Sensitive data abnormal access detection method based on federal learning

Abstract

The invention discloses a sensitive data abnormal access detection method based on federal learning. The method comprises the steps that distributed training is achieved through federal learning, model parameters are collected from local devices of all participants and aggregated on the premise that user privacy is protected, and therefore a global text classification model is constructed, and data privacy and safety are guaranteed; a TextCNN model is constructed and is used for training a text classification model on local equipment, so that local features in a text can be effectively captured, the accuracy and generalization ability of the model on local data are improved, and whether abnormal access is performed on text data such as user input and program logs is judged; compared with a static management method that resources can be continuously accessed by obtaining tokens, the method has the advantages that abnormal access detection can be dynamically performed according to actual conditions through text recognition, and the data size required by model training can be expanded and the generalization ability of the model can be improved under the condition of protecting data security by utilizing federal learning.

Description

Technical Field

[0001] The present invention relates to the technical field of data security management and control, and specifically to a method for detecting abnormal access to sensitive data based on federated learning. Background Art

[0002] With the development of current computer technology, a large amount of data in various industries is stored and processed digitally. This makes data easier to access, copy and transmit, but also increases the risks faced by data. Data leakage may cause serious damage to the privacy, security and property of individuals or organizations. In data management, abnormal access is a key step in controlling data leakage;

[0003] Currently, the following methods are mainly used for data abnormal access detection:

[0004] Rule-based methods: rely on pre-defined rule sets to detect abnormal access behaviors; for example, define some rules on access frequency, time interval, geographic location, etc. When an access behavior violates these rules, it will be marked as abnormal. For example, based on behavioral analysis methods, by modeling and analyzing user behavior characteristics and defining user behavior rules, it can detect a user's frequent access to sensitive data in a short period of time, or detect unauthorized users accessing sensitive data; there are also statistical rule-based methods that use the statistical characteristics of data, including mean, variance, distribution, etc., to formulate pattern rules that are inconsistent with normal access behavior to detect abnormal behavior.

[0005] Machine learning-based methods: Machine learning technology can identify and predict unknown abnormal access by learning from historical data. Common machine learning algorithms include support vector machines (SVM), decision trees, random forests, etc. These algorithms can learn patterns from large amounts of data and automatically adapt to new data situations.

[0006] Identity and event management-based approach: Control access to sensitive data through key and token management systems. Only users with valid keys or tokens can access sensitive data. Regularly audit and monitor system logs and access logs to sensitive data to promptly identify and handle abnormal access behaviors. Auditing and log monitoring can help identify potential security risks and threats and protect the security of sensitive data.

[0007] Patent No. CN107426199B discloses a method and system for detecting and analyzing abnormal network behavior. The invention is based on a machine learning algorithm, and constructs a support vector machine model based on the user's access behavior feature data to predict and analyze the behavior to identify abnormal access behavior.

[0008] Patent No. CN106570131A discloses a method for detecting abnormal access to sensitive data based on cluster analysis. By establishing business rules for illegal scenarios and using cluster analysis algorithms to mine possible illegal accesses, new business rules for illegal scenarios are formed. This solves the problem of being unable to detect new illegal access behaviors due to slow rule updates.

[0009] Patent No. CN110213077B discloses a method, device and system for determining security events in a power monitoring system. By correlating alarm records to build an attack tree, performing aggregation processing and pruning noise reduction, the final attack chain set is obtained to determine the security events in the power monitoring system.

[0010] Patent No. CN109525558B discloses a data leakage detection method, system, device and storage medium. According to tracking and recording the access traffic behavior of each session, a historical traffic log is generated, matched with a preset abnormal feature set, and the category to which the abnormal feature belongs is obtained, and the corresponding strategy is determined by the category.

[0011] The above method has certain limitations:

[0012] Rule-based methods may be limited by the rigidity of the rules themselves and cannot adapt to complex and changing access patterns. If the rule design is not flexible or comprehensive enough, it may lead to omissions or false positives. Behavioral analysis methods require in-depth analysis and understanding of user behavior patterns and characteristics, which requires a large amount of data and complex algorithm models. Moreover, behavioral patterns may change over time, requiring continuous updating and adjustment of analysis models. Auditing and log monitoring require a lot of manpower and resource investment, and the processing and analysis of data requires a certain amount of time and technical support. If the monitoring system is not timely and efficient enough, abnormal access behavior may be ignored or delayed. When using machine learning methods, if the training samples are insufficient or not representative enough, the model's generalization ability may be insufficient and abnormal access behavior cannot be accurately detected.

[0013] The quantity and quality of data determine the accuracy of detection. However, relevant data are usually distributed among different regions or organizations, and one party only has partial dimensional data. The above method needs to share data among different institutions, which usually requires data from different institutions to be transferred to the central server for centralized processing, which may lead to the risk of sensitive information leakage. Once the central server is attacked or fails, all data may be exposed or damaged. Summary of the invention

[0014] In order to solve the above problems, the present invention provides a method for detecting abnormal access to sensitive data based on federated learning, which combines federated learning with the TextCNN model. It allows the model of text data to be trained through local devices while protecting user privacy, and aggregates the parameters to a central server, thereby realizing real-time monitoring and identification of abnormal access to sensitive data and ensuring data security.

[0015] The method comprises the following steps:

[0016] S1. Data preparation: Collect text data of participants, including normal access and abnormal access data.

[0017] S1.1. Label the text data and mark normal access and abnormal access into different categories.

[0018] S2. Model initialization: Initialize the TextCNN (text convolutional neural network) model locally on each participant.

[0019] S3. Local training: Each participant uses local text data to train the TextCNN model, ensuring that sensitive data does not leave the local environment of the participant and protecting the privacy and security of the data.

[0020] S4, Model upload: After each participant completes the training, the updated model parameters are uploaded to the central server. Before uploading the model parameters, the model parameters are encrypted or otherwise processed to ensure secure transmission of the model parameters.

[0021] S5. Parameter aggregation: After receiving the model parameters from each participant, the central server aggregates the parameters. The global model is updated using methods such as federated averaging. The updated global model will contain the local data patterns and features of each participant, as well as the corresponding abnormal access detection capabilities.

[0022] S6. Model evaluation: Evaluate the performance of the updated global model on a central server or locally at each participant.

[0023] S7. Deployment and monitoring: Deploy the updated global model to the actual system and set up a monitoring mechanism to detect text access behavior in real time. The monitoring system will classify text access behavior based on the output of the global model to identify abnormal access behavior.

[0024] S8. Feedback and update: Based on the feedback information from actual applications, the global model is regularly updated and optimized to improve the recognition accuracy and efficiency of the model.

[0025] Furthermore, the participants include different organizations or users.

[0026] Furthermore, the TextCNN includes components such as a convolutional layer, a pooling layer, and a fully connected layer.

[0027] Furthermore, the evaluation includes indicators such as accuracy, recall, and F1 value.

[0028] Beneficial effects of the present invention:

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] Compared with existing rule-based, statistical, and behavioral analysis methods, the TextCNN model does not need to rely on pre-defined rules or statistical models; it can automatically adapt to different types of text data and different abnormal access behaviors by learning patterns and features in text data without manual adjustment.

[0031] Compared with the token management method, using the TextCNN model to recognize text does not require users to perform additional identity authentication. It only needs to monitor and analyze the user's text access behavior, which can reduce the user's operating burden and eliminate the need to obtain and manage tokens, thereby improving the user experience. It has a wide range of applications. TextCNN text recognition can be applied to various types of text data, including text entered by users, log text generated by applications, etc. Token management is usually only applicable to scenarios where user authentication and authorization are required. Dynamic detection: TextCNN text recognition can dynamically detect anomalies based on actual conditions, while token management is usually static. Once a user obtains a valid token, they can continue to access the corresponding resources.

[0032] Compared with machine learning methods, TextCNN can automatically learn features and patterns from text data through end-to-end learning, without artificially defining rules or manually extracting features. This makes the design and deployment of the model simpler and more direct. TextCNN can effectively capture local features and semantic information in text, and has strong adaptability and generalization capabilities. Even if the pattern of text data changes or new anomalies appear, the model can make corresponding adjustments and adaptations.

[0033] The method of using TextCNN based on federated learning to recognize text to detect abnormal access to sensitive data has the advantages of simple training, strong adaptability, high degree of automation, data expansion, and high data privacy security, and can play an important role in practical applications.

[0034] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. DETAILED DESCRIPTION

[0035] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] Example

[0037] Collect text data from participants;

[0038] Build the TextCNN model:

[0039] Define the input layer:

[0040] First, we construct the input layer to receive text data. The function of the input layer is to specify the shape of the input data, that is, the length of the text sequence.

[0041] The embedding layer converts word vectors:

[0042] Create an embedding layer. The purpose of the embedding layer is to convert each word in the input text sequence into a dense vector representation. These vectors represent the semantic information of each word in a high-dimensional space. This step turns each word into a meaningful number.

[0043] Convolutional layer extracts features:

[0044] Use convolution operations to extract features from text. The convolution layer applies multiple convolution kernels of different sizes to the input text to capture features of different sizes. These convolution kernels slide over the text to calculate the feature representation of each position. The convolution operation is a one-dimensional convolution because text data is a one-dimensional sequence, including a sequence of words or a sequence of characters. Each convolution kernel produces a feature map whose size depends on the length of the input text and the size of the convolution kernel. The feature map captures local feature information of different sizes. The output of the convolution layer is activated by a nonlinear activation function (such as ReLU) and then passed to the pooling layer for downsampling to reduce the dimension of the feature and extract the most significant feature information.

[0045] The pooling layer extracts salient features:

[0046] The pooling layer performs a maximum pooling operation on each channel of the feature map generated by each convolution kernel, that is, selects the maximum value in each feature map as the pooling result, which helps to extract the most significant features in each feature map and streamline the existing features, thereby effectively capturing the key information in the text.

[0047] Combine the outputs of multiple pooling layers:

[0048] The output of each pooling layer is connected to form a larger feature vector, which retains feature information of different sizes and fuses feature information of different dimensions, thus enhancing the expressiveness of the model.

[0049] Fully connected layer for classification:

[0050] The feature vectors extracted by the convolutional layer and the pooling layer are input into the fully connected layer. By learning the weight matrix, the fully connected layer maps the input features to the output label space. The weight matrix is ​​updated according to the loss function during the training process to minimize the error between the predicted value and the true label. The output of the fully connected layer passes through the softmax activation function to generate the probability distribution of each category.

[0051] Federated Learning Framework:

[0052] Client device selection: Determine the client devices participating in federated learning, which can be mobile devices of different participants, edge devices, or servers distributed in different geographical locations.

[0053] Local model training: Each client device uses its own data to train the TextCNN model locally. During the training process, the client device only updates the parameters of the model without sharing the original data.

[0054] Sending gradients: Each participant calculates the model gradient locally, uses encryption technologies such as homomorphic encryption and differential privacy to encrypt the gradient information, and sends the encrypted gradient to the aggregation server.

[0055] Model aggregation: The centralized server periodically aggregates the model parameters on each client device. This process can use simple parameter averaging, weighted averaging, or other aggregation strategies.

[0056] Model evaluation and update: The centralized server deploys the aggregated model parameters to each client device, and then the client device uses this model to evaluate the local data. Based on the evaluation results, it determines whether the model architecture or hyperparameters need to be further adjusted, and the updated model is distributed to the client device again.

[0057] Multiple participants with similar data structures jointly train a TextCNN model under the collaboration of the aggregation server. During the entire process, the client device only shares the model parameters but not the original data, thus ensuring the privacy of user data, expanding the data samples, and effectively utilizing distributed resources, so that the model can better learn sensitive data features and improve generalization capabilities.

[0058] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0059] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for detecting abnormal access to sensitive data based on federated learning, characterized in that: The following steps are involved: S1. Data preparation: Collect text data from participants; S1.1, marking the text data into two categories: normal access and abnormal access; S2. Model initialization: Initialize the TextCNN model locally at each participant; S3, local training: each of the participants uses the local text data in S1 to train the TextCNN model of S2; S4, model upload: after each participant completes the training, the model parameters of the updated TextCNN model are uploaded to the central server; S5. Parameter aggregation: The central server aggregates the model parameters of each participant using a federal averaging method to obtain an updated global model of the central server.

2. The method for detecting abnormal access to sensitive data based on federated learning according to claim 1, characterized in that: The following steps are also included: S6. Model evaluation: Evaluate the performance of the updated global model on the central server and locally at each participant. S7, deployment and monitoring: deploy the updated global model to the actual system and set up a monitoring mechanism to detect text access behavior in real time to form a monitoring system; S8. Feedback and update: Based on the feedback information from actual applications, the global model is regularly updated and optimized to improve the recognition accuracy and efficiency of the model.

3. The method for detecting abnormal access to sensitive data based on federated learning according to claim 1, characterized in that: The S2 step also includes S2.1, and the TextCNN model includes convolutional layer, pooling layer and fully connected layer components.

4. The method for detecting abnormal access to sensitive data based on federated learning according to claim 1, characterized in that: The step S4 also includes S4.1, performing security processing on the model parameters before uploading the model parameters.

5. The method for detecting abnormal access to sensitive data based on federated learning according to claim 4, characterized in that: The security processing is encryption.

6. The method for detecting abnormal access to sensitive data based on federated learning according to claim 2, characterized in that: The evaluation includes indicators of accuracy, recall, and F1 value.

7. The method for detecting abnormal access to sensitive data based on federated learning according to claim 2, characterized in that: The monitoring system classifies text access behaviors according to the output results of the global model and identifies abnormal access behaviors.

8. The method for detecting abnormal access to sensitive data based on federated learning according to claim 1, characterized in that: The text data includes but is not limited to text input by the user and log text generated by the application.