Privacy leak detection method and device based on mixed learning

Abstract

The invention provides a privacy leak detection method and device based on mixed learning. In one embodiment of the invention, the privacy leak detection method based on hybrid learning comprises the following steps: acquiring privacy leak detection associated data; performing privacy leakage detection associated feature extraction on the privacy leakage detection associated data by using an unsupervised learning model to obtain privacy leakage detection associated features; determining an initial optimal hyper-parameter combination of the deep learning model by using a hyper-parameter optimization scheme; constructing a privacy leakage detection model, and training the privacy leakage detection model by using the privacy leakage detection association features according to the initial optimal hyper-parameter combination to obtain a trained privacy leakage detection model; and carrying out transfer learning on the trained privacy leakage detection model for a to-be-detected Internet of Things environment. According to the method, the generalization ability and accuracy of the model on complex Internet of Things data can be improved.

Description

Technical Field

[0001] The present application relates to the field of information security technology, and in particular to a privacy leakage detection method and device based on hybrid learning. Background Art

[0002] With the rapid development of the Internet of Things (IoT), billions of devices are connected to each other through the Internet, generating and sharing a large amount of data. These devices include smart home devices, industrial control systems, etc. The data they collect and process often involves sensitive information, such as personal identity information, health data, financial information, etc.

[0003] However, the diversity, distribution, and interconnectedness of IoT devices make privacy data leakage an increasingly prominent problem. The data collected by IoT devices often contains sensitive information, which, once leaked, may pose a serious threat to personal privacy.

[0004] IoT data is characterized by massiveness, diversity, and complexity. How to automatically identify and classify privacy data containing sensitive information in the large amount of data generated by IoT devices has become a current research hotspot. Summary of the invention

[0005] In view of this, the present application provides a privacy leakage detection method and device based on hybrid learning.

[0006] Specifically, the present application is implemented through the following technical solutions: According to a first aspect of an embodiment of the present application, a privacy leakage detection method based on hybrid learning is provided, comprising: Obtaining privacy leakage detection associated data; wherein the privacy leakage detection associated data is obtained by processing data generated by an IoT device; Using an unsupervised learning model, extracting privacy leakage detection associated features from the privacy leakage detection associated data to obtain privacy leakage detection associated features; wherein the unsupervised learning model includes an independent autoencoder and a stacked autoencoder; the privacy leakage detection associated features are obtained by fusing the first privacy leakage detection associated features and the second privacy leakage detection associated features; the first privacy leakage detection associated features are extracted using the independent autoencoder, and the second privacy leakage detection associated features are extracted using the stacked autoencoder; and, Use hyperparameter optimization schemes to determine the optimal hyperparameter combination for deep learning models; Constructing a privacy leakage detection model, and training the privacy leakage detection model based on the initial optimal hyperparameter combination and using the privacy leakage detection associated features to obtain a trained privacy leakage detection model; wherein the trained privacy leakage detection model is used to perform privacy leakage detection on IoT devices; For the Internet of Things environment to be detected, transfer learning is performed on the trained privacy leakage detection model to obtain a privacy leakage detection model to be deployed for privacy leakage detection of the Internet of Things environment to be detected; wherein the privacy leakage detection model to be deployed is used to perform privacy leakage detection on the Internet of Things devices in the Internet of Things environment to be detected.

[0007] According to a second aspect of an embodiment of the present application, an electronic device is provided, including a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor is used to execute the machine executable instructions to implement the method provided in the first aspect.

[0008] The privacy leakage detection method based on hybrid learning in the embodiment of the present application obtains privacy leakage detection related data, uses an unsupervised learning model to extract privacy leakage detection related features from the privacy leakage detection related data, and obtains privacy leakage detection related features; and uses a hyperparameter optimization scheme to determine the optimal hyperparameter combination of the deep learning model; then, a privacy leakage detection model is constructed, and based on the initial optimal hyperparameter combination, the privacy leakage detection model is trained using the privacy leakage detection related features to obtain a trained privacy leakage detection model, and transfer learning is performed on the trained privacy leakage detection model for the Internet of Things environment to be detected to obtain a to-be-deployed privacy leakage detection model for privacy leakage detection in the Internet of Things environment to be detected, and the obtained privacy leakage detection related features are optimized by using an unsupervised learning method. The unsupervised feature extraction method can be used to extract features from the data without relying on manually labeled data. In addition, by combining unsupervised feature extraction with deep learning networks, the generalization ability and accuracy of the model on complex IoT data can be effectively improved, thereby improving the privacy leakage detection performance of the privacy leakage detection model for IoT devices. Furthermore, by jointly extracting privacy leakage associated features through independent autoencoders and stacked autoencoders, the quality of privacy leakage detection associated features is improved, and then the performance of the trained privacy leakage detection model can be optimized. Finally, in the process of deploying the trained privacy leakage detection model, the trained privacy leakage detection model is transferred for the IoT environment to be detected, so that the privacy leakage detection model is more suitable for the IoT environment to be detected, thereby further improving the privacy leakage detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1A flowchart of a privacy leakage detection method based on hybrid learning is shown as an exemplary embodiment of the present application; Figure 2 A schematic diagram of an implementation flow of a privacy leakage detection solution based on hybrid learning is shown as an exemplary embodiment of the present application; Figure 3 A schematic diagram of a privacy leakage detection method and device based on hybrid learning is shown as an exemplary embodiment of the present application; Figure 4 The figure is a schematic diagram of the hardware structure of an electronic device shown as an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0010] In order to enable those skilled in the art to better understand the technical solutions provided by the embodiments of the present application and to make the above-mentioned purposes, features and advantages of the embodiments of the present application more obvious and understandable, the technical solutions in the embodiments of the present application are further described in detail below in conjunction with the accompanying drawings.

[0011] See also Figure 1 , which is a flow chart of a privacy leakage detection method based on hybrid learning provided in an embodiment of the present application, such as Figure 1 As shown, the privacy leakage detection method based on hybrid learning may include: Step S100: Obtain privacy leakage detection associated data; wherein the privacy leakage detection associated data is obtained by processing data generated by an IoT device.

[0012] Exemplarily, the privacy leakage detection-related data includes data used to perform privacy leakage detection on IoT devices.

[0013] In an example, the obtaining of privacy leakage detection related data may include: Obtain data generated by IoT devices; Clean the acquired data and select features related to privacy leakage detection from the cleaned data; The features related to privacy leakage detection are converted into numerical form and normalized to obtain privacy leakage detection related data.

[0014] For example, the data generated by IoT devices may include, but are not limited to, part or all of the data such as device logs, network traffic, and user behavior data.

[0015] For the data generated by the IoT devices, data cleaning can be used to remove invalid, erroneous or incomplete data from the acquired data, and features related to privacy leakage detection can be selected from the cleaned data.

[0016] Exemplarily, the features related to privacy leakage detection may include but are not limited to access frequency and / or data transmission volume.

[0017] Considering that an unsupervised learning model will be used to extract features from privacy leakage detection related data in the future, and the input of the network model usually needs to be numerical data in the range of [0, 1], the features related to privacy leakage detection can also be converted into numerical form and normalized to obtain privacy leakage detection related data.

[0018] For example, the features related to privacy leakage detection may be converted into a numerical form by performing One-Hot encoding (one valid bit) on them.

[0019] Among them, One-Hot encoding is a method of converting categorical variables into numerical form, which is often used in machine learning and deep learning models. Its main idea is to represent each category as a binary vector, in which only one element is 1 and the rest are 0.

[0020] Step S110: Use an unsupervised learning model to extract privacy leakage detection associated features from privacy leakage detection associated data to obtain privacy leakage detection associated features; wherein the unsupervised learning model includes an independent autoencoder and a stacked autoencoder; the privacy leakage detection associated features are obtained by performing feature fusion on the first privacy leakage detection associated features and the second privacy leakage detection associated features; the first privacy leakage detection associated features are extracted using the independent autoencoder, and the second privacy leakage detection associated features are extracted using the stacked autoencoder.

[0021] In an embodiment of the present application, in order to reduce the dependence of manual training on the privacy leakage detection deep learning model (hereinafter referred to as the privacy leakage detection model), an unsupervised learning method can be used to extract features of the privacy leakage detection related data, and the extracted features can be used for subsequent training of the privacy leakage detection model.

[0022] Accordingly, an unsupervised learning model can be used to extract privacy leakage detection related features from privacy leakage detection related data to obtain privacy leakage detection related features.

[0023] In one example, the unsupervised learning model may include: independent autoencoders, and / or, stacked autoencoders.

[0024] Exemplarily, independent autoencoders may include, but are not limited to, one or more of a deep autoencoder (DAE), a deep LSTM (Long Short-Term Memory) autoencoder (LSTM-DAE), or a deep convolutional autoencoder (DCAE).

[0025] Stacked autoencoders may include multiple different types, and different types of stacked autoencoders may participate in different types of stacked autoencoders, and / or the stacking relationships between the autoencoders participating in the stacking may be different.

[0026] For stacked autoencoders, each autoencoder involved in the stacking extracts features independently and uses the extracted features as the input of the next autoencoder. Through layer-by-layer learning, the dimension of the input data can be gradually reduced and more advanced features can be extracted.

[0027] Exemplarily, through independent autoencoders and / or stacked autoencoders, feature extraction is performed on privacy leakage detection related data, and data dimensionality reduction is achieved, and more representative feature dimensions are automatically extracted, thereby improving the training efficiency in the subsequent privacy leakage detection model training process and optimizing the training effect.

[0028] It should be noted that in the embodiments of the present application, the unsupervised learning model is not limited to independent autoencoders or stacked autoencoders, and other models such as deep belief networks can also be used, and their specific implementation is not limited here.

[0029] Exemplarily, in order to optimize the feature extraction effect, the unsupervised learning model may include an independent autoencoder and a stacked autoencoder, and the independent autoencoders extract corresponding privacy leakage detection key features (which may be called first privacy leakage detection associated features), and the stacked autoencoders are used to extract corresponding privacy leakage detection key features (which may be called second privacy leakage detection associated features), and the privacy leakage detection associated features used for model training are obtained by feature fusion.

[0030] Step S120: Determine the initial optimal hyperparameter combination of the deep learning model using a hyperparameter optimization scheme.

[0031] In an embodiment of the present application, in order to ensure the performance of the privacy leakage detection model, a hyperparameter optimization scheme can be used to determine the optimal hyperparameter combination of the deep learning model (which can be called the initial optimal hyperparameter combination).

[0032] Exemplarily, the hyperparameters may include, but are not limited to, one or more of the hyperparameters such as learning rate, number of hidden layers, number of neurons, etc.

[0033] In one example, determining an initial optimal hyperparameter combination of a deep learning model using a hyperparameter optimization solution may include: Based on the defined hyperparameter search range, the Bayesian optimization algorithm is used to determine the initial optimal hyperparameter combination of the deep learning model.

[0034] Exemplarily, in order to improve the efficiency of determining the optimal hyperparameter combination, a Bayesian optimization algorithm can be used to determine the optimal hyperparameter combination of the deep learning model.

[0035] Exemplarily, in the process of determining the optimal hyperparameter combination, the optimal hyperparameter combination can be automatically searched for using a Bayesian optimization algorithm based on a defined search space, that is, a hyperparameter search range.

[0036] For example, the search space for learning rate, the search space for the number of hidden layers, and the search space for the number of neurons can be pre-defined, and the Bayesian optimization algorithm can be used to automatically search for the optimal hyperparameter combination, i.e., the optimal learning rate, the optimal number of hidden layers, and the optimal number of neurons, based on the search space of each defined hyperparameter.

[0037] Step S130: construct a privacy leakage detection model, and train the privacy leakage detection model based on the initial optimal hyperparameter combination using the privacy leakage detection associated features to obtain a trained privacy leakage detection model.

[0038] In an embodiment of the present application, after obtaining the privacy leakage detection associated features in the manner described in the above embodiments and determining the initial optimal hyperparameter combination, a privacy leakage detection model can be constructed, and based on the initial optimal hyperparameter combination, the privacy leakage detection model can be trained using the privacy leakage detection associated features to obtain a trained privacy leakage detection model.

[0039] Exemplarily, when a trained privacy leakage detection model is obtained, the trained privacy leakage detection model can be used to perform privacy leakage detection on IoT devices.

[0040] For example, the trained privacy leakage detection model is deployed to the IoT platform, which monitors the data reported by IoT devices in real time and uses the trained privacy leakage detection model to perform privacy leakage detection.

[0041] Step S140: For the IoT environment to be detected, transfer learning is performed on the trained privacy leakage detection model to obtain a privacy leakage detection model to be deployed for privacy leakage detection of the IoT environment to be detected; wherein the privacy leakage detection model to be deployed is used to perform privacy leakage detection on IoT devices in the IoT environment to be detected.

[0042] In an embodiment of the present application, in the process of deploying the trained privacy leakage detection model, in order to make the privacy leakage detection model better adapt to the Internet of Things environment to be deployed and optimize the privacy leakage detection effect, the trained privacy leakage detection model can be transferred and learned for the Internet of Things environment to be detected, so that the privacy leakage detection model can learn the relevant characteristics of the Internet of Things environment to be detected, and obtain the privacy leakage detection model to be deployed for privacy leakage detection in the Internet of Things environment to be detected.

[0043] Exemplarily, the privacy leakage detection model to be deployed is used to perform privacy leakage detection on IoT devices in the IoT environment to be detected.

[0044] For example, the privacy leakage detection model to be deployed is deployed to the IoT platform of the IoT environment to be detected. The IoT platform of the IoT environment to be detected monitors the data reported by the IoT devices in the IoT environment to be detected in real time, and uses the deployed privacy leakage detection model to perform privacy leakage detection.

[0045] It can be seen that in Figure 1 In the method flow shown, by obtaining privacy leakage detection associated data, using an unsupervised learning model, privacy leakage detection associated features are extracted from the privacy leakage detection associated data to obtain privacy leakage detection associated features; and a hyperparameter optimization scheme is used to determine the initial optimal hyperparameter combination of the deep learning model; then, a privacy leakage detection model is constructed, and based on the initial optimal hyperparameter combination, the privacy leakage detection model is trained using the privacy leakage detection associated features to obtain a trained privacy leakage detection model, and transfer learning is performed on the trained privacy leakage detection model for the Internet of Things environment to be detected to obtain a to-be-deployed privacy leakage detection model for privacy leakage detection in the Internet of Things environment to be detected, and feature extraction is performed on the acquired privacy leakage detection associated data by using an unsupervised learning method. There is no need to rely on manually labeled data; in addition, by combining unsupervised feature extraction with a deep learning network, the generalization ability and accuracy of the model on complex IoT data can be effectively improved, thereby improving the privacy leakage detection performance of the privacy leakage detection model for IoT devices; further, by jointly extracting privacy leakage associated features through independent autoencoders and stacked autoencoders, the quality of privacy leakage detection associated features is improved, and then the performance of the trained privacy leakage detection model can be optimized; finally, in the process of deploying the trained privacy leakage detection model, transfer learning is performed on the trained privacy leakage detection model for the IoT environment to be detected, so that the privacy leakage detection model is more suitable for the IoT environment to be detected, thereby further improving the privacy leakage detection performance.

[0046] In some embodiments, the unsupervised learning model is used to extract privacy leakage detection related features from the privacy leakage detection related data to obtain the privacy leakage detection related features, which may include: Using an independent autoencoder, extracting privacy leakage detection related features from the privacy leakage detection related data to obtain a first privacy leakage detection related feature; and Using stacked autoencoders, extract privacy leakage detection related features from privacy leakage detection related data to obtain second privacy leakage detection related features; The first privacy leakage detection associated feature and the second privacy leakage detection associated feature are feature fused to obtain a final privacy leakage detection associated feature.

[0047] Exemplarily, in order to optimize the feature extraction effect, an independent autoencoder and a stacked autoencoder may be used to jointly perform feature extraction to extract privacy leakage detection related features from privacy leakage detection related data.

[0048] Correspondingly, for the acquired privacy leakage detection associated data, on the one hand, an independent autoencoder can be used to extract privacy leakage detection associated features from the privacy leakage detection associated data to obtain corresponding privacy leakage detection associated features (i.e., the first privacy leakage detection associated features).

[0049] On the other hand, a stacked autoencoder can be used to extract privacy leakage detection associated features from the privacy leakage detection associated data to obtain corresponding privacy leakage detection associated features (i.e., the second privacy leakage detection associated features).

[0050] The obtained first privacy leakage detection associated feature and the second privacy leakage detection associated feature may be feature fused to obtain a final privacy leakage detection associated feature.

[0051] In one example, performing feature fusion on the first privacy leakage detection associated feature and the second privacy leakage detection associated feature to obtain the final privacy leakage detection associated feature may include: According to the maximum dimension of the privacy leakage detection associated feature, the first privacy leakage detection associated feature and the second privacy leakage detection associated feature are aligned by filling in zeros; For any dimension, the non-zero value of the first privacy leakage detection associated feature in this dimension and the non-zero value of the second privacy leakage detection associated feature in this dimension are fused using an average method to obtain the final value of the privacy leakage detection associated feature in this dimension.

[0052] For example, considering that the dimensions of privacy leakage detection associated features obtained by extracting privacy leakage detection associated features from privacy leakage detection associated data using different unsupervised learning models may be different, feature alignment is required in the process of fusing the privacy leakage detection associated features extracted using different unsupervised learning models.

[0053] Exemplarily, the first privacy leakage detection associated feature and the second privacy leakage detection associated feature may be aligned by padding with zeros according to the maximum dimension of the privacy leakage detection associated feature.

[0054] For example, assuming that the maximum dimension of the privacy leakage detection associated feature is 10 dimensions, and the dimension of the first privacy leakage detection associated feature is 8 dimensions (such as the first 8 dimensions), the last two dimensions of the first privacy leakage detection associated feature can be padded with 0; assuming that the dimension of the first privacy leakage detection associated feature is 7 dimensions (such as the last 7 dimensions), the first three dimensions of the second privacy leakage detection associated feature can be padded with 0.

[0055] After completing the feature alignment of the first privacy leakage detection associated feature and the second privacy leakage detection associated feature, for any dimension, the non-zero value of the first privacy leakage detection associated feature in that dimension and the non-zero value of the second privacy leakage detection associated feature in that dimension are fused using an averaging strategy to obtain the final value of the privacy leakage detection associated feature in that dimension.

[0056] Exemplarily, the averaging strategy may include an arithmetic averaging strategy or a weighted averaging strategy.

[0057] Taking the arithmetic average strategy as an example, assuming that for a certain dimension, the value of the first privacy leakage detection associated feature in this dimension is a1 (a1≠0), and the value of the second privacy leakage detection associated feature in this dimension is a2 (a2≠0), then the final value of the privacy leakage detection associated feature in this dimension can be (a1+a2) / 2.

[0058] In one example, for any dimension, a non-zero value of the first privacy leakage detection associated feature in the dimension and a non-zero value of the second privacy leakage detection associated feature in the dimension are fused using an average strategy to obtain a final value of the privacy leakage detection associated feature in the dimension, which may include: For any dimension, if the number of non-zero values ​​in the value of the first privacy leakage detection associated feature of the dimension and the value of the second privacy leakage detection associated feature of the dimension is 1, the non-zero value is determined as the final value of the privacy leakage detection associated feature of the dimension; When there is no non-zero value in the value of the dimension of the first privacy leakage detection associated feature and the value of the dimension of the second privacy leakage detection associated feature, it is determined that the final value of the dimension of the privacy leakage detection associated feature is 0.

[0059] Exemplarily, for a certain dimension, if the number of non-zero values ​​in the value of the first privacy leakage detection associated feature of the dimension and the value of the second privacy leakage detection associated feature of the dimension is 1, the non-zero value can be determined as the final value of the privacy leakage detection associated feature of the dimension.

[0060] For example, assuming that for a certain dimension, the value of the first privacy leakage detection associated feature for this dimension is b1 (b1≠0), and the value of the second privacy leakage detection associated feature for this dimension is 0, then the value of the final privacy leakage detection associated feature for this dimension may be b1.

[0061] Exemplarily, for a certain dimension, if the value of the dimension in the first privacy leakage detection associated feature and the value of the dimension in the second privacy leakage detection associated feature are both 0, then the value of the dimension in the final privacy leakage detection associated feature is also 0.

[0062] For example, assuming that for a certain dimension, the value of the first privacy leakage detection associated feature for this dimension is 0, and the value of the second privacy leakage detection associated feature for this dimension is 0, then the value of the final privacy leakage detection associated feature for this dimension may be 0.

[0063] In one example, the unsupervised learning model includes multiple types of independent autoencoders, and / or multiple types of stacked autoencoders; The first privacy leakage detection associated features include first privacy leakage detection associated features extracted by multiple different independent autoencoders, and / or the second privacy leakage detection associated features include second privacy leakage detection associated features extracted by multiple different stacked autoencoders.

[0064] Exemplarily, in order to further optimize the feature extraction effect and obtain higher quality privacy leakage detection associated features, in the process of extracting privacy leakage detection associated features from privacy leakage detection associated data using independent autoencoders and stacked autoencoders, the types of independent autoencoders involved in feature extraction may include multiple types, and / or the types of stacked autoencoders may include multiple types.

[0065] Accordingly, the first privacy leakage detection associated features may include first privacy leakage detection associated features extracted by multiple different independent autoencoders, and / or the second privacy leakage detection associated features may include second privacy leakage detection associated features extracted by multiple different stacked autoencoders.

[0066] For example, independent autoencoders S1 and S2, as well as stacked autoencoders S3 and S4, can be used to perform privacy leakage detection associated feature extraction on privacy leakage detection associated data to obtain the first privacy leakage detection associated feature 1, the first privacy leakage detection associated feature 2, the second privacy leakage detection associated feature 1, and the second privacy leakage detection associated feature 2, and perform feature fusion on the first privacy leakage detection associated feature 1, the first privacy leakage detection associated feature 2, the second privacy leakage detection associated feature 1, and the second privacy leakage detection associated feature 2.

[0067] Among them, in the process of feature fusion of the first privacy leakage detection associated feature 1, the first privacy leakage detection associated feature 2, the second privacy leakage detection associated feature 1, and the second privacy leakage detection associated feature 2, feature alignment can be performed by filling in 0. For any dimension, the non-zero values ​​of the first privacy leakage detection associated feature 1, the first privacy leakage detection associated feature 2, the second privacy leakage detection associated feature 1, and the second privacy leakage detection associated feature 2 in this dimension can be fused using the averaging strategy.

[0068] For example, assuming that for a certain dimension, the value of the first privacy leakage detection associated feature 1 is c1, the value of the first privacy leakage detection associated feature 2 is c2, the value of the second privacy leakage detection associated feature 1 is 0, and the value of the second privacy leakage detection associated feature 2 is c3, then the final value of the dimension of the privacy leakage detection associated feature can be (c1+ c2+ c3) / 3.

[0069] In some embodiments, the training of the privacy leakage detection model using the privacy leakage detection associated features based on the initial optimal hyperparameter combination may include: Generate training sets and validation sets based on privacy leakage detection related features; Based on the initial optimal hyperparameter combination, the privacy leakage detection model is trained using the training set; Use the validation set to tune the model structure and parameters of the trained privacy leakage detection model; Use the test set to evaluate the performance of the optimized privacy leakage detection model; When the performance evaluation result meets the performance requirements, the current privacy leakage detection model is determined as the trained privacy leakage detection model; When the performance evaluation results do not meet the performance requirements, the current privacy leakage detection model is iteratively trained using the training set.

[0070] Exemplarily, for the privacy leakage detection related features extracted in the above manner, a training set and a validation set can be generated based on them.

[0071] During the training process of the privacy leakage detection model, the privacy leakage detection model can be trained using the generated training set based on the initial optimal hyperparameter combination determined in the above manner, and the model structure and parameters of the trained privacy leakage detection model can be tuned using the generated validation set.

[0072] For a verified privacy leakage detection model, that is, a tuned privacy leakage detection model, the performance of the tuned privacy leakage detection model can be evaluated using a test set.

[0073] Exemplarily, the test set may be an original IoT data set. For example, the test set may be generated by collecting data generated by IoT devices and annotating the data with labels.

[0074] Exemplarily, when the current privacy leakage detection model passes the performance evaluation, that is, when the performance evaluation result meets the performance requirements, the current privacy leakage detection model can be determined as a trained privacy leakage detection model.

[0075] When the performance evaluation result of the current privacy leakage detection model does not meet the performance requirements, the training set can be used to iteratively train the current privacy leakage detection model.

[0076] In some embodiments, the privacy leakage detection method based on hybrid learning provided in the embodiments of the present application may also include: The data generated by the IoT environment to be detected is obtained, and the trained privacy leakage detection model is fine-tuned based on the obtained data; wherein the fine-tuned privacy leakage detection model is used to perform privacy leakage detection on the IoT environment to be detected.

[0077] For example, in order to further improve the detection performance of the privacy leakage detection model, before using the trained privacy leakage detection model for privacy leakage detection, the trained privacy leakage detection model can be fine-tuned using data from the IoT environment to be detected, so that the privacy leakage detection model is more suitable for the IoT environment to be detected.

[0078] Accordingly, the data generated by the IoT environment to be detected can be obtained, and the trained privacy leakage detection model can be fine-tuned based on the obtained data. Then, the fine-tuned privacy leakage detection model can be used to perform privacy leakage detection on the IoT environment to be detected.

[0079] For example, the fine-tuned privacy leakage detection model can be deployed to the Internet of Things platform of the Internet of Things environment to be detected, and the Internet of Things platform uses the deployed privacy leakage detection model to perform privacy leakage detection.

[0080] In order to enable those skilled in the art to better understand the technical solutions provided in the embodiments of the present application, the technical solutions provided in the embodiments of the present application are described below with reference to specific examples.

[0081] In this embodiment, a privacy leakage detection method based on hybrid learning is proposed. Relevant data including device logs, network traffic and user behavior data are collected, and the collected data is preprocessed to prepare the data for unsupervised feature extraction. In the feature extraction stage, the autoencoder model is trained to reduce the dimension and extract the key features of the data, and the hyperparameters of the model are automatically adjusted based on the Bayesian optimization algorithm to determine the optimal hyperparameter combination to improve performance. Then, a deep learning model for privacy leakage detection (i.e., the above-mentioned privacy leakage detection model) is constructed, and the privacy leakage detection model is trained based on the determined optimal hyperparameter combination and the privacy leakage detection associated features obtained by unsupervised feature extraction.

[0082] For example, in order to improve the generalization ability of the model, when the trained privacy leakage detection model is deployed to a new IoT environment, transfer learning technology can be used to fine-tune the model on the new IoT environment or data set, and the fine-tuned privacy leakage detection model can be deployed to the IoT platform for real-time monitoring, and an alarm system can be set up to detect potential privacy data leakage.

[0083] In this embodiment, if Figure 2 As shown, the implementation steps of privacy leakage detection based on hybrid learning may include: Step 1: Data collection.

[0084] Exemplarily, data generated by IoT devices are collected, including device logs, network traffic, user behavior data, etc.

[0085] Step 2: Data preprocessing.

[0086] Exemplarily, data preprocessing may include: Data cleaning: Remove invalid, erroneous or incomplete data records.

[0087] Feature selection: Select features related to privacy data leakage, such as access frequency, data transmission volume, etc.

[0088] Data encoding: One-Hot encoding is performed on the data after feature selection (which can be called classification data) and converted into numerical form.

[0089] Normalization processing: Scale the value of the encoded data to the range of [0, 1].

[0090] Step 3: Unsupervised feature extraction.

[0091] An unsupervised learning model can be used to extract features from the preprocessed data to obtain corresponding features (such as the privacy leakage detection related features mentioned above).

[0092] Taking the unsupervised learning model as an independent autoencoder (autoencoder for short) as an example, the autoencoder can be used to extract features from the preprocessed data to obtain the feature representation after dimensionality reduction.

[0093] The specific implementation of unsupervised feature extraction using autoencoders is described below.

[0094] 3.1. Determine the candidate autoencoder model: Exemplarily, the candidate autoencoder model may include at least: 1) Deep Autoencoder (DAE): An autoencoder containing multiple hidden layers that can learn more complex feature representations.

[0095] 2) Convolutional Autoencoder (CAE): Use convolutional layers to extract features of image data and is suitable for image data.

[0096] 3) Deep Convolutional Autoencoder (DCAE): It uses convolutional layers to extract features of network traffic and is suitable for network traffic data.

[0097] 4) Recurrent Autoencoder (RCAE): It uses recurrent layers to process time series data and is suitable for time series data.

[0098] 5) Deep LSTM Autoencoder (LSTM-DAE): Uses LSTM layer to process the time series features of network traffic, which is suitable for network traffic data.

[0099] 3.2. Build an autoencoder model: 1) Select a suitable autoencoder model structure from the candidate autoencoder models based on the data type.

[0100] 2) Determine the parameters of the autoencoder model, such as the number of hidden layers, the number of neurons in the hidden layers, the activation function, etc.

[0101] 3.3. Training the Autoencoder: 1) Train the autoencoder model using the autoencoder training data.

[0102] 2) The objective function usually uses the reconstruction error, such as mean squared error (MSE) or cross entropy loss.

[0103] 3) Gradient descent algorithm or stochastic gradient descent algorithm can be used for training.

[0104] 3.4. Evaluating the performance of the autoencoder: 1) Use the autoencoder test data to evaluate the performance of the autoencoder, such as reconstruction error, loss function, etc.

[0105] 2) A variety of metrics can be used to evaluate the performance of the autoencoder, such as reconstruction error, reconstructed image quality, clustering results, etc.

[0106] 3.5. Extract features: 1) Use the trained autoencoder to extract features from the preprocessed data and extract the feature representation of the data.

[0107] 2) The feature representation can be the hidden layer output of the autoencoder or the reconstruction error, etc.

[0108] 3) Feature representation can be used for subsequent tasks such as classification and clustering.

[0109] Using autoencoders for unsupervised feature extraction has at least the following advantages: No labeled data required: The autoencoder model can learn the feature representation of data without labeled data, and thus can be applied to many scenarios where labeled data is not available.

[0110] Rich feature representation: The feature representation learned by autoencoders is usually richer than traditional feature selection methods because it can capture the complex structure and patterns of the data.

[0111] Robustness: The autoencoder model is robust to noise and outliers because it is able to learn the intrinsic characteristics of the data.

[0112] Step 4: Hyperparameter optimization.

[0113] Define the search space: Determine the search range for hyperparameters such as learning rate, number of hidden layers, number of neurons, etc.

[0114] Bayesian Optimization: Use the Bayesian optimization algorithm to automatically search for the best hyperparameter combination.

[0115] Step 5: Build a deep learning model.

[0116] Model architecture design: Design the neural network architecture including input layer, hidden layer and output layer.

[0117] Classifier design: Use softmax function in the output layer for multi-classification.

[0118] Step 6: Model training and tuning.

[0119] Training data preparation: The data after preprocessing and feature extraction is used for model training.

[0120] Tuning: Adjust the model structure and parameters based on the performance of the validation set.

[0121] Step 7: Model evaluation.

[0122] Performance metrics: Use metrics such as accuracy, recall, and F1 score to evaluate model performance.

[0123] Confusion Matrix: Analyze the performance of the model on different categories.

[0124] The confusion matrix is ​​used to analyze the performance of the model in identifying different types of privacy data. Each category corresponds to a subcategory of the privacy data type. The confusion matrix shows the model's prediction results for each category in the test dataset, as well as the actual category.

[0125] Exemplarily, the confusion matrix contains the following elements: True Positive (TP): The model correctly labels the actual category as this category.

[0126] False Positive (FP): The model incorrectly labels the actual category as this category.

[0127] False Negative (FN): The model mistakenly labels the actual category as other categories.

[0128] True Negative (TN): The model correctly labels the actual category as other categories.

[0129] By analyzing the confusion matrix, we can evaluate the performance of the model on different categories: Precision: The proportion of samples that the model correctly predicts as this category out of all samples predicted as this category.

[0130] Recall (Recall / Detection Rate): The proportion of samples that the model correctly predicts to be of this category to all samples that are actually of this category.

[0131] F1-score: The harmonic mean of precision and recall, used to balance the two.

[0132] TNR (True Negative Rate, also known as Specificity): The proportion of samples that the model correctly predicts are not of this category.

[0133] FDR (False Discovery Rate): The proportion of samples that are incorrectly labeled as a category.

[0134] Exemplarily, when the model evaluation result meets the requirements, step 8 can be executed; when the model evaluation result does not meet the requirements, the model parameters can be adjusted and step 6 can be executed again.

[0135] Step 8: Transfer learning.

[0136] New dataset adaptation: Apply the model to new IoT environments or datasets.

[0137] Fine-tuning: Fine-tuning on a new dataset to adapt to a different data distribution.

[0138] Step 9: Real-time detection deployment.

[0139] System integration: Integrate the trained model into the IoT platform.

[0140] Monitoring setup: Configure a real-time monitoring system to detect potential privacy data leaks.

[0141] The method provided by the present application is described above. The device provided by the present application is described below: See also Figure 3 , which is a structural diagram of a privacy leakage detection device based on hybrid learning provided in an embodiment of the present application, such as Figure 3 As shown, the privacy leakage detection device based on hybrid learning may include: The acquisition unit 310 is used to acquire privacy leakage detection related data; wherein the privacy leakage detection related data is obtained by processing the data generated by the IoT device; The feature extraction unit 320 is used to extract the privacy leakage detection associated features from the privacy leakage detection associated data using an unsupervised learning model to obtain the privacy leakage detection associated features; wherein the unsupervised learning model includes an independent autoencoder and a stacked autoencoder; the privacy leakage detection associated features are obtained by fusing the first privacy leakage detection associated features and the second privacy leakage detection associated features; the first privacy leakage detection associated features are extracted using the independent autoencoder, and the second privacy leakage detection associated features are extracted using the stacked autoencoder; and, A determination unit 330 is used to determine an optimal hyperparameter combination of the deep learning model using a hyperparameter optimization scheme; A training unit 340 is used to construct a privacy leakage detection model, and train the privacy leakage detection model based on the initial optimal hyperparameter combination and the privacy leakage detection associated features to obtain a trained privacy leakage detection model; The transfer learning unit 350 is used to perform transfer learning on the trained privacy leakage detection model for the IoT environment to be detected, so as to obtain a privacy leakage detection model to be deployed for performing privacy leakage detection on the IoT environment to be detected; wherein the privacy leakage detection model to be deployed is used to perform privacy leakage detection on IoT devices in the IoT environment to be detected.

[0142] The trained privacy leakage detection model is used to perform privacy leakage detection on IoT devices.

[0143] In some embodiments, the acquiring unit 310 acquires the privacy leakage detection related data, including: Obtain data generated by IoT devices; Clean the acquired data and select features related to privacy leakage detection from the cleaned data; The features related to the privacy leakage detection are converted into numerical form and normalized to obtain privacy leakage detection related data.

[0144] In some embodiments, The feature extraction unit 320 uses an unsupervised learning model to extract privacy leakage detection related features from the privacy leakage detection related data to obtain privacy leakage detection related features, including: Using an independent autoencoder, extracting privacy leakage detection related features from the privacy leakage detection related data to obtain a first privacy leakage detection related feature; and Using a stacked autoencoder, extracting privacy leakage detection associated features from the privacy leakage detection associated data to obtain a second privacy leakage detection associated feature; The first privacy leakage detection associated feature and the second privacy leakage detection associated feature are feature fused to obtain a final privacy leakage detection associated feature.

[0145] In some embodiments, the unsupervised learning model includes multiple types of independent autoencoders, and / or multiple types of stacked autoencoders; The first privacy leakage detection associated features include first privacy leakage detection associated features extracted by multiple different independent autoencoders, and / or the second privacy leakage detection associated features include second privacy leakage detection associated features extracted by multiple different stacked autoencoders.

[0146] In some embodiments, the feature extraction unit 320 performs feature fusion on the first privacy leakage detection associated feature and the second privacy leakage detection associated feature to obtain a final privacy leakage detection associated feature, including: According to the maximum dimension of the privacy leakage detection associated feature, the first privacy leakage detection associated feature and the second privacy leakage detection associated feature are aligned by filling in zeros; For any dimension, the non-zero value of the first privacy leakage detection associated feature in this dimension and the non-zero value of the second privacy leakage detection associated feature in this dimension are fused using an averaging strategy to obtain the final value of the privacy leakage detection associated feature in this dimension.

[0147] In some embodiments, the feature extraction unit 320 uses an average strategy to fuse the non-zero value of the first privacy leakage detection associated feature in any dimension and the non-zero value of the second privacy leakage detection associated feature in the dimension to obtain the final value of the privacy leakage detection associated feature in the dimension, including: For any dimension, if the number of non-zero values ​​in the value of the first privacy leakage detection associated feature of the dimension and the value of the second privacy leakage detection associated feature of the dimension is 1, the non-zero value is determined as the final value of the privacy leakage detection associated feature of the dimension; When there is no non-zero value in the value of the dimension of the first privacy leakage detection associated feature and the value of the dimension of the second privacy leakage detection associated feature, it is determined that the final value of the dimension of the privacy leakage detection associated feature is 0.

[0148] In some embodiments, the determining unit 330 determines the initial optimal hyperparameter combination of the deep learning model using a hyperparameter optimization scheme, including: Based on the defined hyperparameter search range, a Bayesian optimization algorithm is used to determine the initial optimal hyperparameter combination of the deep learning model; wherein the hyperparameters include one or more of a learning rate, a number of hidden layers, and a number of neurons.

[0149] In some embodiments, the training unit 340 trains the privacy leakage detection model using the privacy leakage detection associated features according to the initial optimal hyperparameter combination, including: Generate a training set and a validation set based on the privacy leakage detection related features; According to the initial optimal hyperparameter combination, the privacy leakage detection model is trained using the training set; Using the validation set to tune the model structure and parameters of the trained privacy leakage detection model; Use the test set to evaluate the performance of the optimized privacy leakage detection model; When the performance evaluation result meets the performance requirements, the current privacy leakage detection model is determined as the trained privacy leakage detection model; When the performance evaluation result does not meet the performance requirements, the current privacy leakage detection model is iteratively trained using the training set.

[0150] In some embodiments, the transfer learning unit 350 performs transfer learning on the trained privacy leakage detection model for the IoT environment to be detected, and obtains a privacy leakage detection model to be deployed for performing privacy leakage detection on the IoT environment to be detected, including: The data generated by the IoT environment to be detected is obtained, and the trained privacy leakage detection model is fine-tuned based on the obtained data; wherein the fine-tuned privacy leakage detection model is used to perform privacy leakage detection on the IoT environment to be detected.

[0151] An embodiment of the present application provides an electronic device, including a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor is used to execute the machine executable instructions to implement the privacy leakage detection method based on hybrid learning described above.

[0152] See also Figure 4 , is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. The electronic device may include a processor 401 and a memory 402 storing machine executable instructions. The processor 401 and the memory 402 may communicate via a system bus 403. In addition, by reading and executing the machine executable instructions corresponding to the privacy leakage detection logic based on hybrid learning in the memory 402, the processor 401 may execute the privacy leakage detection method based on hybrid learning described above.

[0153] The memory 402 mentioned herein may be any electronic, magnetic, optical or other physical storage device that may contain or store information, such as executable instructions, data, etc. For example, the machine-readable storage medium may be: RAM (RadomAccess Memory, random access memory), volatile memory, non-volatile memory, flash memory, storage drive (such as hard disk drive), solid state drive, any type of storage disk (such as optical disk, DVD, etc.), or similar storage medium, or a combination thereof.

[0154] In some embodiments, a machine-readable storage medium is also provided. Figure 4 The memory 402 in the machine-readable storage medium stores machine executable instructions, and when the machine executable instructions are executed by the processor, the privacy leakage detection method based on hybrid learning described above is implemented. For example, the storage medium can be ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

Claims

1. A privacy leakage detection method based on hybrid learning, characterized in that: include: Obtaining privacy leakage detection associated data; wherein the privacy leakage detection associated data is obtained by processing data generated by an IoT device; Using an unsupervised learning model, extracting privacy leakage detection associated features from the privacy leakage detection associated data to obtain privacy leakage detection associated features; wherein the unsupervised learning model includes an independent autoencoder and a stacked autoencoder; the privacy leakage detection associated features are obtained by fusing the first privacy leakage detection associated features and the second privacy leakage detection associated features; the first privacy leakage detection associated features are extracted using the independent autoencoder, and the second privacy leakage detection associated features are extracted using the stacked autoencoder; and, Use hyperparameter optimization schemes to determine the initial optimal hyperparameter combination of deep learning models; Constructing a privacy leakage detection model, and training the privacy leakage detection model using the privacy leakage detection associated features according to the initial optimal hyperparameter combination to obtain a trained privacy leakage detection model; For the Internet of Things environment to be detected, transfer learning is performed on the trained privacy leakage detection model to obtain a privacy leakage detection model to be deployed for privacy leakage detection of the Internet of Things environment to be detected; wherein the privacy leakage detection model to be deployed is used to perform privacy leakage detection on the Internet of Things devices in the Internet of Things environment to be detected.

2. The method according to claim 1, characterized in that The obtaining of privacy leakage detection related data includes: Access data generated by IoT devices; Clean the acquired data and select features related to privacy leakage detection from the cleaned data; The features related to the privacy leakage detection are converted into numerical form and normalized to obtain privacy leakage detection related data.

3. The method according to claim 1, characterized in that The unsupervised learning model is used to extract privacy leakage detection related features from the privacy leakage detection related data to obtain privacy leakage detection related features, including: Using an independent autoencoder, extracting privacy leakage detection related features from the privacy leakage detection related data to obtain a first privacy leakage detection related feature; and Using a stacked autoencoder, extracting privacy leakage detection associated features from the privacy leakage detection associated data to obtain a second privacy leakage detection associated feature; The first privacy leakage detection associated feature and the second privacy leakage detection associated feature are feature fused to obtain a final privacy leakage detection associated feature.

4. The method according to claim 3, characterized in that The unsupervised learning model includes multiple types of independent autoencoders and / or multiple types of stacked autoencoders; The first privacy leakage detection associated features include first privacy leakage detection associated features extracted by multiple different independent autoencoders, and / or the second privacy leakage detection associated features include second privacy leakage detection associated features extracted by multiple different stacked autoencoders.

5. The method according to claim 3 or 4, characterized in that: The step of fusing the first privacy leakage detection associated feature with the second privacy leakage detection associated feature to obtain a final privacy leakage detection associated feature includes: According to the maximum dimension of the privacy leakage detection associated feature, the first privacy leakage detection associated feature and the second privacy leakage detection associated feature are aligned by filling in zeros; For any dimension, the non-zero value of the first privacy leakage detection associated feature in this dimension and the non-zero value of the second privacy leakage detection associated feature in this dimension are fused using an averaging strategy to obtain the final value of the privacy leakage detection associated feature in this dimension.

6. The method according to claim 5, characterized in that For any dimension, the non-zero value of the first privacy leakage detection associated feature in the dimension and the non-zero value of the second privacy leakage detection associated feature in the dimension are merged using an averaging strategy to obtain a final value of the privacy leakage detection associated feature in the dimension, including: For any dimension, if the number of non-zero values ​​in the value of the first privacy leakage detection associated feature of the dimension and the value of the second privacy leakage detection associated feature of the dimension is 1, the non-zero value is determined as the final value of the privacy leakage detection associated feature of the dimension; When there is no non-zero value in the value of the dimension of the first privacy leakage detection associated feature and the value of the dimension of the second privacy leakage detection associated feature, it is determined that the final value of the dimension of the privacy leakage detection associated feature is 0.

7. The method according to claim 1, characterized in that The method of using a hyperparameter optimization scheme to determine the initial optimal hyperparameter combination of the deep learning model includes: Based on the defined hyperparameter search range, a Bayesian optimization algorithm is used to determine the initial optimal hyperparameter combination of the deep learning model; wherein the hyperparameters include one or more of a learning rate, a number of hidden layers, and a number of neurons.

8. The method according to claim 1, characterized in that The step of training the privacy leakage detection model based on the initial optimal hyperparameter combination and using the privacy leakage detection associated features includes: Generate a training set and a validation set based on the privacy leakage detection related features; According to the initial optimal hyperparameter combination, the privacy leakage detection model is trained using the training set; Using the validation set to tune the model structure and parameters of the trained privacy leakage detection model; Use the test set to evaluate the performance of the optimized privacy leakage detection model; When the performance evaluation result meets the performance requirements, the current privacy leakage detection model is determined as the trained privacy leakage detection model; When the performance evaluation result does not meet the performance requirements, the current privacy leakage detection model is iteratively trained using the training set.

9. The method according to claim 1, characterized in that: The method of performing transfer learning on the trained privacy leakage detection model for the IoT environment to be detected to obtain a privacy leakage detection model to be deployed for performing privacy leakage detection on the IoT environment to be detected includes: The data generated by the IoT environment to be detected is obtained, and the trained privacy leakage detection model is fine-tuned according to the obtained data to obtain a privacy leakage detection model to be deployed for performing privacy leakage detection on the IoT environment to be detected.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores machine executable instructions that can be executed by the processor, and the processor is used to execute the machine executable instructions to implement the method according to any one of claims 1 to 9.

Images