficiency of data processing and model iteration is improved.
[0018] 3. By designing a federated multi-task learning framework, it can simultaneously support multiple analysis tasks in smart classrooms, such as pose estimation, facial expression recognition, and health analysis. Through fairness adjustments by the central server, it ensures balanced performance of each task, effectively meeting the diverse needs of complex smart education scenarios. Attached Figure Description
[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 The flowchart illustrates the data security processing method based on federated learning and differential privacy fusion for smart education provided by this invention. Figure 2 This is a block diagram of a data security processing system based on federated learning and differential privacy fusion for smart education, provided by the present invention. Figure 3 This invention provides an architecture diagram of a data security processing system based on federated learning and differential privacy fusion for smart education. Figure 4 A flowchart of a data security processing algorithm based on federated learning and differential privacy fusion for smart education provided by this invention; Figure 5 The diagram illustrates the system architecture for a data security processing system based on federated learning and differential privacy fusion for smart education, as provided in this invention. Figure 6 The evaluation index diagram is provided by the present invention for the data security processing method based on federated learning and differential privacy fusion for smart education. Detailed Implementation
[0020] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0021] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0022] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0024] This embodiment provides a data security processing method based on federated learning and differential privacy fusion for smart education, including the following steps: preprocessing and feature extraction of local educational data through edge nodes of each educational institution, and training a local model based on the processed data to generate local model parameters; dynamically adding noise to different parameters based on the gradient importance of the local model parameters to achieve differential privacy protection, and then uploading the noisy parameters to the central server; the central server aggregates the parameters uploaded by each edge node to generate global model parameters, and distributes them to each edge node to update the local model; and performs multiple rounds of iterative training, adopting a dynamic privacy budget allocation strategy during the iteration process; and deploying the global model that has reached convergence to the smart education application scenario.
[0025] See Figure 1 Using cross-regional secondary school learning outcome prediction as a scenario, a federated learning network was built in collaboration with five secondary schools. First, each school deployed student tablets, teacher terminals, and smart classroom devices to construct the terminal device layer; at the edge, each school deployed one school server to complete the edge node layer; the central server layer was deployed in the regional education bureau's data center, using two dual-socket servers; the data storage layer adopted a distributed file system, with each school server storing the original data of local students, and the central server storing global model parameters. The data backup strategy consisted of three local copies and one off-site copy, completing the hardware and software setup.
[0026] During data preparation, full educational data for 3,000 students from 5 schools throughout the semester was collected. The data types and fields included basic information: student IDs were anonymized and replaced with a combination of school code and a random 6-digit number; learning behavior data included daily study time, number of classroom interactions (number of times hands were raised / answered per class), timely homework submission rate (number of timely submissions / total number of assignments), and course video viewing progress (watched time / total viewing time); academic performance data included scores from 3 monthly tests, homework scores (score / full marks for each assignment), and scores from classroom quizzes; behavioral data included student posture compliance rate captured by smart classroom cameras and classroom focus (focus time determined by facial expression recognition / total class time).
[0027] refer to Figure 4 The gradient importance is evaluated using one of the following methods: sensitivity evaluation method: quantifying the importance based on the change in model accuracy before and after adding noise to a specific feature; variance evaluation method: calculating feature importance based on the variance information of the model weights. Dynamically adding noise to different parameters follows a Laplace distribution, and its standard deviation σ is calculated as: σ = α / (1 + β × importance), where α and β are hyperparameters, and importance is the calculated gradient importance value. The dynamic privacy budget allocation strategy specifically involves allocating a higher privacy budget in the early stages of training and a lower privacy budget in the later stages, with the entire training process satisfying the Renyi differential privacy criterion of ε ≤ 2.
[0028] Therefore, each school's edge server evaluates the importance of the gradient of the local model parameters and dynamically adds Laplace noise. The specific process is as follows: Sensitivity-based method: Fixed noise (σ=0.5) is added to each of the 12-dimensional features, and the difference in accuracy of the model on the validation set before and after adding noise (acc₊-acc_ref) is calculated. The results show that the accuracy difference of the trend feature of the scores of the last 3 tests is 5%, the timely submission rate of homework is 3%, the classroom focus is 2.5%, the average daily study time is 2%, the weekly study time fluctuation is 1%, and the remaining features are all <1%. Therefore, the trend of the scores of the last 3 tests, the timely submission rate of homework, and the classroom focus are high-importance features, the average daily study time is a medium-importance feature, and the weekly study time fluctuation and other features are low-importance features.
[0029] The variance-based method calculates the importance of the 12-dimensional features using the formula FI_i=Σ1^64 Var(w_ij)×|w_ij| (q=64, the number of hidden layer neurons; w_ij is the weight of feature i at hidden layer neuron j). The results show that the FI_i value for the trend of the scores of the last three tests is 0.8, the timely submission rate of homework is 0.6, the classroom focus is 0.5, the average daily study time is 0.3, the weekly study time fluctuation is 0.2, and the remaining features are all <0.2, which is consistent with the evaluation results of the sensitivity-based method, further confirming the importance level of the features.
[0030] The hyperparameters α=0.5 and β=1 are set, and the noise standard deviation of each feature gradient is calculated according to the formula σ=α / (1+β×importance_ij), where importance_ij is assigned values of "high importance=1.0, medium importance=0.6, and low importance=0.2". The calculation results are as follows: High importance features (trend of scores in the last 3 tests, timely submission rate of assignments, classroom focus): σ=0.5 / (1+1×1.0)=0.278; Medium importance features (average daily study time): σ=0.5 / (1+1×0.6)=0.3125; Low importance features (weekly study time fluctuation and other 7-dimensional features): σ=0.5 / (1+1×0.2)=0.417; Then, noise following the Laplace(0,σ) distribution is generated and superimposed on the gradient of the corresponding feature to complete the privacy protection processing of the local model parameters.
[0031] refer to Figure 5 The central server employs a weighted federated average algorithm for parameter aggregation, with the aggregation weight proportional to the amount of data corresponding to each edge node. The method utilizes a federated multi-task learning framework, where the central server maintains shared parameters and task-specific parameters, and dynamically adjusts the aggregation weight based on the data volume or performance of each task.
[0032] Therefore, it can be seen that each school's edge server uploads the noisy local model parameters to the central server through an encrypted transmission channel. During the upload process, "fragmented transmission, checksum verification" is used to avoid parameter loss or tampering. Each fragment is 1MB in size, and the checksum algorithm is SHA-256. The central server uses the weighted FedAvg algorithm to aggregate the parameters. The aggregation formula is w_global=(n_A w_A+n_B w_B+n_C w_C+n_D w_D+n_E w_E) / (n_A+n_B+n_C+n_D+n_E), where n_A=n_B=n_C=n_D=n_E=600 (number of students per school), and w_A to w_E are the noisy local model parameters uploaded by the five schools, respectively.
[0033] The central server distributes the aggregated global model parameters to the edge servers of five schools through an encrypted transmission channel. Each school's edge server replaces the original parameters of its local model with the global parameters to complete the local model update. During the update process, the old parameters are automatically backed up. If the accuracy of the model validation set drops by more than 5% after the update, it is rolled back to the old parameters.
[0034] The process of local model training and privacy protection and parameter interaction is repeated. During the iteration, the Renyi differential privacy dynamic budget allocation strategy is adopted, as follows: In the early stage of iteration 1-10: a high privacy budget is allocated, and ε=1.0 is set to ensure the privacy and security of the original student data in the initial training stage; In the middle stage of iteration 11-20: a medium privacy budget is allocated, and ε=1.5 is set, and the noise is restored to the σ value calculated in the above steps to balance privacy protection and model accuracy; In the late stage of iteration 21-30: a low privacy budget is allocated, and ε=2.0 is set, corresponding to less noise, and the focus is on improving model accuracy; The iteration termination condition is: when the iteration reaches 30 rounds, the average loss value of the model on the validation set of 5 schools decreases from the initial 0.6 to 0.12, and the loss value fluctuation is <0.001 for 3 consecutive rounds, which meets the convergence condition and stops the iteration.
[0035] The final global model, which met the evaluation criteria, was deployed to the smart education systems of five schools to achieve learning outcome prediction and early warning functions. The specific deployment process is as follows: Real-time data collection: Students upload daily learning time and homework submission status on their tablets in real time; smart classroom cameras collect real-time data on classroom focus and posture; teachers upload test scores on their terminals; the data is encrypted at the terminal device layer and then transmitted to the school's edge server; Real-time preprocessing and prediction: The edge server standardizes and extracts features from the real-time data, inputs it into the global model to obtain the "pass / fail" prediction result and confidence level; Early warning and intervention: When the model predicts "fail" and the confidence level is ≥80%, the system automatically generates an early warning message containing the student's student ID and weak characteristics such as "homework submission time rate is only 40%" and "the trend of declining test scores in the last 3 tests is 15%", and pushes it to the corresponding homeroom teacher and subject teacher; At the same time, personalized intervention plans are generated based on the weak characteristics, such as pushing homework reminder services and remedial videos for weak knowledge points to student tablets, and teachers conduct targeted tutoring based on the early warning information.
[0036] This application also provides a data security processing system that integrates federated learning and differential privacy for smart education: See Figure 2 It is understood that the system includes multiple edge nodes configured in various educational institutions to perform local data processing, model training, and adaptive privacy processing; a central server to aggregate parameters from each edge node, generate and distribute global model parameters, and coordinate multiple rounds of iteration; and an application service module to deploy the trained global model to smart education scenarios and provide services.
[0037] Therefore, the system adopts a clear hierarchical distributed architecture, with its core consisting of multiple edge nodes, a central server, and an application service module. In terms of hardware configuration, the edge nodes are flexibly deployed according to scenario requirements. For example, in cross-school collaboration, each school's local high-performance servers are used, while within the smart classroom, an edge computing gateway based on embedded chips is used. These nodes are equipped with a dedicated data preprocessing engine, a local model trainer, and a core adaptive privacy processing unit, enabling them to complete feature extraction, model training, and dynamically inject noise based on gradient importance at the data source.
[0038] The central server, serving as the system's coordination hub, is typically deployed in regional data centers or the cloud. It is not merely a simple parameter relay station, but rather integrates a series of intelligent modules, including federated scheduling management, secure aggregation, multi-task fairness control, and privacy budget auditing. For example, its secure aggregator employs a weighted federated averaging algorithm to ensure that organizations with large datasets have greater influence in the global model. Simultaneously, the privacy budget auditor acts like an actuary, allocating a higher privacy budget in the early stages of training to enhance protection, and then tightening the budget later to improve model accuracy, ensuring that total privacy consumption remains within acceptable limits throughout the process.
[0039] The application service module acts as the interface for transforming technological achievements into educational value. It encapsulates the trained and evaluated global model into an easily accessible API service and integrates it into specific educational application platforms. Whether it's pushing academic warning reports to teachers or recommending personalized learning resources to students, all are implemented through the business logic processor of this module.
[0040] See Figure 3 and 6 The system's workflow, exemplified by a five-school collaborative academic prediction project, begins with the central server initiating a task and initializing the model. Simultaneously, edge nodes at each school perform local data processing, training, and privacy protection in parallel. Subsequently, the noise-protected parameters are uploaded to the center for secure aggregation, generating a new generation of global models. After multiple iterations, once the model reaches convergence, the application service module quickly deploys it online, providing real-time student learning information services to frontline teachers. This entire process, under strict privacy protection, significantly improves the efficiency and value of collaborative educational data analysis.
[0041] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for federated learning and differential privacy fusion data security processing based on smart education, characterized in that, The method comprises the following steps: Preprocessing and feature extraction of local education data by edge nodes of each education institution, training of a local model based on the processed data, and generation of local model parameters; Based on the gradient importance of the local model parameters, noise is dynamically added to different parameters to achieve differential privacy protection, and then the noisy parameters are uploaded to the central server; The central server aggregates the parameters uploaded by each edge node to generate global model parameters, and distributes them to each edge node to update the local model; and performs multiple rounds of iterative training, and adopts a dynamic privacy budget allocation strategy during the iteration process; The global model that meets the convergence condition is deployed to the smart education application scenario.
2. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 1, characterized in that, The gradient importance is evaluated by any one of the following methods: Sensitivity evaluation method: the importance of a feature is quantified based on the change in model accuracy before and after adding noise to the feature; Variance evaluation method: the importance of a feature is calculated based on the variance information of the model weight.
3. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 2, characterized in that, The noise added to different parameters dynamically obeys a Laplace distribution, and the standard deviation σ is calculated according to the formula: σ = α / (1+β×importance), where α and β are hyperparameters, and importance is the calculated gradient importance value.
4. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 1, characterized in that, The dynamic privacy budget allocation strategy is to allocate a higher privacy budget in the early training stage and a lower privacy budget in the later training stage, and the entire training process satisfies the Renyi differential privacy standard of ε≤2.
5. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 1, characterized in that, The feature extraction includes at least one of learning duration features, classroom participation features, and score trend features.
6. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 1, characterized in that, The central server aggregates parameters using a weighted federated averaging algorithm, and the aggregation weight is proportional to the data volume corresponding to each edge node.
7. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 1, characterized in that, The method uses a federated multi-task learning framework, and the central server maintains shared parameters and task-specific parameters, and dynamically adjusts the aggregation weight of each task according to the data volume or performance of the task.
8. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 1, characterized in that, A lower noise standard deviation is allocated to gradients with high importance, and the privacy budget is reduced in the later training stage.
9. The smart education-oriented federated learning and differential privacy fusion data security processing method according to claim 1, characterized in that, The smart education application scenario includes personalized learning recommendation, learning effect prediction and early warning, and multi-task collaborative analysis in a smart classroom.
10. A smart education-oriented federated learning and differential privacy fusion data security processing system for implementing the method of any one of claims 1-9. The method comprises the following steps: A plurality of edge nodes are configured in each education institution to perform local data processing, model training, and adaptive privacy processing; A central server is used to aggregate parameters from each edge node, generate and distribute global model parameters, and coordinate multiple rounds of iteration; An application service module is used to deploy the trained global model to a smart education scenario and provide services.