User continuous authentication method and system based on reconstruction probability Wasserstein auto-encoder

By adopting the reconstruction probability Wasserstein autoencoder in mobile user authentication, the problem of non-end model design and the difficulty of selecting reconstruction error thresholds is solved, and higher authentication accuracy and consistency are achieved.

CN119026113APending Publication Date: 2024-11-26CHONGQING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411236295.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In the existing mobile user authentication technology, non-end-to-end model modules are difficult to design and assemble, and the reconstruction error threshold based on the autoencoder is difficult to determine, resulting in authentication accuracy and consistency problems.

Method used

Using the Wasserstein autoencoder (RPWAE) based on reconstruction probability is used as an end-to-end model, and user authentication is performed by calculating the reconstruction probability, avoiding the assembly problem of non-end models, and using reconstruction probability instead of reconstruction error threshold to improve Accuracy and consistency of certification.

Benefits of technology

It achieves higher user continuous authentication accuracy and consistency, avoids assembly difficulties of non-end models and uncertainty in the selection of thresholds for reconstruction errors, and provides a clear and objective authentication standard.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026113A_ABST
    Figure CN119026113A_ABST
Patent Text Reader

Abstract

The invention provides a user continuous authentication method and system based on a reconstruction probability Wasserstein auto-encoder, and the method comprises the steps: collecting behavior data of interaction between a user and a mobile device, inputting the sampling data into a pre-trained reconstruction probability Wasserstein auto-encoder for coding and decoding after standardization processing and segmentation, obtaining a reconstruction probability, and carrying out the authentication of the reconstruction probability Wasserstein auto-encoder. According to the method, the user validity is authenticated through comparison with a preset threshold value, the encoder comprises an encoder and a decoder, training is carried out only through legal user data, the Wasserstein distance between model distribution and target distribution can be minimized, authentication judgment is carried out through the reconstruction probability, the problem that a reconstruction error threshold value is difficult to select in a traditional self-encoder-based reconstruction model is solved, and the user validity is improved. And the continuous authentication accuracy of the user is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of continuous authentication, and specifically to a method and system for continuous user authentication based on a reconstructed probabilistic Wasserstein autoencoder. Background Art

[0002] With the rapid development of mobile communication technology, mobile devices have become an indispensable part of people's daily lives. Nowadays, people are accustomed to social networking, online shopping, and bank transfers on mobile devices. With people's frequent use, more and more sensitive information is stored on mobile devices, and the security of mobile devices has become more and more important. User authentication is a key technology to prevent malicious access to mobile devices. Traditional mobile user authentication technologies mainly include password and graphic pattern authentication, but this type of authentication is vulnerable to smudge attacks and password inference attacks, and these authentication methods stay at the entry point. In recent years, biometric authentication technology has also become very mature. This method is mainly based on the user's unique physiological characteristics for authentication, such as face, fingerprint, iris, finger vein, etc. These methods greatly improve the security of the identity authentication system because physiological characteristics are difficult to replicate. However, their disadvantage is that this method still stays at the entry point.

[0003] The emergence of continuous authentication has effectively solved the problem of single authentication. The current mainstream continuous authentication system authenticates users by identifying their behavioral characteristics, such as touch gestures, gait, keystroke dynamics, etc. Specifically, identity authentication is achieved by identifying the characteristics of the data generated by user behavior on the mobile phone sensors. When a user operates a mobile phone or carries a mobile phone while walking, various sensors such as accelerometers, gyroscopes, magnetometers, pressure gauges, etc. will generate various data, which reflect the user's behavioral characteristics to a certain extent. The continuous authentication system extracts features from these data to authenticate the user.

[0004] There are two main types of mainstream authentication models. One is based on feature extractors and single classifiers. This model first uses deep learning or machine learning methods to extract features from these data, and then uses the extracted features to train a single classifier for user identity authentication, and uses the single classifier to classify legitimate users and illegitimate users. The other is based on autoencoder reconstruction. This model uses an autoencoder to extract features from sensor data and reconstruct the data to obtain reconstructed data. During authentication, the reconstruction error between the input data and the reconstructed data is calculated and a threshold is set. If the reconstruction error is less than the threshold, the input data and the reconstructed data are considered to be very close and authenticated as a legitimate user; otherwise, it is an illegal user.

[0005] However, in the prior art, for models based on feature extractors and single classifiers, they are all non-end-to-end models, and these non-end-to-end models have three disadvantages. First, different modules (feature extraction modules and single classifiers) are difficult to design and assemble together. Second, since the modules are trained separately, the errors between each component may overlap, thereby amplifying the errors. Finally, the above method uses both legal and illegal user data when training the classifier, which is not feasible in practical applications because the device can only collect data from legal users. Although the model based on autoencoder reconstruction is an end-to-end model that solves the above problems, it uses the reconstruction error as the threshold. This method is difficult to determine a clear threshold in practical applications. For different test sets, the division of thresholds is likely to be quite different. For example, selecting a threshold of 0.5 on test set A can achieve an accuracy of 95%, but selecting a threshold of 0.1 on test set B can achieve an accuracy of 95%. There is no clear and objective threshold selection. Summary of the invention

[0006] In order to solve the above problems, this application proposes a user continuous authentication method and system based on reconstruction probability Wasserstein autoencoder. Reconstruction probability Wasserstein autoencoder (RPWAE) is an end-to-end model that avoids the shortcomings of non-end-to-end models and uses reconstruction probability for authentication, solving the problem of difficulty in selecting reconstruction error threshold.

[0007] The technical solution of this application is as follows:

[0008] In a first aspect, the present application provides a method for continuous user authentication based on a reconstruction probabilistic Wasserstein autoencoder, comprising: collecting behavioral data of the interaction between a user and a mobile device;

[0009] Standardizing and segmenting the collected behavioral data to obtain a standard data set;

[0010] Sample the standard data set, input the sampled data into the pre-trained reconstruction probability Wasserstein autoencoder for encoding and decoding, and calculate the reconstruction probability;

[0011] Compare the reconstruction probability with the preset threshold. If the reconstruction probability is greater than the preset threshold, the authenticated user is a legitimate user; otherwise, the authenticated user is an illegal user.

[0012] The calculation formula of the Wasserstein distance of the Wasserstein autoencoder is:

[0013]

[0014] Among them, D WAE(P X ,P G ) is the Wasserstein distance, P X is the model distribution, P G is the target distribution; LogProb(X,G(Z)) is the Log value of the reconstruction probability between the target sample and the reconstructed sample, X is the data set collected by the system, G(Z) is the reconstruction probability Wasserstein autoencoder reconstructs X, and Z is the latent variable of the autoencoder; MMD k (Q Z ,P Z ) is the regularization factor, λ is the hyperparameter, Q Z is a latent variable, P Z is the prior distribution.

[0015] Preferably, the calculation formula of the reconstruction probability is:

[0016]

[0017] ReconProb(X,G(Z))=e LogProb(X,G(Z)) ,

[0018] Among them, x (i) is the sample segmented from X, For x (i) The reconstructed mean of For x (i) The reconstruction standard deviation of the original sample x (i) , in the reconstruction mean With standard deviation ReconProb is the exponential of the Log value e of the reconstruction probability, that is, the logarithm is restored to the reconstruction probability.

[0019] Preferably, the reconstructed probability Wasserstein autoencoder can achieve the model distribution P during the training process. G and the target distribution P X Minimize the Wasserstein distance between them.

[0020] Preferably, the reconstructed probability Wasserstein autoencoder includes an encoder and a decoder; only legitimate user data is used for training, and the legitimate user data is input into the encoder and encoded to obtain the corresponding latent variable Q Z , the latent variable is decoded by the decoder to output the reconstruction probability.

[0021] Preferably, the encoder includes four convolutional layers and one linear layer, and the decoder includes one linear layer and four deconvolutional layers.

[0022] Preferably, the user's behavior data is collected through an accelerometer, a gyroscope and a magnetometer in the mobile device.

[0023] In a second aspect, the present application further provides a user continuous authentication system based on a reconstruction probability Wasserstein autoencoder, comprising:

[0024] A data collection module, used to collect behavioral data of the interaction between users and mobile devices;

[0025] A data processing module, used for standardizing and segmenting the collected behavior data to obtain a standard data set;

[0026] The user authentication module is used to sample the standard data set, input the sampled data into the pre-trained reconstruction probability Wasserstein autoencoder for encoding and decoding, and calculate the reconstruction probability;

[0027] The judgment module is used to compare the reconstruction probability with a preset threshold. If the reconstruction probability is greater than the threshold, the user is authenticated as a legitimate user; otherwise, the user is authenticated as an illegal user.

[0028] The calculation formula of the Wasserstein distance of the Wasserstein autoencoder is:

[0029]

[0030] Among them, D WAE (P X ,P G ) is the Wasserstein distance, P X is the model distribution, P G is the target distribution; LogProb(X,G(Z)) is the Log value of the reconstruction probability between the target sample and the reconstructed sample, X is the data set collected by the system, G(Z) is the reconstruction probability Wasserstein autoencoder reconstructs X, and Z is the latent variable of the autoencoder; MMD k (Q Z ,P Z ) is the regularization factor, λ is the hyperparameter, Q Z is a latent variable, P Z is the prior distribution.

[0031] This application has at least the following beneficial effects:

[0032] 1. The reconstructed probabilistic Wasserstein autoencoder (RPWAE) used in this application is an end-to-end model, which is more convenient and avoids the problem that the feature extraction module and the single classifier in the non-end-to-end model are difficult to design and assemble. It only uses legitimate user data for training, which is more in line with the situation that the device can only collect legitimate user data in actual applications, and the trained model has a higher accuracy rate.

[0033] 2. Most studies use mean square error to measure the quality of reconstruction. The reconstruction result obtained by this method is only the reconstruction of the input variable itself, not the reconstruction of the input variable distribution. This is not conducive to continuous mobile authentication. Unlike anomaly detection, in the authentication process, we do not need to find multiple abnormal data points in the data, but need to identify whether the distribution of data generated by the device comes from legitimate users. To this end, we use reconstruction probability to authenticate users. This method determines whether it belongs to the data distribution of legitimate users by reconstructing the probability density of the data. It is difficult to determine a clear threshold when using traditional reconstruction error for detection. For different test sets, the division of thresholds is likely to be quite different. For example, selecting a threshold of 0.5 on test set A can achieve an accuracy of 95%, but a threshold of 0.1 needs to be selected on test set B to achieve an accuracy of 95%. There is no clear and objective threshold selection. If the reconstruction probability of this application is used, for any data or test set, a probability of 0.9 is 0.9, so it is more objective and reasonable. Using reconstruction probability for authentication solves the problem that the reconstruction error threshold is difficult to select in the traditional autoencoder-based reconstruction model. Different test sets may require different thresholds, and the reconstructed probability authentication method makes the error selection more reasonable and improves the accuracy of user continuous authentication.

[0034] 3. In the existing technology, continuous user authentication usually relies on algorithms such as password authentication and machine learning. The reconstructed probabilistic Wasserstein autoencoder can extract representative low-dimensional feature vectors from user behavior data through its powerful feature extraction capabilities. These feature vectors can not only effectively compress the original data, but also retain key information, providing strong support for subsequent authentication decisions. The application of the reconstructed probabilistic Wasserstein autoencoder in this technical field improves the accuracy of continuous user authentication. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method for continuous authentication of users of this application;

[0036] Figure 2 It is the composition diagram of the reconstructed probability Wasserstein autoencoder;

[0037] Figure 3It is a flowchart of data standardization processing and segmentation;

[0038] Figure 4 is a flow chart of data collection;

[0039] Figure 5 It is the training flow chart of the reconstruction probability Wasserstein autoencoder;

[0040] Figure 6 This is a diagram of the structure of the user continuous authentication system of this application. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. It is understood that the specific embodiments described herein are only used to explain the related applications, rather than to limit the applications. It should also be noted that, for ease of description, only the parts related to the related applications are shown in the drawings.

[0042] In order to better understand the technical solution of the present application, the reconstruction probability Wasserstein autoencoder and its basic theory are explained below.

[0043] First of all, the autoencoder (AE) is a neural network model that learns to effectively represent data through unsupervised learning. Its core idea is to compress the input data into a low-dimensional latent representation through the encoder, and then reconstruct the original input data from the representation through the decoder. Autoencoders have shown excellent performance in data dimensionality reduction, feature extraction, data denoising, anomaly detection, and generative models. With the development of deep learning, many variants of autoencoders have emerged to adapt to different application scenarios and needs. Among them, the variational autoencoder (VAE): introduces randomness in the latent space to make the generated samples diverse, and realizes continuous sampling of the latent space by optimizing the variational lower bound (ELBO). VAE is a generative model that can generate new data samples with similar distribution to the original data, and has a wide range of applications in image generation, text generation and other fields.

[0044] The autoencoder consists of two parts: the encoder and the decoder. Encoder: responsible for mapping the input data to a low-dimensional latent space. This process is usually implemented through a multi-layer neural network. Each layer will transform and compress the data to obtain a low-dimensional encoded representation. This encoded representation is a compressed form of the input data. It contains the main feature information of the input data, but removes redundancy and noise. Decoder: responsible for mapping this low-dimensional representation back to the original input space, thereby reconstructing the input data. In contrast to the encoder, the decoder upsamples and transforms layer by layer through a multi-layer neural network to restore the low-dimensional representation to an approximate form of the original input data. The goal of the decoder is to make the reconstructed data as close as possible to the original input data, that is, to minimize the reconstruction error.

[0045] Among them, the loss function is a key component used to evaluate the performance of the model during the autoencoder training process. It usually aims to minimize the difference between the original input and the autoencoder output (i.e., the reconstructed input). Depending on the data type and distribution, the reconstruction loss can be calculated in different ways, such as the mean squared error (MSE) and cross-entropy (Cross-Entropy) in the reconstruction error, which are indicators that measure the difference between the reconstructed signal and the original signal. The smaller the reconstruction error, the closer the reconstructed signal is to the original signal, that is, the higher the reconstruction quality.

[0046] The reconstruction probability represents the probability density of the original data under the reconstruction variance and reconstruction standard deviation. The larger the reconstruction probability, the higher the probability that the input data belongs to the legitimate user data distribution, and vice versa. This means that the reconstruction probability can provide high-quality reconstruction results in most cases, with high stability and reliability. In summary, the reconstruction error is the reconstruction of the data itself, while the reconstruction probability is the reconstruction of the input variable distribution.

[0047] The Wasserstein Auto-Encoders (WAE) introduce the Wasserstein distance into the loss function of the autoencoder. Wasserstein distance is a mathematical method for measuring the difference between two probability distributions. It can be achieved by finding an optimal transmission plan (i.e., mass movement plan) that minimizes the total movement cost from one distribution to another. The model is trained by minimizing the loss function of the Wasserstein distance, which solves the gradient discontinuity problem in the traditional loss function and better supports the training of generative models.

[0048] In the prior art, the traditional formula uses mean square error (MSE) as the main term for calculating Wasserstein distance. The smaller the MSE, the closer the original data and the reconstructed data are. The specific calculation formula is:

[0049]

[0050] where x (i) Represents the original data, represents the corresponding reconstructed data, is a regularization matrix used to penalize the difference between the distribution of the hidden code Z and the target distribution.

[0051] Therefore, the present application combines reconstruction probability and Wasserstein autoencoder, proposes a generative model algorithm that combines Wasserstein distance and autoencoder ideas, and improves the Wasserstein distance calculation formula. The distance formula of the present application uses reconstruction probability as the main item for calculating Wasserstein distance. The greater the reconstruction probability, the greater the probability that the input data belongs to the distribution of legitimate users. Training is to minimize the distance, so a negative sign is added, i.e., negative correction. The present application not only focuses on the accuracy of the reconstruction probability, but also optimizes the distribution of generated data by introducing Wasserstein distance, thereby further improving the quality and diversity of generated samples.

[0052] Embodiment 1

[0053] like Figure 1-Figure 3 As shown, this embodiment is a user continuous authentication method based on reconstruction probability Wasserstein auto-encoder (Reconstruction Rrobability Wasserstein Auto-Encoders Authentication, referred to as RPWAEAuth), including the following steps:

[0054] S20: Collect behavioral data on the interaction between users and mobile devices;

[0055] S21: performing standardization processing and segmentation on the collected behavior data to obtain a standard data set;

[0056] S22: Sampling the standard data set, inputting the sampled data into the pre-trained reconstruction probability Wasserstein autoencoder for encoding and decoding, and calculating the reconstruction probability;

[0057] S23: If it is determined that the reconstruction probability is greater than a preset threshold, the user is a legitimate user;

[0058] The calculation formula of the Wasserstein distance of the Wasserstein autoencoder is:

[0059]

[0060] Among them, D WAE (P X ,P G ) is the Wasserstein distance, P X is the model distribution, P G is the target distribution; LogProb(X,G(Z)) is the Log value of the reconstruction probability between the target sample and the reconstructed sample, X is the data set collected by the system, G(Z) is the reconstruction probability Wasserstein autoencoder reconstructs X, and Z is the latent variable of the autoencoder; MMD k (Q Z ,P Z ) is the regularization factor, λ is the hyperparameter, Q Z is a latent variable, P Z is the prior distribution.

[0061] Preferably, Figure 2 As shown, the reconstruction probability Wasserstein autoencoder includes an encoder (A31) and a decoder (A33); only legitimate user data is used for training, and after input into the encoder (A31), encoding is performed to obtain the corresponding latent variable (A32), and the latent variable (A32) is decoded by the decoder (A33) to output the reconstruction probability (A34). The encoder includes four convolutional layers (A311-A314) and one linear layer (A315), and the decoder includes one linear layer (A331) and four deconvolutional layers (A332-A335).

[0062] Preferably, the preset threshold is the average reconstruction probability when the reconstruction probability Wasserstein autoencoder minimizes the Wasserstein distance during the training process.

[0063] Preferably, the data standardization program (A21) and the data segmentation program (A22) perform data standardization and segmentation as follows: Figure 3 The data normalization procedure (A21) is as follows: The data collected in each cycle t is normalized using a d×n matrix D i The final complete data set can be represented by D. Then, each axis of each sensor is normalized to (0,1] by the maximum and minimum normalization method, that is, Then through d norm=(d-0.5) / 0.5) to normalize the data to (-1,1), where i represents the axis of the sensor. Finally, the normalized data set D norm The data segmentation procedure (A22) is used to train the data set D with a window size of 200 sampling points (2 seconds). norm The split data will be used for registration training and certification testing.

[0064] Preferably, Figure 4 As shown, the user's behavior data is collected through the accelerometer (A11), gyroscope (A12) and magnetometer (A13) in the mobile device.

[0065] Preferably, the calculation formula of the reconstruction probability is:

[0066]

[0067]

[0068] ReconProb(X,G(Z))=e LogProb(X,G(Z))

[0069] Here, we assume that X and Y represent the training and testing datasets of RPWAE, x (i) and (i) represents the samples split from X and Y. In addition, we use Q and G as encoder and decoder, and Z as the latent variable. During training, each x (i) Input into the model and get x (i) The reconstructed mean and the reconstruction standard deviation Then, we calculate the logarithm of the reconstruction probability to ensure that there is a larger gradient around zero. Where LogProb represents the Log value of the reconstruction probability, X is the data set collected by the system, G(Z) is the data after the reconstruction probability Wasserstein autoencoder reconstructs X, and Z is the latent variable of the autoencoder. (i) is the sample segmented from X, For x (i) The reconstructed mean of For x (i) The reconstruction probability P is the original sample x (i) In reconstructing the mean With standard deviation ReconProb represents the conditional probability of the reconstruction probability after the log value of the reconstruction probability is raised to an exponential of e.

[0070] Preferably, the pre-trained reconstruction probability Wasserstein autoencoder is divided into two stages: registration stage and certification stage. The registration stage is mainly to train the RPWAE, and the process is as follows: Figure 5 As shown. Specifically including the following steps:

[0071] S10: Collect behavioral data on the interaction between users and mobile devices. This is the basis of the entire process, because all subsequent steps rely on the collected data.

[0072] S11: The collected data needs to be standardized and segmented to form a standard data set. This step is to ensure the consistency and comparability of the data and facilitate subsequent analysis and processing.

[0073] S12: Sample from the standard data set and encode the sampled samples into latent variables to obtain latent variables. Latent variable encoding is a data dimensionality reduction technique that can convert high-dimensional data into low-dimensional latent representations to facilitate capturing key information in the data.

[0074] S13: Sample from the prior distribution and calculate the maximum average difference between the prior distribution and the latent variable, and take its square root as the regularization term. Regularization terms are often used in machine learning to prevent overfitting and improve the generalization ability of the model. Here, it can be used to control the complexity of the latent variable encoding process to ensure that the model is not too complex and loses its generalization ability.

[0075] S14: Decode the latent variables to obtain the reconstruction probability, and take the log value of the reconstruction probability as the loss value. This step is to evaluate the reconstruction ability of the model, that is, whether the model can recover the key information of the original data from the latent variables. The smaller the loss value, the stronger the reconstruction ability of the model.

[0076] S15: Train the authentication module based on the loss value and regularization term. Authentication is the ultimate goal of the entire process. It recognizes the user's identity or behavior characteristics by learning the data features of the user's interaction with the mobile device. By minimizing the loss value and regularization term, an authentication module with better recognition ability can be trained.

[0077] The average reconstruction probability when the reconstruction probability Wasserstein autoencoder minimizes the Wasserstein distance during the training process completes the registration stage; then the newly collected data is input into the reconstruction probability Wasserstein autoencoder for verification and judgment, entering the authentication stage, that is, Figure 1 process.

[0078] During training, unlike ordinary autoencoders, RPWAEAuth minimizes the model distribution P G and the target distribution PX To this end, the decoder accurately reconstructs the sample based on the reconstruction probability loss function, while the encoder is responsible for two opposing goals: 1) forcing the latent variable Q Z becomes a continuous distribution and ensures that Q Z Close to the prior distribution P Z . This ensures that for every input x (i) In the latent variable space, it corresponds to a z (i) , different z (i) Keep a certain distance between them. The latent variable codes of different training samples are far apart, ensuring a good reconstruction effect; 2) Ensure that the generated code contains enough feature information so that the decoder can accurately reconstruct the sample. Therefore, the calculation formula of the training objective of RPWAEAuth is expressed as:

[0079]

[0080] LogProb(X,G(Z)) measures the difference in reconstruction probability between the target sample and the reconstructed sample, λ is a hyperparameter, MMD k (Q Z ,P Z ) represents the latent variable distribution Q Z With the prior distribution P Z The main reason for using the square root of the MMD is that it brings the two distributions closer together.

[0081] Table 1

[0082]

[0083] Specifically, an actual training process of RPWAEAuth is shown in Table 1.

[0084] Before training, we predefine the regularization coefficient λ and select a characteristic positive definite kernel k. During the training process, a batch of training set samples {x (1) ,…,x (n)} Input RPWAE, from the prior standard normal distribution P Z Sampling {z (1) ,…,z (n) Then, from the encoder Q φ (Z|x (i) ) in which the latent variables are sampled From the decoder Get the reconstruction mean from and variance Finally, the loss function is calculated (the loss function is the Wasserstein distance, which is essentially the same parameter), and the model parameters are updated in the direction of minimizing the loss function, thereby minimizing the model distribution P G and the target distribution P X The Wasserstein distance between them.

[0085] Certification stages such as process Figure 1 As shown. Given y in the test set Y (i) , we use the reconstruction probability to authenticate the user, which represents the probability density of the original data under the reconstruction variance and reconstruction standard deviation. If the "reconstruction probability" is greater than the threshold α, then y (i) The data is generated by a legitimate user and authenticated by the legitimate user. Otherwise, y (i) The data is considered to be generated by the impostor and the user is authenticated as the impostor.

[0086] In general, the Reconstruction Probabilistic Wasserstein Autoencoder (RPWAEAuth) is an end-to-end model that is easier to train and only uses data from legitimate users, which is more in line with the actual situation. RPWAE uses reconstruction probability for authentication, which solves the problem of difficulty in selecting the reconstruction error threshold, making the error selection more reasonable and having a higher accuracy rate.

[0087] In order to verify the accuracy and reliability of the user continuous authentication method based on the reconstruction probabilistic Wasserstein autoencoder of the present application, three experiments are specifically proposed for testing.

[0088] Experiment 1,As shown in Table 2, the comparison between reconstruction probability and reconstruction error:,It can be seen that under different numbers of unknown users,,whether it is accuracy (Accuracy) or equal error rate (EER), the,performance of reconstruction probability (Prob) in authentication is always due to,the reconstruction error (MSE).

[0089] Table 2

[0090]

[0091] Experiment 2,As shown in Table 3, the comparison of different autoencoders:,It can be seen that the performance of RPWAEAuth,is better than the other two autoencoders (AE, VAE) under different unknown users.

[0092] Table 3

[0093]

[0094] Experiment 3,As shown in Table 4, comparison of different authentication systems:,It can be seen that our proposed system RPWAEAuth,,whether using data from two sensors or data from three sensors, has better,equal error rate and accuracy than other systems.

[0095] Table 4

[0096]

[0097] Embodiment 2

[0098] like Figure 6 As shown, this embodiment is a user continuous authentication system based on a reconstruction probability Wasserstein autoencoder, comprising:

[0099] A data collection module, used to collect behavioral data of the interaction between users and mobile devices;

[0100] A data processing module, used for standardizing and segmenting the collected behavior data to obtain a standard data set;

[0101] The user authentication module is used to sample the standard data set, input the sampled data into the pre-trained reconstruction probability Wasserstein autoencoder for encoding and decoding, and calculate the reconstruction probability;

[0102] The judgment module is used to compare the reconstruction probability with a preset threshold. If the reconstruction probability is greater than the threshold, the user is authenticated as a legitimate user; otherwise, the user is authenticated as an illegal user.

[0103] The calculation formula of the Wasserstein distance of the Wasserstein autoencoder is:

[0104]

[0105] Among them, D WAE (P X ,P G ) is the Wasserstein distance, P X is the model distribution, P G is the target distribution; LogProb(X,G(Z)) is the Log value of the reconstruction probability between the target sample and the reconstructed sample, X is the data set collected by the system, G(Z) is the reconstruction probability Wasserstein autoencoder reconstructs X, and Z is the latent variable of the autoencoder; MMD k (Q Z ,P Z ) is the regularization factor, λ is the hyperparameter, Q Z is a latent variable, P Z is the prior distribution.

[0106] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A user continuous authentication method based on reconstruction probability Wasserstein autoencoder, characterized in that: include: Collect behavioral data about users’ interactions with mobile devices; Standardizing and segmenting the collected behavioral data to obtain a standard data set; Sample the standard data set, input the sampled data into the pre-trained reconstruction probability Wasserstein autoencoder for encoding and decoding, and calculate the reconstruction probability; Compare the reconstruction probability with the preset threshold. If the reconstruction probability is greater than the preset threshold, the authenticated user is a legitimate user; otherwise, the authenticated user is an illegal user. The calculation formula of the Wasserstein distance of the Wasserstein autoencoder is: Among them, D WAE (P X ,P G ) is the Wasserstein distance, P X is the model distribution, P G is the target distribution; LogProb(X,G(Z)) is the Log value of the reconstruction probability between the target sample and the reconstructed sample, X is the data set collected by the system, G(Z) is the reconstruction probability Wasserstein autoencoder reconstructs X, and Z is the latent variable of the autoencoder; MMD k (Q Z ,P Z ) is the regularization factor, λ is the hyperparameter, Q Z is a latent variable, P Z is the prior distribution.

2. The user continuous authentication method according to claim 1, characterized in that: The calculation formula of the reconstruction probability is: ReconProb(X,G(Z))=e LogProb(X,G(Z)) , Among them, x (i) is the sample segmented from X, For x (i) The reconstructed mean of For x (i) The reconstruction standard deviation of the original sample x (i) , in the reconstruction mean With standard deviation ReconProb is the exponential of the Log value e of the reconstruction probability, that is, the logarithm is restored to the reconstruction probability.

3. The user continuous authentication method according to claim 1, characterized in that: The reconstructed probabilistic Wasserstein autoencoder can achieve the model distribution P during training. G and the target distribution P X Minimize the Wasserstein distance between them.

4. The user continuous authentication method according to claim 1, characterized in that: The reconstructed probability Wasserstein autoencoder includes an encoder and a decoder; only legitimate user data is used for training, and the legitimate user data is input into the encoder for encoding to obtain the corresponding latent variable Q Z , the latent variable is decoded by the decoder to output the reconstruction probability.

5. The user continuous authentication method according to claim 1, characterized in that: The encoder includes four convolutional layers and one linear layer, and the decoder includes one linear layer and four deconvolutional layers.

6. The user continuous authentication method according to any one of claims 1 to 5, characterized in that: Collect user behavior data through the accelerometer, gyroscope and magnetometer in the mobile device.

7. A user continuous authentication system based on reconstruction probability Wasserstein autoencoder, characterized in that: include: A data collection module, used to collect behavioral data of the interaction between users and mobile devices; A data processing module, used for standardizing and segmenting the collected behavior data to obtain a standard data set; The user authentication module is used to sample the standard data set, input the sampled data into the pre-trained reconstruction probability Wasserstein autoencoder for encoding and decoding, and calculate the reconstruction probability; The judgment module is used to compare the reconstruction probability with a preset threshold. If the reconstruction probability is greater than the threshold, the user is authenticated as a legitimate user; otherwise, the user is authenticated as an illegal user. The calculation formula of the Wasserstein distance of the Wasserstein autoencoder is: Among them, D WAE (P X ,P G ) is the Wasserstein distance, P X is the model distribution, P G is the target distribution; LogProb(X,G(Z)) is the Log value of the reconstruction probability between the target sample and the reconstructed sample, X is the data set collected by the system, G(Z) is the reconstruction probability Wasserstein autoencoder reconstructs X, and Z is the latent variable of the autoencoder; MMD k (Q Z ,P Z ) is the regularization factor, λ is the hyperparameter, Q Z is a latent variable, P Z is the prior distribution.

8. The user continuous authentication system according to claim 7, characterized in that: The calculation formula of the reconstruction probability is: ReconProb(X,G(Z))=e LogProb(X,G(Z)) , Among them, x (i) is the sample segmented from X, For x (i) The reconstructed mean of For x (i) The reconstruction standard deviation of the original sample x (i) , in the reconstruction mean With standard deviation ReconProb is the exponential of the Log value e of the reconstruction probability, that is, the logarithm is restored to the reconstruction probability.

9. The user continuous authentication method according to claim 7, characterized in that: The reconstructed probabilistic Wasserstein autoencoder can achieve the model distribution P during training. G and the target distribution P X Minimize the Wasserstein distance between them.

10. The user continuous authentication method according to claim 7, characterized in that: The reconstructed probability Wasserstein autoencoder includes an encoder and a decoder; only legitimate user data is used for training, and the legitimate user data is input into the encoder for encoding to obtain the corresponding latent variable Q Z , the latent variable is decoded by a decoder to output a reconstruction probability; the encoder includes four convolutional layers and one linear layer, and the decoder includes one linear layer and four deconvolutional layers.