An identity authentication method and apparatus for mobile terminals

Abstract

This invention provides an identity authentication method and apparatus for mobile terminals, relating to the field of identity recognition technology. The method includes the following steps: detecting a user's typing behavior on a password input interface; verifying the user's input password to obtain a password verification result; activating sensor monitoring on the password input interface to collect sensor data corresponding to the user's typing behavior; inputting the sensor data into a typing behavior prediction model to obtain an authentication distance output by the model; and performing user authentication based on the password verification result and the authentication distance output by the model. This method not only improves security but also enables seamless dual authentication in the background while the user is entering their password, and allows for different risk control measures to be taken based on the dual authentication results, thus enhancing the user experience.

Description

Technical Field

[0001] This invention relates to the field of identity recognition technology, and in particular to an identity authentication method and apparatus for mobile terminals. Background Technology

[0002] Biometric features such as facial recognition, fingerprints, iris scans, and voiceprints are collectively known as hard biometrics. Currently, regulations governing hard biometrics are becoming increasingly stringent, and the leakage of hard biometric data can have extremely long-term, even lifelong, consequences for users. Hard biometrics require a high level of authorization, generally demanding user cooperation for detection to be effective, resulting in a poor user experience.

[0003] Therefore, most users still prefer to use the traditional password login method. However, most password login interfaces on mobile devices do not have a seamless authentication method to accompany them. If the password is leaked or someone else operates the account on your behalf, the logged-in account is actually being operated by someone other than the actual controller, which poses a very high risk.

[0004] How to maximize user privacy during the identity authentication process, so that data breaches will not cause long-term damage, and how to make the authorization process universally applicable without requiring high-level permissions are all considerations for future identity authentication. Summary of the Invention

[0005] In view of this, the present invention provides an identity authentication method and device for mobile terminals, which adopts a dual authentication method of password verification and user operation behavior verification. This method can not only improve security, but also complete dual identity authentication in the background without the user's awareness when entering the password, and can take different risk control measures for the dual identity authentication results, thereby improving the user experience.

[0006] To achieve the above objectives, the present invention provides an identity authentication method for mobile terminals, comprising the following steps:

[0007] The system detects the user's typing behavior on the password input interface, verifies the password entered by the user, and obtains the password verification result.

[0008] Sensor monitoring is enabled on the password input interface to collect sensor data corresponding to the user's typing behavior on the password input interface;

[0009] The sensor data is input into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model;

[0010] User authentication is performed based on the password verification result and the authentication distance output by the model:

[0011] If the password verification passes and the authentication distance output by the model passes the threshold comparison, then the user's identity authentication is deemed successful.

[0012] If the password verification passes, but the authentication distance output by the model fails to compare with the threshold, it is determined that there may be an unauthorized user login, and the first risk control measure is activated to control the user or biometric authentication is activated.

[0013] If the password verification fails, and the authentication distance output by the model passes the threshold comparison, it is determined that the user may have entered the wrong password, and the number of password input errors is increased.

[0014] If password verification fails, and the authentication distance output by the model fails to compare with the threshold, it is determined to be a high-risk unauthorized user login. The number of password input errors will be reduced, a second risk control measure will be activated, or biometric authentication will be activated.

[0015] Optionally, the collection of sensor data corresponding to the user's typing behavior on the password input interface specifically includes the following steps:

[0016] Determine if the mobile terminal is currently in a password input interface;

[0017] If so, the sensors are activated for listening, and the user's input behavior data is collected using the sensors of the mobile terminal; wherein, the sensors include a linear accelerometer and a gyroscope installed in the mobile terminal;

[0018] The input behavior data is divided into small time windows, and the long-term data is divided into several small time window samples. These small time window samples are used as training samples or samples to be predicted for the typing behavior prediction model. Furthermore, each small time window sample is labeled with the user identity to which it belongs.

[0019] Optionally, the typing behavior prediction model is trained based on template samples from password input users and template samples from other users; specifically, the typing behavior prediction model is trained through the following steps:

[0020] Based on the template samples of this user and the template samples of other users, construct training sample pairs;

[0021] The training sample pairs are used as input data for training, and deep learning is used to train the model to obtain the typing behavior prediction model used to generate the identity recognition result of the sensor data.

[0022] Optionally, the typing behavior prediction model is trained based on features extracted from template samples of password input users and template samples of other users. Specifically, the typing behavior prediction model is trained through the following steps:

[0023] Based on the template samples of this user and the template samples of other users, construct training sample pairs;

[0024] Extract the training features from the training sample pairs;

[0025] The training feature pairs are used as input data for training, and deep learning is employed to train the model to obtain the typing behavior prediction model used to generate the identity recognition result of the sensor data.

[0026] Optionally, the training features include statistical features, local features, signal features, frequency domain features, and cross features.

[0027] Optionally, after the step of constructing training sample pairs based on the template samples of the user and other users, the method further includes the following steps:

[0028] The training sample pairs are preprocessed; wherein the preprocessing methods include smoothing filtering, median filtering, averaging filtering, and Kalman filtering.

[0029] Optionally, in the training phase of the typing behavior prediction model, the training samples of the typing behavior prediction model include template samples, positive samples, and negative samples. The template samples, positive samples, and negative samples are obtained by collecting sensor data from each user at the password input interface. The authentication distance refers to the mean square distance between the template sample and the positive sample and the negative sample, respectively, to learn the behavioral differences between the password input habits of the user and the password input habits of unauthorized users. In the prediction phase of the typing behavior prediction model, the authentication distance refers to the mean square distance between the sample to be judged and the template sample. The authentication distance is compared with a threshold, specifically including the following steps:

[0030] The sensor data of the current user is input into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model; the current user is either the user corresponding to the training sample in the typing behavior prediction model or a new user that has not been trained in the typing behavior prediction model.

[0031] Determine the relationship between the authentication distance and the preset distance threshold;

[0032] If the authentication distance is greater than a preset distance threshold, the identity authentication result is that the comparison between the authentication distance and the threshold has failed.

[0033] If the authentication distance is less than a preset distance threshold, the identity authentication result is that the authentication distance passes the threshold comparison.

[0034] The present invention also provides an identity authentication device for a mobile terminal, comprising the following steps:

[0035] The first verification module is used to detect the user's typing behavior on the password input interface, verify the password entered by the user, and obtain the password verification result;

[0036] The second verification module is used to enable sensor monitoring at the password input interface, collect sensor data corresponding to the user's typing behavior at the password input interface, and input the sensor data into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model.

[0037] The data analysis module is used to perform user authentication based on the password verification result and the authentication distance output by the model.

[0038] If the password verification passes and the authentication distance output by the model passes the threshold comparison, then the user's identity authentication is deemed successful.

[0039] If the password verification passes, but the authentication distance output by the model fails to compare with the threshold, it is determined that there may be an unauthorized user login, and the first risk control measure is activated to control the user or biometric authentication is activated.

[0040] If the password verification fails, and the authentication distance output by the model passes the threshold comparison, it is determined that the user may have entered the wrong password, and the number of password input errors is increased.

[0041] If password verification fails, and the authentication distance output by the model fails to compare with the threshold, it is determined to be a high-risk unauthorized user login. The number of password input errors will be reduced, a second risk control measure will be activated, or biometric authentication will be activated.

[0042] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the authentication method for a mobile terminal as described above.

[0043] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the authentication method applied to a mobile terminal as described above.

[0044] This invention employs a dual authentication method combining password verification and user action behavior verification. This method not only improves security but also enables seamless dual authentication in the background when the user enters their password. Furthermore, it allows for different risk control measures to be taken based on the results of the dual authentication, thereby enhancing the user experience.

[0045] As can be seen from the above, the identity authentication method and apparatus for mobile terminals provided by this invention complete a series of verifications seamlessly in the background of the mobile terminal when the user uses the mobile terminal, for example, when using an application (App) installed on the mobile terminal, without requiring the interruption of App operation for dedicated biometric authentication like face or fingerprint authentication. Using sensor-collected data for identity authentication is more privacy-friendly, serving both verification and minimizing harm to the user in the event of leakage. Furthermore, sensor data is time-sensitive; after a period of time, it automatically becomes invalid due to changes in input habits or other objective factors, improving the accuracy of identity authentication. The model built based on sensor data is universal and also meets the requirements of accuracy. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0047] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this invention should have the ordinary meaning understood by those skilled in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0048] As a preferred embodiment of the present invention, the present invention provides an identity authentication method for a mobile terminal, comprising the following steps:

[0049] The system detects the user's typing behavior on the password input interface, verifies the password entered by the user, and obtains the password verification result.

[0050] Sensor monitoring is enabled on the password input interface to collect sensor data corresponding to the user's typing behavior on the password input interface;

[0051] The sensor data is input into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model;

[0052] User authentication is performed based on the password verification result and the authentication distance output by the model:

[0053] If the password verification passes and the authentication distance output by the model passes the threshold comparison, then the user's identity authentication is deemed successful.

[0054] If the password verification passes, but the authentication distance output by the model fails to compare with the threshold, it is determined that there may be an unauthorized user login, and the first risk control measure is activated to control the user or biometric authentication is activated.

[0055] If the password verification fails, and the authentication distance output by the model passes the threshold comparison, it is determined that the user may have entered the wrong password, and the number of password input errors is increased.

[0056] If password verification fails, and the authentication distance output by the model fails to compare with the threshold, it is determined to be a high-risk unauthorized user login. The number of password input errors will be reduced, a second risk control measure will be activated, or biometric authentication will be activated.

[0057] Wherein, the risk control level of the first risk control method is lower than that of the second risk control method. For example, the first risk control method uses any one of the following: telephone or SMS notification to the user, SMS verification code verification, etc., while the second risk control method uses any one of the following: bank card authentication, ID card authentication, facial liveness detection, etc.

[0058] The password of this invention can be a fixed password or a dynamic password. In a preferred embodiment, the password is a combination of a first password and a second password. The first password consists of one or more numbers, letters, and symbols, while the second password is formed by a specific trajectory. This specific trajectory is formed by connecting lines from user touch operations (e.g., inputting numbers, letters, symbols, etc.) or by user swiping operations. Verification of the user-input password includes verification of both the first and second passwords. Enabling sensor monitoring on the password input interface means triggering sensor monitoring after the first password is input. The sensor data corresponding to the typing behavior refers to the sensor data corresponding to the user's input of the second password. This allows for accurate capture of the user's password input behavior, avoiding interference from other operations. Furthermore, the combination of this sensor data and the specific trajectory greatly improves the robustness of the sensor data.

[0059] This mobile terminal authentication method and device seamlessly performs a series of verifications in the background of the mobile terminal when the user uses it, such as using an app installed on the mobile terminal, without requiring the interruption of app operation for biometric authentication like facial recognition or fingerprint authentication. Using sensor-collected data for authentication is more privacy-friendly, serving both verification and minimizing harm to users in the event of leakage. Furthermore, sensor data is time-sensitive; it automatically becomes invalid after a period of time due to changes in input habits or other objective factors, improving the accuracy of authentication. The model built based on sensor data is both universal and meets accuracy requirements.

[0060] It is understandable that mobile terminals include, but are not limited to, mobile phones, laptops, tablets, and wearable devices. Linear accelerometers and gyroscopes are used to collect data such as changes in movement acceleration and rotational angular velocity when a user uses an app on the mobile terminal. Before data collection, the sampling frequency of each sensor needs to be set, such as 50Hz or 100Hz. When the user on the mobile terminal engages in input, the sensors are activated to listen and collect sensor data from the user's typing.

[0061] Since there is a lack of template data for the user in the early stages, sensor data collected from the same session that has already passed facial and fingerprint verification can be stored in a database as a sample template for the authenticated user. The collection of sample templates is for subsequent judgment of unknown samples, serving as benchmark data for authentication comparison. Simultaneously, in the early data collection phase, other biometric authentication methods such as fingerprints and facial recognition can also be combined to label the user's data.

[0062] In this embodiment, the collection of sensor data corresponding to the user's typing behavior on the password input interface specifically includes the following steps:

[0063] Determine if the mobile terminal is currently in a password input interface;

[0064] If so, the sensors are activated for listening, and the user's input behavior data is collected using the sensors of the mobile terminal; wherein, the sensors include a linear accelerometer and a gyroscope installed in the mobile terminal;

[0065] The input behavior data is divided into small time windows, splitting long periods of data into several small time window samples. These small time window samples are used as training samples or samples to be predicted for the typing behavior prediction model. Furthermore, each small time window sample is labeled with the user's identity. For example, when a user's input behavior data accumulates to a certain duration, it is divided into small time windows, such as 0.2 seconds or 0.5 seconds, thus dividing the long period of data into many small time window samples.

[0066] In the preferred embodiment A, the typing behavior prediction model is trained through the following steps:

[0067] A110. Based on the template samples of this user and the template samples of other users, construct training sample pairs.

[0068] A120. Using training sample pairs as input data for training, a deep learning approach is employed to train the model, resulting in a typing behavior prediction model used to generate identity recognition results from sensor data.

[0069] In this method, a template sample is needed in the construction of the sample pair. There are two main ways to select the template sample. The first method is the single-sample method. This method refers to a sample that is representative of the user as the template sample or registration sample. A common way to obtain the template sample is to ask the user to input a predetermined input content before the user's first use of typing for identity authentication. The sensor data at the time of inputting this content is used as the user's template sample. It can be understood that this is one way to generate template samples. Alternatively, the optimal template sample can be selected from the authenticated historical data through calculation. One method is to use a fast Fourier transform to convert the small time window samples to the time domain and frequency domain, calculate the frequency domain features, traverse the existing samples, calculate the similarity of each sample with respect to other samples, and select the sample with the highest overall similarity as the template sample. When using the single-sample method, the template sample does not change in the short term, and there is only one template sample in the short term, hence the name single-sample method. The second method is the random sample method, which randomly selects authenticated historical samples as template samples. After constructing the template samples, positive and negative samples for training and testing need to be built, using the positive and negative sampling approach. The template samples are combined with other samples to obtain many sample pairs, further constructing a complete sample set. Other samples refer to samples from the user themselves or samples from other users. If it is a sample from the user themselves, it is considered a positive sample, as the behavioral pattern between the template sample and other samples matches; if it is a sample from other users, it is considered a negative sample, as the behavioral pattern between the template sample and other samples does not match. Thus, when forming training sample pairs, each sample pair is labeled with a corresponding matching tag based on whether they match.

[0070] In another preferred embodiment B, the typing behavior prediction model is further trained through the following steps:

[0071] B110. Based on the template samples of this user and the template samples of other users, construct training sample pairs. The process of constructing training sample pairs in step B110 is the same as that in step A110, and will not be elaborated on here.

[0072] B120. Extract training features from training sample pairs.

[0073] In this method, the training sample pairs are processed using two approaches: the first is a feature engineering approach corresponding to step B120, and the second is an end-to-end approach corresponding to step A120. The feature engineering approach involves manually extracting features before building the model. Extractable features include general statistical features, local features, signal features, frequency domain features, and cross-feature features, such as minimum, maximum, mean, variance, frequency domain features, spectral entropy, amplitude, rolling features, zero-crossing rate, number of peaks, and rate of change. The second approach involves directly inputting the raw data into the deep learning model, allowing the model to directly extract features.

[0074] B130. Using the training feature pairs as input data for training, a deep learning approach is employed to train the model, resulting in a typing behavior prediction model used to generate identity recognition results from sensor data.

[0075] In this method, model construction primarily employs metric learning, one approach of which involves constructing a Siamese network structure. The Siamese network has two inputs, including two feature extraction networks, whose weights are shared. Due to the two approaches used in processing the training sample pairs, the sub-network structure differs depending on the approach. If feature engineering has been performed (step B120), the sub-network structure mainly consists of fully connected layers. If an end-to-end approach (step A120) is used, where the model input is the template sample and the original data from other samples, the sub-network structure needs to employ a structure with temporal model extraction capabilities, such as recurrent neural networks, temporal convolutional structures, temporal transformer structures, and fully connected structures. Combinations of different structures can achieve feature extraction capabilities.

[0076] In this embodiment, the sub-network structure of the first method uses a multi-layer stack of one-dimensional convolution and WaveNet network, with one-dimensional convolution combined with a bidirectional long short-term memory network. Preferably, it is a one-dimensional convolution and WaveNet network structure.

[0077] The model for one-dimensional convolution and WaveNet network schemes is as follows:

[0078] First layer: 16-kernel one-dimensional convolution, kernel size is 1, padding method is same, activation function is mish;

[0079] The second layer: a 16-core Wavenet module with a core size of 3 and 16 stacking layers inside the Wavenet module;

[0080] The third layer: 32-kernel one-dimensional convolution, kernel size is 1, padding method is same, activation function is mish;

[0081] Fourth layer: 32-core Wavenet module, with a core size of 3 and an internal stacking layer of 8.

[0082] Fifth layer: 64-kernel one-dimensional convolution, kernel size 1, padding method is same, activation function is mish;

[0083] The sixth layer: a 64-core Wavenet module with a core size of 3 and an internal stacking layer of 4.

[0084] The seventh layer: 128-kernel one-dimensional convolution, kernel size 1, padding method is same, activation function is mish;

[0085] Layer 8: 128-core Wavenet module, with a core size of 3 and an internal stacking layer of 1.

[0086] Ninth layer: One-dimensional global average pooling.

[0087] Therefore, there are two ways to output the model's features. The first way is to concatenate, subtract, or otherwise manipulate the features extracted from the template sample and other samples, then connect them to a fully connected layer or other discriminative model. The final output is a binary classification: if the template sample and other samples belong to the same user, the output is a match; if the template sample and other samples do not belong to the same user, the output is a mismatch. The second way is to optimize the network using a pair-based loss function, such as triplet loss or constant loss. Taking triplet loss as an example, a training set consisting of template samples, positive samples, and negative samples is constructed. A Siamese network with shared weights is used to extract deep features from the sensor data on the training set. After successful extraction, the extracted deep embedding features are input into the triplet loss function for error calculation. The error is then backpropagated to train the Siamese network. In this embodiment, the preferred loss function is triplet loss. Simultaneously, if triplet loss is chosen, a main network needs to be defined during the training phase. Within the main network, a base network needs to be defined. The base network is a weight-sharing model, primarily employing a time-series-based neural network architecture to extract features from samples. The application of the base network varies depending on the loss function. The main network simultaneously inputs template samples, positive samples, and negative samples, using the weight-sharing base network to extract features from these three types of samples respectively. The extracted features from these three types of samples are then used as the output of the main network. The main network uses triplet loss to calculate the loss on the outputs of these three types of samples. During the training of the base and main networks, activation functions involved include ReLU, Mish, sigmoid, and tanh. The optimizer used is the Ranger optimizer, and the optimization strategy is Batch Normalization (BN). L2 weight decay regularization, early stopping, and Dropout are employed, and Bayesian optimization is used to optimize the parameters of both the main and base networks.

[0088] Following steps A110 and B110, the method further includes the following steps:

[0089] A111 / B111 Preprocess the training sample pairs; among them, some filtering and noise reduction techniques are used for preprocessing to filter noise. For example, noise reduction methods such as smoothing filtering, median filtering, average filtering, and Kalman filtering can be used to denoise the training sample pairs.

[0090] The outputs of the base network for different types of samples form the outputs of the main network. Taking triplet loss as an example, triplet loss controls the features learned by the base network, making the features of positive samples closer to the features of the template sample, and making the features of negative samples farther away from the features of the template sample, thus enabling the base network to learn the feature differences between sensor data from different people. In the prediction stage, taking a login scenario as an example, when an account is logged in, a sensor data point to be authenticated is generated. A template sample is extracted from the authenticated data of the user of that account. The trained typing behavior prediction model is used to calculate the embedded feature representations of the typing behavior prediction model for both the sensor data and the authenticated template sample. Then, the feature distance is calculated. The authentication distance is compared with the preset distance threshold set during the verification of the typing behavior prediction model, i.e., the authentication pass threshold. If the authentication distance is greater than the preset distance threshold, the authentication is considered unsuccessful, and the identity authentication result is "yes, not the user" or "other user". If the authentication distance is less than the preset distance threshold, the authentication is considered successful, and the identity authentication result is "yes, not the user".

[0091] Based on the preferred embodiment B described above, another preferred embodiment of the present invention may also be:

[0092] In the training phase of the typing behavior prediction model, the training samples include template samples, positive samples, and negative samples. These template samples, positive samples, and negative samples are obtained by collecting sensor data from each user at the password input interface. The authentication distance refers to the mean square distance between the template sample and the positive and negative samples, respectively, to learn the behavioral differences between the password input habits of the original user and those of unauthorized users. In the prediction phase of the typing behavior prediction model, the authentication distance refers to the mean square distance between the sample to be judged and the template sample. The authentication distance is compared with a threshold, specifically including the following steps:

[0093] The sensor data of the current user is input into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model; the current user is either the user corresponding to the training sample in the typing behavior prediction model or a new user that has not been trained in the typing behavior prediction model.

[0094] The present invention determines the relationship between the authentication distance and a preset distance threshold, specifically as follows:

[0095] If the authentication distance is greater than a preset distance threshold, the identity authentication result is that the comparison between the authentication distance and the threshold has failed.

[0096] If the authentication distance is less than a preset distance threshold, the identity authentication result is that the authentication distance passes the threshold comparison.

[0097] In this invention, during model training, the distances (e.g., mean squared distances) between template samples and positive and negative samples are calculated on the validation set. Then, a reasonable distance threshold is defined based on an indicator. Samples with distances greater than the threshold are considered authentication failures, while those less than the threshold are considered successful. During prediction, the distance between the sample to be judged and the template sample is calculated, and a judgment is made based on the threshold. The template sample at this stage is selected from those corresponding to the login account. Because a metric learning method is used, the model built in the previous steps learns the differences between the user's own actions and the password input habits of non-users. Through extensive training data, the final judgment distance between the user and others is significantly greater, enabling generalization to novel users. For example, for a new user, even if samples exist in the database but were not included in model training, the constructed model can still identify the differences in input patterns between this user and other users, meaning that the model can make judgments for new users without retraining the previous model.

[0098] The above steps predict the distance between the input habits of the unknown person entering the password and the account holder based on the model prediction. The final judgment needs to be combined with the password authentication situation.

[0099] There are four possible outcomes for two-factor authentication:

[0100] 1. Password verification passes, and the model output distance is compared with the threshold: In this case, it is proven that the user has passed the dual authentication and is allowed to pass.

[0101] 2. Password verification fails, but the model output distance passes the threshold comparison: In this case, it is possible that the user entered the wrong password or forgot the password. In this case, the number of password input errors can be increased appropriately to prevent the account from being frozen due to the user's own reasons. This allows for dynamic adjustment of the number of errors to improve the user experience.

[0102] 3. Password verification passes, but model output distance and threshold comparison fails: In this case, there may be a situation where someone other than the account holder has illegally logged into the account, which poses a high risk. Different risk control measures can be used to control this. For example, a more lenient risk control measure is to immediately send an SMS or make a phone call to notify the actual controller of the account, while a more stringent risk control measure is to immediately call other hard biometric authentication methods to re-authenticate the account.

[0103] IV. Password verification failed, and the model output distance failed to compare with the threshold: In this case, it can be considered that there is an extremely high risk of unauthorized login. Risk control measures can be upgraded by reducing the number of password input errors or immediately calling hard biometric authentication.

[0104] The identity authentication device provided by the present invention is described below. The identity authentication device described below can be referred to in correspondence with the identity authentication method described above.

[0105] The device in this embodiment specifically includes:

[0106] The first verification module is used to detect the user's typing behavior on the password input interface, verify the password entered by the user, and obtain the password verification result;

[0107] The second verification module is used to enable sensor monitoring at the password input interface, collect sensor data corresponding to the user's typing behavior at the password input interface, and input the sensor data into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model.

[0108] The data analysis module is used to perform user authentication based on the password verification result and the authentication distance output by the model.

[0109] If the password verification passes and the authentication distance output by the model passes the threshold comparison, then the user's identity authentication is deemed successful.

[0110] If the password verification passes, but the authentication distance output by the model fails to compare with the threshold, it is determined that there may be an unauthorized user login, and the first risk control measure is activated to control the user or biometric authentication is activated.

[0111] If the password verification fails, and the authentication distance output by the model passes the threshold comparison, it is determined that the user may have entered the wrong password, and the number of password input errors is increased.

[0112] If password verification fails, and the authentication distance output by the model fails to compare with the threshold, it is determined to be a high-risk unauthorized user login. The number of password input errors will be reduced, a second risk control measure will be activated, or biometric authentication will be activated.

[0113] This method uses a data acquisition module to perform a series of verifications seamlessly in the background of the mobile terminal when the user uses it, such as an app installed on the mobile terminal, without requiring the interruption of app operation for dedicated biometric authentication like facial recognition or fingerprint recognition. Using sensor-collected data for identity authentication is more privacy-friendly, serving both verification and minimizing harm to users in the event of leakage. Furthermore, sensor data is time-sensitive; it automatically becomes invalid after a period of time due to changes in input habits or other objective factors, improving the accuracy of identity authentication. The model built based on sensor data through a behavior prediction module is both versatile and meets accuracy requirements.

[0114] This invention utilizes linear accelerometers and gyroscopes to collect data such as movement acceleration and rotational angular velocity changes when a user uses a mobile terminal app. Before data collection, the sampling frequency of each sensor needs to be set, such as 50Hz or 100Hz. When the user on the mobile terminal engages in input behavior, the sensors are activated to listen and collect sensor data during typing.

[0115] Since there is a lack of template data for the user in the early stages, sensor data collected from the same session that has already passed facial and fingerprint verification can be stored in a database as a sample template for the authenticated user. The collection of sample templates is for subsequent judgment of unknown samples, serving as benchmark data for authentication comparison. Simultaneously, in the early data collection phase, other biometric authentication methods such as fingerprints and facial recognition can also be combined to label the user's data.

[0116] The outputs of the base network for different types of samples form the outputs of the main network. Taking triplet loss as an example, triplet loss controls the features learned by the base network, making the features of positive samples closer to the features of template samples, and making the features of negative samples farther from the features of template samples, thus enabling the base network to learn the feature differences between sensor data from different people. In the prediction stage, taking a login scenario as an example, when an account is logged in, a sensor data point to be authenticated is generated. A template sample is extracted from the authenticated data of the user of that account. The trained typing behavior prediction model is used to calculate the embedded feature representations of the typing behavior prediction model for both the sensor data and the authenticated template sample, and then the feature distance is calculated. The authentication distance is compared with the preset distance threshold set during the verification of the typing behavior prediction model, i.e., the authentication pass threshold. If the authentication distance is greater than the preset distance threshold, the authentication is considered unsuccessful, and the identity authentication result is that the user is not the intended user, but another user. If the authentication distance is less than the preset distance threshold, the authentication is considered successful, and the identity authentication result is that the user is the intended user.

[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0118] The identity authentication method and apparatus of the present invention have the following beneficial effects:

[0119] 1. This method can perform dual authentication in the background without the user's awareness when entering the password, and can take different risk control measures based on the dual authentication results, thus improving the user experience.

[0120] 2. Using sensor data for identity authentication is more privacy-friendly for users, as it can both verify identity and minimize harm to users in the event of a leak.

[0121] 3. Using sensor input behavior for identity authentication has extremely high accuracy. Under the test set constructed in this scheme, the AUC can reach 0.81, which has high accuracy.

[0122] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0123] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0124] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An identity authentication method applied to a mobile terminal, characterized in that, Includes the following steps: The system detects the user's typing behavior on the password input interface, verifies the password entered by the user, and obtains the password verification result. Sensor monitoring is enabled on the password input interface to collect sensor data corresponding to the user's typing behavior on the password input interface; The sensor data is input into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model; User authentication is performed based on the password verification result and the authentication distance output by the model: If the password verification passes and the authentication distance output by the model passes the threshold comparison, then the user's identity authentication is deemed successful. If the password verification passes, but the authentication distance output by the model fails to compare with the threshold, it is determined that there may be an unauthorized user login, and the first risk control measure is activated to control the user or biometric authentication is activated. If the password verification fails, and the authentication distance output by the model passes the threshold comparison, it is determined that the user may have entered the wrong password, and the number of password input errors is increased. If password verification fails, and the authentication distance output by the model fails to compare with the threshold, it is determined to be a high-risk unauthorized user login. The number of password input errors will be reduced, a second risk control measure will be activated, or biometric authentication will be activated.

2. The identity authentication method for mobile terminals according to claim 1, characterized in that, The process of collecting sensor data corresponding to the user's typing behavior on the password input interface specifically includes the following steps: Determine if the mobile terminal is currently in a password input interface; If so, the sensors are activated for listening, and the user's input behavior data is collected using the sensors of the mobile terminal; wherein, the sensors include a linear accelerometer and a gyroscope installed in the mobile terminal; The input behavior data is divided into small time windows, and the long-term data is divided into several small time window samples. These small time window samples are used as training samples or samples to be predicted for the typing behavior prediction model. Furthermore, each small time window sample is labeled with the user identity to which it belongs.

3. The identity authentication method applied to a mobile terminal according to claim 1, characterized in that, The typing behavior prediction model is trained based on template samples from password input users and template samples from other users; specifically, the typing behavior prediction model is trained through the following steps: Based on the template samples of this user and the template samples of other users, construct training sample pairs; The training sample pairs are used as input data for training, and deep learning is used to train the model to obtain the typing behavior prediction model used to generate the identity recognition result of the sensor data.

4. The identity authentication method applied to a mobile terminal according to claim 1, characterized in that, The typing behavior prediction model is trained based on features extracted from template samples of password input users and template samples of other users. Specifically, the typing behavior prediction model is trained through the following steps: Based on the template samples of this user and the template samples of other users, construct training sample pairs; Extract the training features from the training sample pairs; The training feature pairs are used as input data for training, and deep learning is employed to train the model to obtain the typing behavior prediction model used to generate the identity recognition result of the sensor data.

5. The identity authentication method applied to a mobile terminal according to claim 4, characterized in that, The training features include statistical features, local features, signal features, frequency domain features, and cross features.

6. The identity authentication method applied to a mobile terminal according to any one of claims 3-5, characterized in that, After the step of constructing training sample pairs based on the template samples of the user and other users, the method further includes the following steps: The training sample pairs are preprocessed; wherein the preprocessing methods include smoothing filtering, median filtering, averaging filtering, and Kalman filtering.

7. The identity authentication method applied to a mobile terminal according to claim 1, characterized in that, In the training phase of the typing behavior prediction model, the training samples include template samples, positive samples, and negative samples. These template samples, positive samples, and negative samples are obtained by collecting sensor data from each user at the password input interface. The authentication distance refers to the mean square distance between the template sample and the positive and negative samples, respectively, to learn the behavioral differences between the password input habits of the original user and those of unauthorized users. In the prediction phase of the typing behavior prediction model, the authentication distance refers to the mean square distance between the sample to be judged and the template sample. The authentication distance is compared with a threshold, specifically including the following steps: The sensor data of the current user is input into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model; the current user is either the user corresponding to the training sample in the typing behavior prediction model or a new user that has not been trained in the typing behavior prediction model. Determine the relationship between the authentication distance and the preset distance threshold; If the authentication distance is greater than a preset distance threshold, the identity authentication result is that the comparison between the authentication distance and the threshold has failed. If the authentication distance is less than a preset distance threshold, the identity authentication result is that the authentication distance passes the threshold comparison.

8. An identity authentication device for a mobile terminal, characterized in that, Includes the following steps: The first verification module is used to detect the user's typing behavior on the password input interface, verify the password entered by the user, and obtain the password verification result; The second verification module is used to enable sensor monitoring on the password input interface and collect sensor data corresponding to the user's typing behavior on the password input interface. The sensor data is input into the typing behavior prediction model to obtain the authentication distance output by the typing behavior prediction model; The data analysis module is used to perform user authentication based on the password verification result and the authentication distance output by the model. If the password verification passes and the authentication distance output by the model passes the threshold comparison, then the user's identity authentication is deemed successful. If the password verification passes, but the authentication distance output by the model fails to compare with the threshold, it is determined that there may be an unauthorized user login, and the first risk control measure is activated to control the user or biometric authentication is activated. If the password verification fails, and the authentication distance output by the model passes the threshold comparison, it is determined that the user may have entered the wrong password, and the number of password input errors is increased. If password verification fails, and the authentication distance output by the model fails to compare with the threshold, it is determined to be a high-risk unauthorized user login. The number of password input errors will be reduced, a second risk control measure will be activated, or biometric authentication will be activated.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the identity authentication method applied to a mobile terminal as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the identity authentication method applied to a mobile terminal as described in any one of claims 1 to 7.