An implicit authentication method for smart device users based on physiological behavior characteristics

Abstract

The present invention discloses an implicit authentication method for a user of a smart device based on physiological behavior characteristics, including: collecting screen pressure data and pressing area data reflecting the physiological characteristics of a user's finger; collecting acceleration data, magnetic field data and gyroscope data reflecting the characteristics of the user's behavioral habits; preprocessing the collected data to generate a certain amount of sample data; building an implicit authentication model based on a twin neural network, and training the implicit authentication model using the sample data; collecting authentication data for multiple times for users of registered smart devices to generate a corresponding reference data set; preprocessing the real-time authentication data input by the user and importing it into the trained implicit authentication model, so that the implicit authentication model authenticates the user's identity based on the user's reference data set. The present invention can complete a non-sensing authentication operation for the user when the user takes the mobile phone.

Description

Technical Field

[0001] The present invention relates to the technical field of implicit authentication, and in particular to a method for implicitly authenticating a user of a smart device based on physiological behavior characteristics. Background Art

[0002] The convenience and increasing functionality of mobile smart devices enable users to collect and store a variety of data, but most of it is highly personal or sensitive information, such as photos, call logs, chat messages, location information, and even confidential business documents. Therefore, protecting mobile smart devices such as smartphones from unauthorized access is crucial to protecting users' personal data and privacy. Accordingly, identity authentication for smart device users is particularly important. A common user authentication method for smartphones is to use a PIN or pattern lock when reactivating the screen. Attacks against this type of access authentication are currently mainly divided into visual-based attacks, such as shoulder surfing attacks, smudge attacks, and sensor-based side channel attacks. In order to defend against the above attacks, implicit authentication is one of the effective methods. Implicit authentication is different from explicit authentication. It does not require users to enter authentication credentials. It uses the user's behavioral characteristics during the interaction process to authenticate the user, or uses the user's habits of using mobile smart devices to authenticate the user, which is usually imperceptible to the user. Implicit authentication can be used as a basic authentication method to replace common password authentication, eliminating the burden of users entering passwords and preventing users from forgetting or losing keys; it can also be used as a supplement to common password authentication, making the final authentication have the effect of two-factor authentication, which is more secure and reliable. Whether users use desktop computers or smart devices to access certain services on the Internet that require identity authentication, users enter login information in the usual way, and the server also uses implicit authentication technology to further confirm the user's identity, which makes user identity authentication more reliable.

[0003] However, in order to obtain the user's behavioral characteristics, the previous implicit authentication technology requires the user to do many additional authentication actions, such as swinging the arm twice, which undoubtedly increases the user's authentication burden and reduces the usability of implicit authentication. The problem to be solved by the present invention is to complete the authentication operation without the user's feeling when the user picks up the mobile phone.

[0004] The invention with patent number CN110472389A discloses an identity authentication system based on touch gestures and its working method, including a gesture perception module and an identity authentication module; the gesture perception module obtains the touch screen and inertial sensor data of the mobile device based on the time series, and extracts the feature vector representing the user's touch screen behavior habits from the acquired time series data; in the identity authentication module, the effectiveness of the feature representing the user's behavior habits is analyzed based on the collected feature vector data, the most effective feature data is selected to build a gesture identity model, and identity authentication is performed based on the gesture identity model. The identity authentication system of this invention can not only obtain the content input when the screen is touched, but also perceive the implicit gesture behavior habits of the user when the screen is touched, which can ensure that even if an illegal user peeks into the gesture, the verification will fail because the user cannot completely imitate the gesture behavior habits of the legitimate user, thereby achieving higher security while maintaining a good user experience. However, this invention or similar implicit authentication inventions only rely on touch gestures, and cannot achieve the purpose of truly imperceptible authentication under the premise of ensuring authentication security. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides an implicit authentication method for a user of a smart device based on physiological behavior characteristics, which can perform a non-perceptible authentication operation on the user when the user picks up the mobile phone.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] The embodiment of the present invention proposes an implicit authentication method for a smart device user based on physiological behavior characteristics, the authentication method comprising:

[0008] When a user touches the touch screen of a smart device, screen pressure data and pressing area data reflecting the physiological characteristics of the user's finger are collected; when a user picks up a smart device, acceleration data, magnetic field data, and gyroscope data reflecting the user's behavioral habits are collected;

[0009] Preprocess the collected data to generate a certain amount of sample data;

[0010] An implicit authentication model is constructed based on the twin neural network, and the implicit authentication model is trained using sample data;

[0011] Multiple authentication data collections are performed for registered smart device users to generate corresponding reference data sets; the authentication data includes acceleration data, magnetic field data, and gyroscope data when the user of the smart device holds the smart device in different postures, as well as screen pressure data and pressing area data when the user touches the touch screen of the smart device;

[0012] The real-time authentication data input by the user is preprocessed and then imported into the trained implicit authentication model, so that the implicit authentication model authenticates the user's identity based on the user's reference data set.

[0013] Optionally, the preprocessing includes: performing time alignment operation on time series data of all parameters using interpolation method; applying Kalman filter and performing noise reduction processing on attitude data in combination with gyroscope data.

[0014] Optionally, the process of performing a time alignment operation on the collected time series data of all parameters using an interpolation method comprises the following steps:

[0015] S11, obtain the timestamp t when the user presses the screen base As the starting point, and the time stamp t′ when the user lifts his finger, calculate t end =t′-(t′-t base )mod10 as the end reference;

[0016] S12, traverse each data (t,y) in the remaining time series data records and find |tt base |Take the record with the smallest value as the respective benchmark t base ;

[0017] S13, traverse all time series data, for each (tt base ) mod 10 = 0. If there is already a value at that position, skip it until t>t end until;

[0018] S14, keep all time series (tt base )mod10=0, delete the rest of the points, and replace the t base Alignment.

[0019] Optionally, the process of applying a Kalman filter to perform noise reduction processing on the attitude data in combination with the gyroscope data includes the following steps:

[0020] S21, using attitude angle data as observation data and gyroscope data as process data, establish the state equation and observation equation of the system:

[0021]

[0022] Z k =HX k

[0023] In the formula, is the estimation of the state quantity at time k based on the state quantity at time k-1, X kis the state vector of the system at time k, A is the state transfer matrix, B is the input control weight matrix, Z k is the observed quantity of the system at time k, and H is the state transfer matrix;

[0024] S22, for the Kalman filter of attitude angle denoising, the discrete system state equation of gyroscope data and the discrete system observation equation of attitude angle data are:

[0025]

[0026] The state vector of the system includes the attitude angle data and the drift value of the gyroscope data. k is the data input by the gyroscope at time k, dt is the time interval between each sampling, Z k is the calculated angle value; in the Kalman filter for attitude angle denoising, H=[1 0],

[0027] S23, performing a priori estimation on the error covariance of the system state quantity according to the following formula to calculate the Kalman gain:

[0028] P k|k-1 =AP k-1 A T +Q

[0029] Where P k|k-1 is the estimated error covariance matrix, P k-1 is the error covariance matrix of the previous state, Q is the covariance of the system process, that is, Simplify Q to Q Angle and Q Gyro is the variance of the attitude angle data and the gyroscope data, which is obtained by measuring the attitude angle data and the gyroscope data when the smart device is placed still and calculating their variance;

[0030] S24, a weighted average is performed between the measurement result and the empirical estimate, and the most likely true value is calculated using the Kalman gain matrix. The calculation method of the Kalman gain matrix is:

[0031] K k =P k|k-1 H T (HP k|k-1 H T +R) -1

[0032] Where K k is the Kalman gain matrix at time k, R is the covariance of the measurement noise;

[0033] S25, by observing the deviation And the weighted average of the state estimation value is used to estimate the current state. The optimal state estimation calculation method is:

[0034]

[0035] In the formula, is the estimated value of the current state, is the attitude angle data output after filtering and noise reduction processing, is the estimated value of the current gyroscope drift;

[0036] S26, updates the covariance in state k, so that the Kalman filter continues to run until the system process ends:

[0037] P k =(IK k H)P k|k-1

[0038] Where I is the unit matrix; R is the measurement noise; the initial value and hyperparameters of the Kalman filter are: .

[0039] Optionally, the sample data is an 11-dimensional time series, which are X-, Y- and Z-axis data of posture angle, linear acceleration, angular acceleration, and pressure data and area data of the user pressing the screen.

[0040] Optionally, the implicit authentication model includes two feature extraction networks with identical structures and shared parameters and a similarity calculation module;

[0041] The feature extraction network includes a one-dimensional convolutional neural network and a pyramid pooling layer connected to each other, the convolutional neural network is used to automatically extract relevant features from sample data in a multi-dimensional time series, and the pyramid pooling layer is used to output a feature sequence of a fixed length;

[0042] The similarity calculation module is used to calculate the similarity of feature sequences output by two feature extraction networks.

[0043] Optionally, the one-dimensional convolutional neural network is composed of two one-dimensional convolutional layers and two Relu activation layers, the input channel of the first one-dimensional convolutional layer is 11, which is the same as the dimension of the sample data; the feature vector processed by the first one-dimensional convolutional layer enters the second one-dimensional convolutional layer after passing through the first Relu activation layer, the input channel of the second one-dimensional convolutional layer is the same as the output channel of the first one-dimensional convolutional layer, and the feature vector processed by the second one-dimensional convolutional layer passes through the second Relu activation layer and enters the pyramid pooling layer.

[0044] Optionally, the number of layers of the pyramid pooling layer is 3;

[0045] The window size Win_size of pyramid pooling is related to the number of layers and the length of the feature vector:

[0046]

[0047] Where Length is the length of the feature vector and level is the number of levels.

[0048] Optionally, the similarity calculation module uses the L2 norm to calculate the similarity of two feature vectors. For two feature vectors Net(w, T1) and Net(w, T2), the similarity coefficient E w (T1,T2) is:

[0049] E w (T1,T2)=||Net(w,T1)-Net(w,T2)||2

[0050] The implicit authentication model uses a contrast loss function to calculate the similarity coefficient E w The calculation result of (T1, T2) is optimized to reduce the similarity coefficient of two samples belonging to the same user and to increase the similarity coefficient of samples not belonging to the same user;

[0051] The contrastive loss function is defined as follows:

[0052]

[0053] Where Y is the label of the training set. When two samples are from the same user, Y = 0. When two samples are from different users, Y = 1.

[0054] When two samples belong to the same user, the loss function is:

[0055]

[0056] That is, the optimization goal is to reduce the similarity coefficient Ew;

[0057] When two samples do not belong to the same user, the loss function is:

[0058] arg w minL(w,(Y,T i ,T j ))=∑{max(0,mE w (T i ,T j ))} 2

[0059] Where m is the interval parameter, which is a hyperparameter. w (T i ,T j)The network will only perform back propagation when it is within this interval; the overall target value of the network minimizes the loss function L(w,(Y,T i ,T j ), when the samples belong to the same user, E w (T i ,T j ) decreases, and when the samples belong to different users, E w (T i ,T j ) increases.

[0060] Optionally, the process of pre-processing the real-time authentication data input by the user and importing it into the trained implicit authentication model so that the implicit authentication model authenticates the user identity based on the reference data set of the user includes the following steps:

[0061] S31, using the implicit authentication model to calculate the similarity between each sample in the reference set and the authentication sample, and obtain a similarity set D;

[0062] S32, each similarity coefficient in the similarity set and the threshold T D Compare, if the value is greater than the threshold, it is considered that the reference sample voted for authentication failure, and if it is less than or equal to the threshold, it is considered that the vote was successful, and the voting set C is obtained;

[0063] S33: Count the votes in voting set C. If the proportion of the votes that are successfully authenticated is greater than the threshold T C If the authentication sample is successful, the authentication will fail.

[0064] The present invention proposes an implicit authentication method for smart device users based on physiological behavioral characteristics. When a user uses a smart device, the method collects data from the smart device sensor, analyzes the user's physiological behavioral characteristics, and compares them with the registered data to authenticate the user. In most cases, the user will take the mobile phone out of his pocket before unlocking the smart device. For this taking action, the physiological differences and behavioral habits between different individuals will cause the behavioral characteristics of the action such as speed and angle, as well as the physiological characteristics such as the pressure and area of ​​the finger on the screen to be different, which becomes the theoretical basis for implicit authentication. The whole process is completely transparent to the user. If the authentication is passed, the user directly unlocks the device. If the authentication fails, the authentication credentials are entered. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of a method for implicitly authenticating a user of a smart device based on physiological behavior characteristics according to an embodiment of the present invention.

[0066] Figure 2 It is a structural diagram of the implicit authentication model of an embodiment of the present invention.

[0067] Figure 3 Schematic diagram of the structure of a convolutional neural network according to an embodiment of the present invention.

[0068] Figure 4 Schematic diagram of the structure of the pyramid pooling layer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The present invention will now be described in further detail with reference to the accompanying drawings.

[0070] It should be noted that the terms such as "upper", "lower", "left", "right", "front", "back", etc. cited in the invention are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments in their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0071] Figure 1 FIG1 is a flow chart of an implicit authentication method for a smart device user based on physiological behavior characteristics according to an embodiment of the present invention. Referring to FIG1 , the authentication method includes:

[0072] When a user touches the touch screen of a smart device, screen pressure data and pressed area data reflecting the physiological characteristics of the user's fingers are collected; when a user picks up a smart device, acceleration data, magnetic field data and gyroscope data reflecting the user's behavioral habits are collected.

[0073] Preprocess the collected data to generate a certain amount of sample data.

[0074] An implicit authentication model is constructed based on the twin neural network, and the implicit authentication model is trained using sample data.

[0075] Multiple authentication data collections are performed for users of registered smart devices to generate corresponding reference data sets; the authentication data includes acceleration data, magnetic field data, and gyroscope data when the user of the smart device holds the smart device in different postures, as well as screen pressure data and pressing area data when the user touches the touch screen of the smart device.

[0076] The real-time authentication data input by the user is preprocessed and then imported into the trained implicit authentication model, so that the implicit authentication model authenticates the user's identity based on the user's reference data set.

[0077] The basic workflow of implicit authentication based on physiological and behavioral characteristics is as follows: Figure 1 As shown in the figure, it is mainly divided into three stages, namely the model training stage, the registration stage of legitimate users after the model deployment, and the implicit authentication stage. The three stages also include two common steps, namely the data acquisition step and the data preprocessing step. The basic tasks of each step are as follows:

[0078] (1) The data acquisition step is mainly responsible for collecting the smart terminal’s attitude sensor data, gyroscope sensor data, and touch screen data when the user picks up the mobile smart terminal, providing a reliable data source for subsequent data processing steps.

[0079] (2) Data preprocessing step: Perform certain preprocessing operations on the collected raw sensor data: Since the sampling frequencies of different types of sensors are different, the interpolation method is used to perform time alignment operations on the data of multiple sensors; and the sensor itself has measurement errors. External magnetic field disturbances will also cause noise in the attitude data. The Kalman filter can be applied in combination with the gyroscope data to reduce noise.

[0080] In addition to the above two common steps, there is also an implicit authentication model training step based on the twin neural network in the training phase of the authentication model. This step uses the pre-acquired data to train the twin neural network, and stores the relevant network parameters in an appropriate form. In the implicit authentication phase, there is also an authentication result response step, which inputs the pre-processed data of the user authentication into the implicit authentication model, and the model calculates and gives the authentication result. If the user passes the authentication, he is allowed to continue to perform subsequent access operations, otherwise explicit authentication is performed.

[0081] 1. Training phase

[0082] 1.1 Data Collection

[0083] In the training phase, in order to train the authentication model, it is first necessary to obtain sensor data that reflects the user's physiological behavior characteristics. The screen pressure data and pressed area data reflect the physiological characteristics of the user's fingers; the acceleration data, magnetic field data, and gyroscope data reflect the change of the direction of the smart device over time during the movement of the smart device, reflecting the movement trajectory of the device in space, and to a certain extent, reflecting the user's behavioral habits.

[0084] 1.2 Data Preprocessing

[0085] In order to obtain reliable data for implicit authentication of users, the acquired sensor data needs to be preprocessed. The interpolation method is used to align the time of data from multiple sensors with different sampling frequencies; due to the measurement error of the sensor itself, the influence of the external magnetic field, etc., the attitude data is noisy. The Kalman filter is applied to reduce the noise of the attitude angle data of the smart device in combination with the gyroscope data.

[0086] (1) Time alignment of multi-sensor data

[0087] Since the sensor sampling frequencies of mobile smart terminals are different and the collected time series data are not aligned, each time series data needs to be preprocessed before further calculation.

[0088] After investigation, it was found that the sampling rates of different mobile smart terminal sensors are different. The time series data are aligned to 1000Hz using linear interpolation. Linear interpolation is an interpolation method for one-dimensional data. It estimates the value based on the two data points on the left and right of the point to be interpolated in the one-dimensional data sequence. For the known points (t1, y2) and (t2, y2), the interpolation point (t0, y0) is:

[0089]

[0090] Where t is the timestamp of the time series, in milliseconds. The process of time series data alignment is:

[0091] 1) Get the timestamp t when the user presses the screen base As the starting point, and the time stamp t′ when the user lifts his finger, t end =t′-(t′-t base )mod10 as the end reference;

[0092] 2) Traverse each data (t,y) in the remaining time series data records and find |tt base |Take the record with the smallest value as the respective benchmark t base ;

[0093] 3) Traverse all time series data, for each (tt base ) mod 10 = 0. If there is already a value at that position, skip it until t>t end until;

[0094] 4) Keep all time series (tt base )mod10=0, delete the rest of the points, and replace the t base Alignment.

[0095] (2) Attitude angle data denoising

[0096] The attitude data obtained by directly reading the smart device sensor has a large amount of noise. The noise mainly comes from: the acceleration sensor and magnetic field sensor are affected by external vibrations and the sensor itself has high-frequency noise. The attitude angle data calculated from the acceleration data and magnetic field data has a small drift, but it is not accurate enough in a short time; while the gyroscope is more accurate in a short time, but has the characteristic of zero drift. If the gyroscope data is simply integrated to estimate the attitude angle data, a large cumulative error will be generated. Therefore, the collected attitude angle data and gyroscope data are combined, and the Kalman filter is used for noise reduction to obtain accurate attitude angle data.

[0097] The Kalman filter is a recursive filter proposed by Kalman for time-varying linear systems. The filter uses the observed value of the system as input data and the estimated value of the system state as the output data of the filter, and uses the statistical characteristics of the system noise and observation noise for optimal estimation. The attitude angle and gyroscope data both have x-axis, y-axis and z-axis data. The following is an example of one of the axes. Using the attitude angle data as the observation data and the gyroscope data as the process data, the state equation and observation equation of the system can be established:

[0098]

[0099] Z k =HX k (3)

[0100] In the formula, is the estimation of the state quantity at time k based on the state quantity at time k-1, X k is the state vector of the system at time k, A is the state transfer matrix, B is the input control weight matrix, Z k is the observed value of the system at time k, and H is the state transfer matrix.

[0101] For the Kalman filter of attitude angle denoising, the discrete system state equation of gyroscope data and the discrete system observation equation of attitude angle data are as follows:

[0102]

[0103] The state vector of the system includes the attitude angle data and the drift value of the gyroscope data. k is the data input by the gyroscope at time k, dt is the time interval between each sampling, Z k is the calculated angle value. Therefore, in the Kalman filter for attitude angle denoising, H=[1 0],

[0104] In the Kalman filter, it is necessary to make a priori estimation of the error covariance of the system state quantity to calculate the Kalman gain, which is calculated as follows:

[0105] P k|k-1 =AP k-1 A T +Q (5)

[0106] Where P k|k-1 is the estimated error covariance matrix, P k-1 is the error covariance matrix of the previous state, Q is the covariance of the system process, that is, Since the gyroscope drift noise and attitude angle noise are independent of each other, Q can be simplified to Where Q Angle and Q Gyr o is the variance of the attitude angle data and the gyroscope data, which can be obtained by measuring the attitude angle data and the gyroscope data and calculating their variance when the smart terminal is placed still.

[0107] The Kalman filter performs a weighted average between the measurement results and the empirical estimate, and uses the Kalman gain matrix to calculate the most likely true value. The calculation method of the Kalman gain matrix is:

[0108] K k =P k|k-1 H T (HP k|k-1 H T +R) -1 (6)

[0109] Where K k is the Kalman gain matrix at time k, and R is the covariance of the measurement noise.

[0110] The Kalman filter uses the observation deviation And the weighted average of the state estimation value is used to estimate the current state. The optimal state estimation calculation method is:

[0111]

[0112] In the formula, is the estimated value of the current state, is the attitude angle data output after filtering and noise reduction processing, is the estimated value of the current state gyroscope drift value. After estimating the optimal state, in order to keep the Kalman filter running until the system process ends, it is also necessary to update the covariance in the k state:

[0113] P k =(IK k H)P k|k-1 (8)

[0114] Where I is the unit matrix. The value of the measurement noise R will affect the filtering effect and response speed: the smaller the value of R, the faster the response and the faster the convergence, but the noise suppression ability will be affected. According to experience, the initial value and hyperparameters of the Kalman filter are set as: .

[0115] 1.3 Training and Authentication Model

[0116] After obtaining the posture angle, linear acceleration and angular acceleration time series data that reflect the user's behavioral characteristics, as well as the screen pressure data and screen contact area data that reflect the user's physiological characteristics, the authentication model needs to be trained to obtain the model parameters.

[0117] For time series data, a twin neural network is used to calculate the similarity between two samples. The following describes the method of similarity calculation from the perspective of network structure, loss function, and network training.

[0118] (1) Twin Neural Network Structure

[0119] The twin neural network consists of two feature extraction networks with the same structure and shared parameters and a similarity calculation module, such as Figure 2 As shown, the feature extraction network consists of a one-dimensional convolutional neural network and a pyramid pooling layer. The convolutional neural network can automatically extract features from multi-dimensional time series. The pyramid pooling layer can output a feature sequence of fixed length. The similarity calculation module calculates the similarity of two output feature sequences.

[0120] For each input sample T, it is an 11-dimensional time series, which includes the X, Y and Z axis data of the posture angle, linear acceleration, angular acceleration, and the pressure and face data of the user pressing the screen. For multi-dimensional time series, a one-dimensional convolutional neural network is used to extract feature values. The specific structure of the convolutional neural network is as follows: Figure 3 As shown in the figure, since the structures and parameters of the two sub-networks of the twin neural network are exactly the same, only the structure of one of the sub-networks is described here.

[0121] The convolutional neural network consists of two one-dimensional convolutional layers and two Relu activation layers. The input channel of the first one-dimensional convolutional layer is 11, which is the same as that of sample T. The processed feature vector passes through the Relu activation layer and enters the second one-dimensional convolutional layer. The input channel is the same as the output channel of the first convolutional layer. Then the feature vector passes through the second Relu activation layer and enters the pyramid pooling layer.

[0122] The function of the pyramid pooling layer is to reduce the features and parameters. At the same time, since the lengths of the samples T are different, the lengths of the generated feature vectors are also different. The pyramid pooling layer can generate feature vectors with fixed lengths, which is convenient for the subsequent similarity calculation module to calculate the feature vectors. In the present invention, the number of pyramid pooling layers is 3, and the structure is as follows: Figure 4 shown.

[0123] The pyramid pooling layer uses the maximum value pooling method to extract features from the feature vector input by the convolutional neural network. Unlike the traditional pooling layer, the window size Win_size of the pyramid pooling is related to the number of layers and the length of the feature vector:

[0124]

[0125] Where Length is the length of the feature vector and level is the number of layers. The pyramid pooling layer extracts the maximum value in the window and combines it into a feature vector of fixed length.

[0126] The similarity calculation module is used to calculate the similarity coefficient of the two feature vectors output by the twin neural network. The L2 norm is used to calculate the similarity of the two feature vectors. For the two feature vectors Net(w,T1) and Net(w,T2), the similarity coefficient E w (T1,T2) is:

[0127] E w (T1,T2)=||Net(w,T1)-Net(w,T2)||2 (10)

[0128] (2) Loss Function

[0129] Use the contrast loss function to calculate the similarity coefficient E w The calculation result of (T1, T2) is optimized. The contrast loss function can reduce the similarity coefficient of two samples belonging to the same user and expand the similarity coefficient of samples that do not belong to the same user. The loss function is defined as follows:

[0130]

[0131] Where Y is the label of the training set. When two samples are samples of the same user, Y = 0, and when two samples are samples of different users, Y = 1. When two samples belong to the same user, the loss function is:

[0132]

[0133] That is, the optimization goal is to reduce the similarity coefficient Ew. When two samples do not belong to the same user, the loss function is:

[0134] argw minL(w,(Y,T i ,T j ))=∑{max(0,mE w (T i ,T j ))} 2 (13)

[0135] Where m is the interval parameter, which is a hyperparameter. w (T i ,T j ) The network performs back propagation only when it is within this interval. According to experience, m is set to 2 in the present invention. The overall target value of the network minimizes the loss function L(w,(Y,T i ,T j ), when the samples belong to the same user, E w (T i ,T j ) decreases, and when the samples belong to different users, E w (T i ,T j ) increases.

[0136] (3) Twin Neural Network Training

[0137] Data needs to be collected in advance to train the twin neural network. Considering that in actual application scenarios, users will take out their mobile phones in both standing and sitting postures, authentication data in both postures are collected for each volunteer. Each volunteer collects a certain amount of data in both standing and sitting postures according to their own habits. In this process, the sensor records each authentication action of each volunteer, and preprocesses the sensor data as described in 1.2 to obtain verification samples for each volunteer. For authentication samples in each posture belonging to the same volunteer, pairs are extracted and labeled with 0. In order to ensure that the number of positive and negative examples in the data set is the same, only the same number of sample pairs of different users and different postures of the same user are extracted and labeled with 1.

[0138] 2. Registration stage

[0139] In the registration phase, legitimate users need to be registered to obtain their reference data sets. After multiple acquisitions of legitimate user authentication data, preprocessing such as multi-sensor time alignment and noise reduction is performed, and the preprocessed data is saved in the implicit authentication model for storage. When the user performs authentication operations, it is used as a reference set to compare and authenticate with the authentication data.

[0140] 3. Certification stage

[0141] In the user authentication stage, firstly, the user authentication data is collected as described in 1.1 above, and the collected authentication data is preprocessed as described in 1.2 to obtain the user authentication sample. And it is also necessary to obtain 10 samples of each legitimate user in advance to create a reference set. The user authentication sample is authenticated using an implicit authentication method based on voting. The specific process of implicit authentication is as follows:

[0142] 1) Use the authentication model to calculate the similarity between each sample in the reference set and the authentication sample to obtain the similarity set D.

[0143] 2) For each similarity coefficient in the similarity set and the threshold T D By comparison, if the number is greater than the threshold, it is considered that the reference sample voted for authentication failure, and if it is less than or equal to the threshold, it is considered that the vote was successful, and the voting set C is obtained.

[0144] 3) Count the votes in voting set C. If the proportion of successfully authenticated votes to the total number of votes is greater than the threshold T C If the authentication sample is successfully authenticated, otherwise the authentication fails

[0145] 4. Example verification

[0146] This experiment uses Samsung Galaxy s20 smart terminal, and recruits 20 volunteers to collect 50 groups of authentication samples in standing and sitting postures. The samples are divided into training set and test set at a ratio of 4:1. The training set trains the implicit authentication model, and the test set tests the performance of the implicit authentication model. First, the samples are preprocessed, such as sensor data denoising, and similarity calculation based on twin neural network is performed, and finally implicit authentication is performed. The experimental results show that the accuracy of the implicit authentication system is 95.84%, the false rejection rate is 13.57%, and the false acceptance rate is 3.56%.

[0147] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A method for implicit authentication of smart device users based on physiological behavior characteristics, characterized in that: The authentication method comprises: When a user touches the touch screen of a smart device, screen pressure data and pressing area data reflecting the physiological characteristics of the user's finger are collected; when a user picks up a smart device, acceleration data, magnetic field data, and gyroscope data reflecting the user's behavioral habits are collected; Preprocess the collected data to generate a certain amount of sample data; An implicit authentication model is constructed based on the twin neural network, and the implicit authentication model is trained using sample data; Multiple authentication data collections are performed for registered smart device users to generate corresponding reference data sets; the authentication data includes acceleration data, magnetic field data, and gyroscope data when the user of the smart device holds the smart device in different postures, as well as screen pressure data and pressing area data when the user touches the touch screen of the smart device; After preprocessing the real-time authentication data input by the user, the data is imported into the trained implicit authentication model, so that the implicit authentication model authenticates the user's identity based on the user's reference data set; The preprocessing includes: using interpolation method to perform time alignment operation on time series data of all parameters; applying Kalman filter to perform noise reduction processing on attitude data in combination with gyroscope data; The process of performing time alignment operation on the time series data of all parameters collected by interpolation method includes the following steps: S11, obtain the timestamp t when the user presses the screen base As the starting point, and the time stamp t′ when the user lifts his finger, calculate t end =t′-(t′-t base )mod10 as the end reference; S12, traverse each data (t,y) in the remaining time series data records and find |tt base |Take the record with the smallest value as the respective benchmark t base ; S13, traverse all time series data, for each (tt base ) mod 10 = 0. If there is already a value at that position, skip it until t>t end until; S14, keep all time series (tt base )mod10=0, delete the rest of the points, and replace the t base Alignment.

2. According to claim 1, the method for implicitly authenticating a user of a smart device based on physiological behavior characteristics is characterized in that: The process of applying the Kalman filter to denoise the attitude data in combination with the gyroscope data includes the following steps: S21, using attitude angle data as observation data and gyroscope data as process data, establish the state equation and observation equation of the system: Z k =HX k In the formula, is the estimation of the state quantity at time k based on the state quantity at time k-1, X k is the state vector of the system at time k, A is the state transfer matrix, B is the input control weight matrix, Z k is the observed quantity of the system at time k, and H is the state transfer matrix; S22, for the Kalman filter of attitude angle denoising, the discrete system state equation of gyroscope data and the discrete system observation equation of attitude angle data are: The state vector of the system includes the attitude angle data and the drift value of the gyroscope data. k is the data input by the gyroscope at time k, dt is the time interval between each sampling, Z k is the calculated angle value; in the Kalman filter for attitude angle denoising, H = [1 0], S23, performing a priori estimation on the error covariance of the system state quantity according to the following formula to calculate the Kalman gain: P k|k-1 =AP k-1 From T +Q Where P k|k-1 is the estimated error covariance matrix, P k-1 is the error covariance matrix of the previous state, Q is the covariance of the system process, that is, Simplify Q to Q Angle and Q Gyro is the variance of the attitude angle data and the gyroscope data, which is obtained by measuring the attitude angle data and the gyroscope data when the smart device is placed still and calculating their variance; S24, a weighted average is performed between the measurement result and the empirical estimate, and the most likely true value is calculated using the Kalman gain matrix. The calculation method of the Kalman gain matrix is: K k =P k|k-1 H T (HP k|k-1 H T +R) -1 Where K k is the Kalman gain matrix at time k, R is the covariance of the measurement noise; S25, by observing the deviation And the weighted average of the state estimation value is used to estimate the current state. The optimal state estimation calculation method is: In the formula, is the estimated value of the current state, is the attitude angle data output after filtering and noise reduction. is the estimated value of the current gyroscope drift; S26, updates the covariance in state k, so that the Kalman filter continues to run until the system process ends: P k =(I-K k H)P k|k-1 Where I is the unit matrix; R is the measurement noise; the initial value and hyperparameters of the Kalman filter are: R=0.

5.

3. According to claim 1, the method for implicit authentication of smart device users based on physiological behavior characteristics is characterized in that: The sample data is an 11-dimensional time series, including the X-, Y-, and Z-axis data of the posture angle, linear acceleration, angular acceleration, and the pressure data and area data of the user pressing the screen.

4. According to claim 1, the method for implicit authentication of smart device users based on physiological behavior characteristics is characterized in that: The implicit authentication model includes two feature extraction networks with the same structure and shared parameters and a similarity calculation module; The feature extraction network includes a one-dimensional convolutional neural network and a pyramid pooling layer connected to each other, the convolutional neural network is used to automatically extract relevant features from sample data in a multi-dimensional time series, and the pyramid pooling layer is used to output a feature sequence of a fixed length; The similarity calculation module is used to calculate the similarity of feature sequences output by two feature extraction networks.

5. According to claim 4, the method for implicit authentication of smart device users based on physiological behavior characteristics is characterized in that: The one-dimensional convolutional neural network consists of two one-dimensional convolutional layers and two Relu activation layers. The input channel of the first one-dimensional convolutional layer is 11, which is the same as the dimension of the sample data. The feature vector processed by the first one-dimensional convolutional layer enters the second one-dimensional convolutional layer after passing through the first Relu activation layer. The input channel of the second one-dimensional convolutional layer is the same as the output channel of the first one-dimensional convolutional layer. The feature vector processed by the second one-dimensional convolutional layer passes through the second Relu activation layer and enters the pyramid pooling layer.

6. The method for implicitly authenticating a user of a smart device based on physiological behavior characteristics according to claim 4 is characterized in that: The number of layers of the pyramid pooling layer is 3; The window size Win_size of pyramid pooling is related to the number of layers and the length of the feature vector: Where Length is the length of the feature vector and level is the number of levels.

7. The method for implicitly authenticating a user of a smart device based on physiological behavior characteristics according to claim 4, characterized in that: The similarity calculation module uses the L2 norm to calculate the similarity of two feature vectors. For two feature vectors Net(w, T1) and Net(w, T2), the similarity coefficient E w (T1,T2) is: E w (T1,T2)=||Net(w,T1)-Net(w,T2)||2 The implicit authentication model uses a contrast loss function to calculate the similarity coefficient E w The calculation result of (T1, T2) is optimized to reduce the similarity coefficient of two samples belonging to the same user and to increase the similarity coefficient of samples not belonging to the same user; The contrastive loss function is defined as follows: Where Y is the label of the training set. When two samples are from the same user, Y = 0. When two samples are from different users, Y = 1. When two samples belong to the same user, the loss function is: That is, the optimization goal is to reduce the similarity coefficient Ew; When two samples do not belong to the same user, the loss function is: arg w minL(w,(Y,T i ,T j ))=∑{max(0,m-E w (T i ,T j ))} 2 Where m is the interval parameter, which is a hyperparameter. w (T i ,T j )The network will only perform back propagation when it is within this interval; the overall target value of the network minimizes the loss function L(w,(Y,T i ,T j ), when the samples belong to the same user, E w (T i ,T j ) decreases, and when the samples belong to different users, E w (T i ,T j ) increases.

8. The method for implicitly authenticating a user of a smart device based on physiological behavior characteristics according to claim 1, characterized in that: The process of preprocessing the real-time authentication data input by the user and importing it into the trained implicit authentication model so that the implicit authentication model authenticates the user identity based on the user's reference data set includes the following steps: S31, using the implicit authentication model to calculate the similarity between each sample in the reference set and the authentication sample, and obtain a similarity set D; S32, each similarity coefficient in the similarity set and the threshold T D Compare, if the value is greater than the threshold, it is considered that the reference sample voted for authentication failure, and if it is less than or equal to the threshold, it is considered that the vote was successful, and the voting set C is obtained; S33: Count the votes in voting set C. If the proportion of the votes that are successfully authenticated is greater than the threshold T C If the authentication sample is successful, the authentication will fail.

Images