WO2019192235A1

WO2019192235A1 - User identity authentication method and system based on mobile device

Info

Publication number: WO2019192235A1
Application number: PCT/CN2019/070668
Authority: WO
Inventors: 伍楷舜; 赵猛; 王丹; 邹永攀
Original assignee: 深圳大学
Priority date: 2018-04-04
Filing date: 2019-01-07
Publication date: 2019-10-10
Also published as: WO2019192253A1; CN108537014A; CN108537014B

Abstract

Provided are a user identity authentication method and system based on a mobile device. The user identity authentication method comprises the following steps: step S1, collecting, by means of a mobile device, a sensor signal regarding the progress of the mobile device from being at a pick-up position to being in a stationary state; step S2, filtering the sensor signal; step S3, pre-processing the filtered sensor signal, and extracting a time domain feature and a frequency domain feature in a pick-up event of the mobile device; and step S4, putting the time domain feature and the frequency domain feature into a neural network recognition algorithm for training to recognize and determine user data. According to the present invention, existing mobile devices extracting behavior features of a user and using a neural network recognition algorithm of a machine learning recognition model to recognize the user identity can effectively simplify the process of user identity authentication, and can greatly improve the user experience, without the need to increase any hardware costs. The system is simple and effective, is convenient to use and has good practicability.

Description

User identity authentication method and system based on mobile device

Technical field

The invention relates to a user identity authentication method, in particular to a mobile device based user identity authentication method, and to a user identity authentication system using the mobile device based user identity authentication method.

Background technique

With the continuous advancement of intelligent social processes and the continuous development of emerging smart devices, smart devices are becoming more and more important in daily life, but there are many insecure and inconvenient data privacy protections. Now unlock smart devices. It is roughly divided into the following aspects.

Traditional protection methods are as follows: on smartphones, PIN codes, passwords or patterns and fingerprint recognition are often used; others use techniques such as fingerprint recognition, vein recognition, iris recognition and face recognition.

First of all, the commonly used PIN code, password and pattern protection and fingerprint authentication, after the investigation by the authority, this protection method is quite insecure and inconvenient. First of all, these protection measures are very easy to be discovered by voyeurs, and it is easy Recorded, then the data privacy is quite unsafe. At the same time, the development of intelligence, everyone has a lot of passwords to remember, it is easy to forget, causing a series of troubles. The most inconvenient is that if the hand is dirty, such as touching some oil and water, the skin on the hand is seriously damaged, which seriously affects the smart device authentication process. It is very likely that the verification is unsuccessful and the failure effect is achieved.

Those who have used these smart devices know that the fastest authentication step on smart devices is to first get the smart device (first step) and then put the fingerprint in the authentication office for authentication (step 2). At present, most of the authentication methods are: first get the smart device you want to use (first step), then turn on the power button of the smart device (the second step), and finally put the fingerprint on the verification verification (third step) . From the above verification steps, it is known that the fastest two-step verification is completed, and the three-step verification is often used.

In addition, there are some biometric identification technologies with high recognition accuracy at home and abroad. At the same time, the recognition accuracy is high, and some cost should be paid. Vein recognition and iris recognition are also based on each person's physiological characteristics and then take physiological characteristics signals, and then a series of analysis and processing of the signal, and finally achieve the role of recognition, these two biometric technologies can achieve high recognition Accuracy, but also with some drawbacks, veins in the back of the hand in vein recognition may still change with age and physiology, and the permanent has not been confirmed. Because the collection method is limited by its own characteristics, the product is difficult to be miniaturized and cannot be In mass production on smart devices, at the same time, the collection equipment has special requirements, the design is relatively complicated, and the manufacturing cost is high.

Summary of the invention

The technical problem to be solved by the present invention is to provide a mobile device-based user identity authentication method that takes user behavior as a feature, thereby simplifying the user identity authentication process and improving the user experience, and provides the mobile device-based user identity authentication. Method of user identity authentication system.

In this regard, the present invention provides a mobile device-based user identity authentication method, including the following steps:

Step S1, collecting, by the mobile device, a sensor signal of the mobile device from the picking position to the stationary state;

Step S2, filtering the sensor signal;

Step S3: performing signal preprocessing on the filtered sensor signal, and extracting time domain features and frequency domain features in the picking event of the mobile device;

In step S4, the time domain feature and the frequency domain feature are put into a neural network recognition algorithm for training, and the user data is recognized and judged.

A further improvement of the present invention is that the method further includes a step S5 of placing a training sample whose confidence level reaches a preset confidence threshold into a data set used for training in the machine learning algorithm, and the original one The class recognition model is upgraded to a second class recognition model.

A further improvement of the present invention is that, in the step S5, in the prediction of the correct training samples, the ranking is performed in ascending order, and the first 5% of the training samples are selected as positive samples of the legal users in the second-class recognition model; In the training samples that predict errors, the order is sorted in descending order, and 5% of the training samples are selected as negative samples of illegal users in the second-class recognition model.

A further improvement of the invention is that said step S2 comprises the following substeps:

Step S201, analyzing characteristics of the time domain and the frequency domain of the sensor signal, and obtaining a distribution law of sensor signal energy in a frequency range;

Step S202: Filtering according to the frequency range obtained in step S201 by using a band pass filter.

A further improvement of the invention is that said step S3 comprises the following substeps:

Step S301, performing event detection on the filtered sensor signal;

Step S302, performing signal preprocessing on the obtained event signal of the mobile device, where the signal preprocessing includes framing and windowing processing;

Step S303, acquiring an event signal, and extracting time domain features and frequency domain features of each event signal respectively.

According to a further improvement of the present invention, in the step S301, the filtered sensor signal is windowed, and then the average and standard deviation calculation is performed on each windowed window signal, and then the detection and interception are performed. The event signal of the mobile device.

A further improvement of the present invention is that, in the step S302, the event signal obtained in the step S301 is framed, and then the window is processed again by the Hamming window, wherein the coverage time between the frame and the frame is smaller than each One-half of the length of the framing.

A further improvement of the present invention is that, in the step S303, the time domain feature and the frequency domain feature of each event signal are separately extracted, and then the time domain feature and the frequency domain feature of each event signal are combined into a feature vector; wherein, extracting The time domain characteristics of each event signal include: mean, variance, standard deviation, maximum, minimum, number of zero crossings, difference between maximum and minimum, and mode; extracting frequency domain characteristics of each event signal includes: DC component, mean, variance, standard deviation, slope, kurtosis, mean, variance, standard deviation, slope, kurtosis, and energy.

A further improvement of the present invention is that in the step S4, the time domain feature and the frequency domain feature are put into the Autoencoder neural network recognition algorithm for training, and the three-layer neural system is used in the algorithm model of the Autoencoder neural network recognition algorithm. Network, the number of nodes in the hidden layer is 10, the encoding function used is the satlin function, and the decoding function used is the purelin function. In the training of the Autoencoder neural network recognition algorithm, the maximum error value in the training sample is used as the threshold value. Determine the identified user data.

The present invention also provides a mobile device-based user identity authentication system that employs a mobile device-based user identity authentication method as described above.

Compared with the prior art, the beneficial effects of the present invention are: extracting the user's behavior characteristics by the existing mobile device, and using the neural network recognition algorithm of the machine learning recognition model to identify the user identity, which process actually moves the user. The process of picking up the device is completed, which can effectively simplify the process of user identity authentication, greatly improving the user experience, and not affecting the user's own good use of the mobile device in data privacy protection; and the present invention does not need to add any hardware cost. The system is simple, effective and easy to use, and can accurately identify legitimate users and illegal users, and has good practicability. The invention provides a user identity authentication method based on a novel non-intrusive behavioral activity, which improves the user experience and realizes data privacy protection of mobile devices.

DRAWINGS

1 is a schematic diagram of a workflow of an embodiment of the present invention.

detailed description

The preferred embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

As shown in FIG. 1 , this example provides a mobile device-based user identity authentication method, including the following steps:

Step S2, filtering the sensor signal;

The mobile device described in this example refers to an existing mobile smart device, including: a smart phone, a smart watch, a smart tablet, and the like. Before use, take out the mobile device from the pick-up position. The pick-up position described in this example is defined by default as the user's trouser pocket (which can be defined according to the user's habits), the table and the bag. Take the process of mobile devices: first get the mobile device, then take the mobile device to the more comfortable position that the user can see at normal speed. The normal range is 20cm-40cm.

For different users, the mobile device can be moved from the pick-up position to the position that it can comfortably watch. This process is also different; then, this example first uses the acceleration sensor and the gyro sensor in the mobile device to collect The mobile device picks up the sensor signal from the position to the stationary state, then performs a series of data processing and analysis, and uses the neural network recognition algorithm to train, and obtains a behavior-based authentication method corresponding to the user identity. In the future, the identification of the user's identity can be realized by collecting the sensor signal between the pick-up position and the stationary state each time the mobile device matches the recognized recognition model of the neural network recognition algorithm; The accuracy is related to the threshold set by the recognition model, which can be set and adjusted according to actual needs.

The stationary state in step S1 of this example refers to a state in which the instantaneous data of the acceleration sensor and the gyro sensor is 0; of course, in practical applications, the instantaneous data of the acceleration sensor and the gyro sensor are not true every time. The hold value is 0, and a floating range can be set, that is, the instantaneous data of the acceleration sensor and the gyro sensor defaults to a static state within a preset floating range, and the floating range can be customized according to the actual situation of the user.

As shown in FIG. 1 , the example further includes a step S5, where the training sample whose confidence level reaches a preset confidence threshold is put into the data set used for training in the machine learning algorithm, and the original One type of recognition model is upgraded to a second type of recognition model. The confidence threshold is a preset standard value for determining the confidence level, and the confidence threshold can be adjusted and set according to actual conditions.

The setting of step S5 in this example is necessary because the accuracy of identifying the user's data into a class of recognition algorithm model may not be very high due to an additional small amount of noise. Therefore, this example is based on Autoencoder neural network identification. The confidence evaluation algorithm of the recognition model used in the algorithm puts some training samples with high confidence (such as training samples whose confidence reaches the preset confidence threshold) into the data set used for training, and turns the original One type of recognition model is upgraded to a second-class recognition model, which can effectively improve the accuracy of recognition. The Autoencoder neural network recognition algorithm is a deep neural learning algorithm with a total of three layers and a neural network with 13 nodes per layer.

More preferably, in step S5 of the present example, in predicting the correct training samples, sorting is performed in order from small to large, and the first 5% of the training samples are selected as positive samples of legal users in the second-class recognition model; In the training samples that predict errors, the order is sorted in descending order, and 5% of the training samples are selected as negative samples of illegal users in the second-class recognition model. In this example, the positive and negative samples are placed in the second-class recognition model to further improve the accuracy, for example, by the Decision Tree decision tree algorithm, and the Decision Tree decision tree algorithm uses the CART algorithm, which can be very good. Identifying legitimate users and illegal users; the CART algorithm is a decision tree algorithm.

Step S2 described in this example includes the following sub-steps:

Step S201, analyzing the characteristics of the time domain and the frequency domain of the collected sensor signals by the acceleration sensor and the gyro sensor, and obtaining the distribution law of the sensor signal energy in the frequency range, that is, obtained by time domain analysis and frequency domain analysis. In which frequency range the energy is mainly distributed, the distribution range of the frequency accompanying energy is obtained. In this example, the frequency range is preferably set from 10 Hz to 60 Hz; in short, the acceleration sensor and the gyro sensor are used for the collected sensor signals. Time domain and frequency domain analysis to obtain the frequency of the signal set;

Step S202: Filtering according to the frequency range obtained in step S201 by using a band pass filter, for example, by Butterworth bandpass filtering, the preferred frequency range of the band pass is 10 Hz to 60 Hz, which can effectively remove noise interference. .

Step S3 described in this example includes the following sub-steps:

Step S301, performing event detection on the filtered sensor signal;

Preferably, the filtered signal is used to detect an event using a CFAR algorithm. The CFAR algorithm is a constant false alarm rate algorithm. In the CFAR algorithm, some parameters are needed, such as Winsize (length of the window), step (current window). Distance from the length of the next window), Mu (average of data within the window), Sigma (variance of data within the window), lamda (first function), and lamda2 (second function). According to the characteristics of the user behavior signal, the parameters adopted in this example are: Winsize=256, Step=1, Mu and Sigma are based on known parameters, lamda=2, lamda2=5, and can be effectively detected by using this CFAR algorithm. And intercepting events to the mobile signal.

The Chinese name of the CFAR algorithm is called the constant false alarm rate algorithm (an algorithm in the common language radar). The role of this algorithm is to use this algorithm to detect the start and end positions of the event. The role of Step is that since the signals processed by the time are all discrete, in order to ensure the short-term stability of the signal, this example adds a window, which is the front of the windowsize. When processing the data point of the next window, this example needs Move the current window backwards by a small amount of data to continue processing the data points in this window. This step is the length of the moving window in this example. Mu is the average of the data in this window, and Sigma is the variance of the data in this window.

In the instant search step S301, the CFAR algorithm can be used to detect the filtered sensor signal, as follows: Since the signal collected in this example is a time-varying signal, the processing is not very well processed, and the default is short-time. The signal is a time-invariant signal, and a series of operations can be performed on the signal. In order to achieve a short-time constant signal, a window function is added to a signal, that is, a part of the signal is processed, which is a window. The concept, this example separately calculates the mean and standard deviation of the signal in this window, the formula is as follows:

Where i is the sample point, μ(i) is the average value within the window, σ(i) is the variance within the window, W is the window size, and A(i) and B(i) are calculated as follows:

Where S(k) is the original signal in this window and k is the sample point.

Then this example performs detection events based on the data calculated above:

|S(i)|*|S(i)|>μ(i)+γ ₁ σ(i)

If the signal satisfies the above requirement, this example considers this to be the starting point of the event. If the requirements of the following formula are met, this example is considered to be the end point.

γ ₁ and γ ₂ are constant parameters depending on noise, Winsize is the length of this window, and step is the length of time from the current window to the next window.

Of course, the step S301 in this example is to describe how to implement the event detection by using the CFAR algorithm. In actual applications, other algorithms or methods may be used, as long as event detection can be implemented. As the detection of the event, the focus is on how to obtain the starting point and focus of the event, so that the event can be intercepted effectively. Therefore, there are many methods, and the CFAR algorithm is not limited. For example, in step S301 in this example, The filtered sensor signal is windowed, and then the mean and standard deviation calculations are performed on each windowed window signal, thereby detecting and intercepting the event signal of the mobile device, which is also fully achievable. .

In the step S302, the event signal obtained in the step S301 is framed, and then the window is processed again by the Hamming window, wherein the coverage time between the frame and the frame is less than the duration of each frame. One of the points.

Preferably, in the step S302, according to the above, the signal preprocessing is performed on the obtained mobile signal event, including framing and windowing processing. In this example, each framing of the signal is 5 milliseconds, in order to To avoid moving signal features spanning two frames, this example uses a frame-to-frame coverage of 2 milliseconds to achieve effective feature extraction. The window function added in this example is a hamming window.

In step S303, the time domain feature and the frequency domain feature of each event signal are separately extracted, and then the time domain feature and the frequency domain feature of each event signal are combined into a feature vector; wherein each event signal is extracted. The time domain features include: mean, variance, standard deviation, maximum, minimum, number of zero crossings, difference between maximum and minimum, and mode; the frequency domain characteristics of each event signal are extracted: DC component, graphic Mean, variance, standard deviation, slope, kurtosis, mean, variance, standard deviation, slope, kurtosis, and energy; and then combine a total of 20-dimensional eigenvectors.

That is to say, in this example, after each event is acquired, it is preferable to perform a mathematical operation of the feature vector on the entire segment in each event, and then combine the values together, and the total length is 21. Of course, the more these time domain features and frequency domain features, the more accurate the feature vectors are, and the most important feature vectors include mean and variance.

Step S4 in this example obtains event-related features according to the above, and then puts these features into a machine learning recognition algorithm, such as an Autoencoder neural network recognition algorithm, to obtain user data identification and judgment.

More specifically, in step S4 of this example, the time domain feature and the frequency domain feature are put into the Autoencoder neural network recognition algorithm for training, and the three-layer neural system is used in the algorithm model of the Autoencoder neural network recognition algorithm. Network, the number of nodes in the hidden layer is 10, the encoding function used is the satlin function, and the decoding function used is the purelin function. In the training of the Autoencoder neural network recognition algorithm, the maximum error value in the training sample is used as the threshold value. Judging the identified user data; that is, if the error is smaller than the maximum value of the training sample, it is determined to be a legitimate user, otherwise it is determined to be an illegal user.

The Autoencoder neural network recognition algorithm in this example uses a three-layer neural network model. The purpose of this three-layer neural network model is to reconstruct the input data, ie: input (2.1, 3, 5.5, 2), then through the middle Constantly adjusting the weights makes it possible to re-reconstruct these data points. For example, the last reconstructed data is (2, 3.1, 5.3, 2.1). Reconstructing the data points is actually re-encoding the original data through an encoding function. In this example, a decoding function is used to decode the data. The encoding function and the decoding function are mentioned above.

After the experiment, we can see that they are very close, there is a little error. In this case, according to these errors, the weight between the adjustments is constantly adjusted to make this error the smallest. There are three layers. The first layer is the data input in this example. Constantly adjust the weight between the layers to dynamically reduce the error. Finally, the model outputs the reconstructed data points. Then, in this case, the largest error among these error points is selected as a threshold for this example. This example is larger than this threshold. It is considered to be not a legitimate user. In this case, the value is less than this threshold. It is considered to be its own legitimate user. In the identification and judgment, the L2 parameter can be used as the precision parameter to be obtained. The weight of the L2 parameter is preferably 0.01.

This example also provides a mobile device-based user identity authentication system that employs a mobile device-based user identity authentication method as described above.

In summary, this example provides a neural network recognition algorithm that uses the machine learning recognition model to identify the user's identity by extracting the user's behavior characteristics from the existing mobile device. This process is actually completed when the user picks up the mobile device. The process of user identity authentication can be effectively simplified, the user experience is greatly improved, and the user's own good use of the mobile device during data privacy protection is not affected; and the invention does not need to add any hardware cost, and the system is simple, effective, and convenient to use. It can accurately identify legitimate users and illegal users, and has good practicability.

That is to say, this example provides a new non-intrusive, behavioral activity-based user identity authentication method that enhances the user experience and implements mobile device data privacy protection.

The above is a further detailed description of the present invention in connection with the specific preferred embodiments, and the specific embodiments of the present invention are not limited to the description. It will be apparent to those skilled in the art that the present invention may be practiced without departing from the spirit and scope of the invention.

Claims

A mobile device-based user identity authentication method, comprising the steps of:

Step S1, collecting, by the mobile device, a sensor signal of the mobile device from the picking position to the stationary state;

Step S2, filtering the sensor signal;

Step S3: performing signal preprocessing on the filtered sensor signal, and extracting time domain features and frequency domain features in the picking event of the mobile device;

In step S4, the time domain feature and the frequency domain feature are put into a neural network recognition algorithm for training, and the user data is recognized and judged.
The mobile device-based user identity authentication method according to claim 1, further comprising a step S5 of placing a training sample whose confidence level reaches a preset confidence threshold into the machine learning algorithm for training The data set and upgrade the original one type of recognition model to the second type of recognition model.
The mobile device-based user identity authentication method according to claim 2, wherein in the step S5, in the prediction of the correct training samples, the ranking is performed in ascending order, and the first 5% of the training samples are selected. As a positive sample of the legal user in the second-class recognition model; in the training samples that predict the error, the order is sorted in descending order, and 5% of the training samples are selected as the negative samples of the illegal users in the second-class recognition model.
The mobile device-based user identity authentication method according to any one of claims 1 to 3, wherein the step S2 comprises the following sub-steps:

Step S201, analyzing characteristics of the time domain and the frequency domain of the sensor signal, and obtaining a distribution law of sensor signal energy in a frequency range;

Step S202: Filtering according to the frequency range obtained in step S201 by using a band pass filter.
The mobile device-based user identity authentication method according to any one of claims 1 to 3, wherein the step S3 comprises the following sub-steps:

Step S301, performing event detection on the filtered sensor signal;

Step S302, performing signal preprocessing on the obtained event signal of the mobile device, where the signal preprocessing includes framing and windowing processing;

Step S303, acquiring an event signal, and extracting time domain features and frequency domain features of each event signal respectively.
The mobile device-based user identity authentication method according to claim 5, wherein in the step S301, the filtered sensor signal is first windowed, and then the windowed signal is filtered for each window. Mean and standard deviation calculations are performed to detect and intercept event signals to the mobile device.
The mobile device-based user identity authentication method according to claim 5, wherein in the step S302, the event signal obtained in the step S301 is framed, and then the window is processed again through the Hamming window. The frame duration between frames is less than one-half of the length of each frame.
The mobile device-based user identity authentication method according to claim 5, wherein in step S303, time domain features and frequency domain features of each event signal are separately extracted, and then each event signal is timed. The domain feature and the frequency domain feature form a feature vector; wherein the time domain features of each event signal are extracted: mean, variance, standard deviation, maximum value, minimum value, number of zero crossing points, difference between maximum value and minimum value, and The frequency domain characteristics of each event signal are extracted: DC component, mean, variance, standard deviation, slope, kurtosis, mean, variance, standard deviation, slope, kurtosis, and energy.
The mobile device-based user identity authentication method according to any one of claims 1 to 3, wherein in the step S4, the time domain feature and the frequency domain feature are put into an Autoencoder neural network recognition algorithm. For training, a three-layer neural network is used in the algorithm model of the Autoencoder neural network recognition algorithm. The number of nodes in the hidden layer is 10. The coding function used is a satlin function, and the decoding function used is a purelin function in the Autoencoder nerve. In the training of the network identification algorithm, the maximum value of the error in the training sample is used as a threshold to judge the identified user data.
A mobile device-based user identity authentication system, characterized in that the mobile device-based user identity authentication method according to any one of claims 1 to 9 is employed.