Mobile Device Owner Authentication Method Based on Denoising and Optimized Time-Domain Convolutional Network

Abstract

The present invention belongs to the technical fields of network security and deep learning, and discloses a mobile device owner authentication method based on denoising and optimizing a temporal convolutional network, including: obtaining a training sample set, where the training sample set contains multiple training samples, and each training sample covers the feature data used by the mobile device during the owner authentication process and the label corresponding to the feature data; denoising the training sample set by using a joint denoising method; optimizing and updating an optimized temporal convolutional network by using the denoised training sample set to obtain an owner authentication model, where the optimized temporal convolutional network includes a self-attention layer; obtaining the feature data of the user of the mobile device to be authenticated as the input of the optimized and updated owner authentication model, and obtaining an authentication result on whether the user of the mobile device to be authenticated is the owner based on the output of the owner authentication model. The present invention can achieve implicit, accurate, and real-time identification of mobile devices.

Description

Technical Field

[0001] The present invention belongs to the technical fields of network security and deep learning, and particularly relates to a mobile device owner authentication method based on denoising and optimizing a time-domain convolutional network. Background Art

[0002] The rise and popularity of the mobile Internet benefit from the rapid development of mobile communication technology and the Internet. With the growth of the usage rate of mobile devices, more and more private data and personal authentication information are collected and used for the operation of mobile devices. Given the hardware, software, and application scenarios of mobile devices, it is crucial to develop a suitable and reliable authentication mode to protect user information security and prevent unauthorized access to local private and financial information stored on mobile devices. Implicit authentication collects and calculates data based on the perception of the mobile terminal to complete the evaluation of user authenticity, and this process is based on the user's biometric data or behavioral data for identification.

[0003] Currently, the research on implicit real-time mobile device owner authentication globally is gradually developing towards real-time intelligence and integration. The specific technologies involved in this research include multi-sensor fusion, machine learning, mobile computing, human-computer interaction, and other fields. Some progress has been made in technologies such as behavior perception, data collection, and behavioral data processing. However, the current research is still in its infancy, mainly focusing on academic exploration and prototype verification. In the key technology directions, there is still a lack of a systematic research system, and there is also a large gap between the research and actual applications.

[0004] For the above reasons, it is necessary to propose a more ideal and practical mobile device owner authentication technology to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a mobile device owner authentication method based on denoising and optimizing a time-domain convolutional network to achieve implicit, accurate, and real-time identification of mobile devices.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A mobile device owner authentication method based on denoising and optimizing a time-domain convolutional network, the mobile device owner authentication method based on denoising and optimizing a time-domain convolutional network includes:

[0008] Obtain a training sample set, the training sample set includes multiple training samples, and each training sample covers the feature data used by the mobile device during the owner authentication process and the label corresponding to the feature data;

[0009] Denoise the training sample set using the joint denoising method, where the joint denoising method includes generating signals, signals, signals and signals for the training sample set by using ensemble empirical mode decomposition, and extracting the criterion to extract the signal and the detail part of the signal. Simultaneously denoise and unmix the detail part of the signal, the signal, the signal and the detail part of the signal by using the Fast-ICA algorithm to obtain the component, the component, the component and the component. Apply wavelet threshold denoising to the component to obtain the denoised training sample set;

[0010] Optimize and update the optimized time-domain convolutional network by using the denoised training sample set to obtain the owner authentication model, where the optimized time-domain convolutional network includes a self-attention layer;

[0011] Obtain the feature data of the mobile device user to be authenticated as the input of the optimized and updated owner authentication model, and obtain the authentication result of whether the mobile device user to be authenticated is the owner based on the output of the owner authentication model.

[0012] The following also provides several optional methods, which are not additional limitations to the above overall solution, but are only further supplements or optimizations. Without technical or logical contradictions, each optional method can be combined with the above overall solution alone, or multiple optional methods can be combined with each other.

[0013] Preferably, the obtaining of the training sample set includes:

[0014] When the mobile device is in the owner authentication process, collect the three-axis data of the acceleration sensor, the three-axis data of the gyroscope, and the three-axis data of the gravity sensor;

[0015] Add labels to each collected data point, where the labels include the user number, sensor type, axis type, and timestamp when the data point is collected.

[0016] Preferably, the denoising of the training sample set using the joint denoising method includes:

[0017] Take the three-axis data of the acceleration sensor, the three-axis data of the gyroscope, and the three-axis data of the gravity sensor in the training sample set to obtain a total of nine-axis time-series data, and denoise the time-series data of each axis using the joint denoising method.

[0018] Preferably, the structure of the optimized time-domain convolutional network includes an input layer, a time-domain convolutional layer, a self-attention layer, and a fully connected layer connected in sequence.

[0019] Preferably, the structure of the time-domain convolutional layer includes a plurality of residual blocks, and a convolutional layer and a ReLU activation function are added after every two residual blocks.

[0020] Preferably, the residual block includes a first branch and a second branch. The first branch sequentially includes a convolutional layer, a ReLU activation function, and a convolutional layer. The second branch includes a downsampling layer. The outputs of the first branch and the second branch are added together as the output of the residual block.

[0021] Preferably, the calculation process of the attention value of the self-attention layer is as follows:

[0022] Calculate the correlation between the query vector and the key vector:

[0023] ;

[0024] In the formula, represents the query vector, represents the -th element value of the key vector, represents the correlation between the query vector and the -th element value of the key vector;

[0025] Calculate the weight of the value vector:

[0026] ;

[0027] In the formula, represents the weight of the -th element in the value vector, represents the SoftMax function, is the dimension of the key vector, represents the correlation between the query vector and the -th element value of the key vector;

[0028] Calculate the attention value:

[0029] ;

[0030] In the formula, is the attention value, represents the -th element value of the value vector.

[0031] Preferably, the owner authentication model runs on a mobile device for local owner authentication, or the owner authentication model is deployed on the server side to perform cloud owner authentication based on the feature data uploaded by the mobile device.

[0032] Preferably, the output of the owner authentication model is the similarity between the user of the mobile device to be authenticated and the owner. If the similarity is greater than the threshold, the authentication result is that the identity matches the owner and the authentication is successful; otherwise, the authentication result is that the identity does not match the owner and the authentication fails.

[0033] A mobile device owner authentication method based on denoising and optimizing the time-domain convolutional network provided by the present invention. To achieve implicit real-time owner authentication of mobile devices and apply it in practice, the present invention provides a mobile device owner authentication method based on denoising and optimizing the time-domain convolutional network. This method uses a joint denoising method to denoise the training sample set and introduces a self-attention mechanism into the time-domain convolutional network, thereby realizing implicit, accurate, and real-time identification of mobile devices.

[0034] The beneficial effects of the present invention are mainly reflected in: (1) By using the self-designed ICA-EEMD-wavelet threshold sensor signal joint denoising method to filter sensor noise, reducing the impact of the training sample set on the results during the training stage, thereby improving the robustness of the model. (2) Using an optimized time-domain convolutional network for efficient calculation, effectively capturing the relative relationships and local dependencies between time series data by stacking convolutional layers and increasing the receptive field of the convolutional kernel and the self-attention mechanism, understanding deeper features, improving the analysis ability of the model, and enhancing the accuracy and feasibility of implicit owner authentication. Description of the Drawings

[0035] Figure 1 It is a flowchart of a mobile device owner authentication method based on denoising and optimizing the time-domain convolutional network of the present invention;

[0036] Figure 2 It is a flowchart of the joint denoising method of the present invention;

[0037] Figure 3 It is a schematic structural diagram of the optimized time-domain convolutional network of the present invention. Detailed Embodiments

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this invention belongs. The terms used in the description of the present invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention.

[0040] As Figure 1 shown, this embodiment provides a mobile device owner authentication method based on denoising and optimizing a time-domain convolutional network, including the following steps:

[0041] (1) Obtain a training sample set, which contains multiple training samples. Each training sample covers the feature data used in the owner authentication process of the mobile device and the label corresponding to the feature data.

[0042] When the mobile device is in the owner authentication process, collect the three-axis data of the acceleration sensor, the three-axis data of the gyroscope, and the three-axis data of the gravity sensor, and add a label to each collected data point. The label includes the user number, sensor type, axis type, and timestamp when the data point is collected. This embodiment provides a specific acquisition method as follows:

[0043] (1-1) Preparations and parameter construction in advance: Each sensor reading includes values corresponding to the x, y, and z axes:

[0044] ;

[0045] ;

[0046] ;

[0047] wherein, represents the th time step in the data acquisition time period, represents the data collected on the x-axis at the th time step of the acceleration sensor, represents the data collected on the y-axis at the th time step of the acceleration sensor, represents the data collected on the z-axis at the th time step of the acceleration sensor, represents the data collected on the x-axis at the th time step of the gyroscope, represents the data collected on the y-axis at the th time step of the gyroscope, represents the data collected on the z-axis at the th time step of the gyroscope, Denote the data collected by the x-axis of the gravity sensor at the th time step, denote the data collected by the y-axis of the gravity sensor at the th time step, denote the data collected by the z-axis of the gravity sensor at the th time step.

[0048] Taking Android as an example, Android allows developers to refresh sensor data at fixed and customized intervals / delays after registering the sensor as a register listener(). There are four delays:

[0049] SENSOR_DELAY_FASTEST = 0s;

[0050] SENSOR_DELAY_GAME = 0:02s (50HZ);

[0051] SENSOR_DELAY_NORMAL = 0.06s;

[0052] SENSOR_DELAY_UI = 0.2s.

[0053] Considering battery consumption, SENSOR_DELAY_GAME delay is selected for data collection. In this embodiment, an existing mobile application is used for data collection. This application continuously detects during the loading stage when the user actively uses the mobile device, and presents the characteristics of the user using the mobile device through sensor readings within a specified time.

[0054] Generally, the duration of the loading stage lasts for 2 to 4 seconds. The selection of the duration is based on sensitivity analysis. When the time is less than 3 seconds, the final accuracy increases with the increase of time. When the time is greater than 3 seconds, the final accuracy will decrease with the increase of time. In this embodiment, 3 seconds is used as the data collection duration.

[0055] (1 - 2) Data collection process: Taking Android as an example, the software uses the BroadcastReceiver() method to capture the system event of turning on the screen of the device, and then starts a service to regularly query the current application in the foreground. If the currently active application is different from the application in the last query, the software will recognize that a new application has started. If both conditions are met, the data collection will last for 3 seconds:

[0056] The first condition is: the screen of the mobile device is in the on state;

[0057] The second condition is: a new application is running on the home page.

[0058] The boundaries of the three-dimensional gravity sensor readings are as follows. Data that does not fall within this range will be discarded: .

[0059] (2)The training sample set is denoised using the joint denoising method. The joint denoising method includes using ensemble empirical mode decomposition to generate signals, signals, signals and signals for the training sample set, and using the criterion to extract the detailed parts of the signals and signals. The Fast-ICA algorithm is used to simultaneously denoise and unmix the detailed parts of the signals, signals, signals and signals to obtain the components, components, components and components. The wavelet threshold denoising is applied to the components to obtain the denoised training sample set.

[0060] As Figure 2 shown, the proposed joint denoising method of ICA (Fast-ICA algorithm)-EEMD (Ensemble Empirical Mode Decomposition)-wavelet threshold in this embodiment performs denoising processing on the data of each dimension, that is, the three-axis data of the acceleration sensor, the three-axis data of the gyroscope, and the three-axis data of the gravity sensor in the training sample set are taken, and a total of nine-axis time series data are obtained. The joint denoising method is used to denoise the time series data of each axis. The specific denoising steps are as follows:

[0061] (2-1)The training sample set is decomposed into four layers of IMF signals using EEMD. It is found through the spectrum that the source signal waves are concentrated on the signals and signals. These two layers of IMF signals are retained. The EEMD decomposition process is as follows:

[0062] (2-1-1)Add white Gaussian noise with the same height to the original signal to obtain the new signal as shown below:

[0063] ;

[0064] In the formula, is the signal after adding noise for the th time in the original signal , is the The white Gaussian noise added each time should meet the following conditions:

[0065] ;

[0066] In the formula, is the standard deviation of the input signal, and are the intensity and number of times of the added noise respectively.

[0067] (2-1-2) Solve the mean value by fitting the upper and lower envelopes of the signal . During fitting, all local maximum points and local minimum points need to be identified from the preprocessed signal (the signal when first executed in step (2-1-1), and the detected signal obtained in step (2-1-3) in subsequent steps). Using the cubic spline curve interpolation method, connect all local maximum points and all local minimum points respectively, so as to obtain the upper envelope and the lower envelope of the signal. The upper envelope is above all local maximum points, and the lower envelope is below all local minimum points. The calculation is carried out using the following formula:

[0068] ;

[0069] In the formula, is the mean value sequence.

[0070] (2-1-3) Remove the mean value sequence from the preprocessed signal to obtain the detected signal, and judge whether the detected signal meets the IMF conditions. If not, continue to repeat step (2-1-2) based on the detected signal until the detected signal meets the IMF conditions, and output the detected signal as the intrinsic mode function IMF component.

[0071] (2-1-4)Use the signal and the intrinsic mode function to calculate the remaining signal, and the formula is as follows:

[0072] ;

[0073] In the formula, is the remaining signal of the th addition of the equal-height Gaussian white noise, is the th intrinsic mode function of the th addition of the equal-height Gaussian white noise, so is the first intrinsic mode function of the th addition of the Gaussian white noise.

[0074] (2-1-5) Repeat steps (2-1-2) and (2-1-3) for the remaining signal to obtain the IMF components successively;

[0075] (2-1-6) Re-add white noise to the original signal and repeat the above steps;

[0076] (2-1-7) Perform ensemble averaging on the IMF components of the same order obtained by decomposition. Then, the decomposition result of EEMD is:

[0077] ;

[0078] where is the th intrinsic mode function of the th addition of Gaussian white noise, and is the total number of times of adding Gaussian white noise.

[0079] (2-2) Use the criterion to extract the detailed parts of the signal and the signal. The

[0080] criterion is based on the equal-precision repeated measurement of the normal distribution and is adopted when the interference or noise of singular data is difficult to satisfy the normal distribution. The specific steps are as follows: According to the principle of ensemble empirical mode decomposition, the signal and the

[0081] ;

[0082] ;

[0083] where represents the signal, represents the signal, represents the remaining information after removing the noise from the signal, represents the remaining information after removing the noise from the signal, represents the noise in the signal, represents the noise in the signal, and , .

[0084] (2-2-2) According to the criterion, the noise distribution satisfies , where represents the th time point. From the above, the probability that the noise falls within the interval is 0.9973. Therefore, if the signal value at the th time point of the signal does not fall within , it is considered that necessarily contains significant errors and there must be information that needs to be retained. Using the criterion to perform detail extraction on the signal is expressed as:

[0085] ;

[0086] In the formula, is the detail part extracted from the

[0087] (2-2-3) The noise variance of the

[0088] ;

[0089] In the formula, is a function that returns the median, and is the high-frequency subband wavelet coefficient of the

[0090] (2-2-4) Noise is distributed opposite to . Using the criterion, we can obtain:

[0091] ;

[0092] In the formula, is the detail part extracted from the

[0093] (2-2-5) The noise variance of the

[0094] ;

[0095] In the formula, is the high-frequency subband wavelet coefficient of the

[0096] (2-3) Use the Fast-ICA algorithm to simultaneously denoise and demix the four IMF signals to obtain the best estimated source signals. The basic model of ICA is as follows:

[0097] (2-3-1) Assume is a set of dimensional mutually independent source signals, is dimensional measured observation signals, The components in are composed of

[0098] ;

[0099] In the formula, is the unknown mixing matrix, satisfying that the number of observation points is greater than or equal to the number of source signal points .

[0100] (2-3-2) The purpose of ICA analysis is to estimate the inverse matrix of the coefficient matrix according to a certain optimization criterion without prior knowledge, and obtain the independent source signals by solving the inverse matrix of , that is to say, . Among them, the inverse matrix can be represented by the separation matrix as follows:

[0101] ;

[0102] In the formula, is the approximate estimate of the source signal obtained by separation.

[0103] (2-4) Use wavelet soft threshold to further denoise the IC1 reconstructed signal and improve the denoising effect and performance indicators. The principle of wavelet threshold denoising is as follows: The basic principle of wavelet threshold denoising is to set a critical threshold , if the wavelet coefficient is less than , then this coefficient is mainly generated by noise, and this part of the coefficient is removed; if the wavelet coefficient is greater than or equal to , then this coefficient is mainly generated by the signal, and this part of the coefficient is retained. Finally, the inverse transform is performed on the processed wavelet coefficients to obtain the denoised signal. The steps of the wavelet threshold denoising method are as follows:

[0104] (2-4-1) Perform wavelet transform on the IC1 original signal to the wavelet domain to obtain a set of wavelet decomposition coefficients.

[0105] After threshold processing in the wavelet domain (2-4-2), smaller wavelet coefficients mainly containing random noise are obtained. In this embodiment, the soft threshold method is used for denoising, and its threshold function is:

[0106] ;

[0107] In the formula, is the wavelet coefficient after threshold processing, is the sign function, is the threshold, is the wavelet coefficient.

[0108] (2-4-3) Using the processed wavelet coefficients for signal reconstruction, the denoised signal is obtained. After denoising processing on the 9-axis time series data, a denoised training sample set is obtained.

[0109] (3) Using the denoised training sample set to optimize and update the optimized time-domain convolutional network, a host authentication model is obtained. The optimized time-domain convolutional network includes a self-attention layer.

[0110] The time-domain convolutional network is an architecture based on the convolutional neural network, which effectively captures the long-term dependencies in the time series by stacking convolutional layers; while the self-attention mechanism can dynamically adjust its weights according to the similarity between each time step in the input sequence. As Figure 3 shown, the implementation of the optimized time-domain convolutional network in this embodiment is as follows: including an input layer, a time-domain convolutional layer, a self-attention layer, and a fully connected layer connected in sequence. Among them, the structure of the time-domain convolutional layer includes multiple residual blocks (8 residual blocks are taken in this embodiment), and a convolutional layer and a ReLU activation function are added after every two residual blocks. The residual block includes a first branch and a second branch. The first branch sequentially includes a convolutional layer, a ReLU activation function, and a convolutional layer. The second branch includes a downsampling layer (the downsampling layer can be replaced by a convolutional layer as needed). The outputs of the first branch and the second branch are added as the output of the residual block. The detailed description of each layer is as follows:

[0111] 1. Input layer: The input of the optimized time-domain convolutional network is a tensor with a shape of (batch_size, input_channels, sequence_length), where batch_size is the batch size, input_channels is the number of input channels (or features), and sequence_length is the sequence length.

[0112] 2. Temporal Convolution (TCN) Layer: First, it passes through a series of residual blocks (ResidualBlock), and each residual block contains two one-dimensional convolutional layers (Conv1d) and a ReLU activation function. These convolutional layers may have different strides and dilations to capture features at different time scales. By introducing residual blocks, the network can learn the identity mapping of the input data, which helps to alleviate the vanishing gradient problem in deep networks, enabling the network to be deeper and thus extract richer features.

[0113] 3. Convolutional Layer: Each convolutional layer is followed by a ReLU activation function to increase non-linearity, enabling the network to fit more complex functional relationships.

[0114] 4. Downsampling Layer: If the stride is not equal to 1 or the dilation rate is greater than 1, the input is downsampled through a one-dimensional convolutional layer (downsample) to ensure dimension matching during residual connections, which can ensure dimension matching during residual connections, reduce the computational amount, and improve the model efficiency.

[0115] 5. Additional Convolutional Layer: After every two residual blocks, an additional one-dimensional convolutional layer and ReLU activation function are added to provide more flexibility for further feature extraction.

[0116] 6. Self-Attention Layer (SelfAttentionLayer): After passing through a series of temporal convolutional layers, a self-attention layer is applied to capture the dependencies between different positions in the sequence. The self-attention mechanism can capture the dependencies between different positions in the sequence, which is particularly important for time series data. It allows the network to dynamically focus on important parts of the sequence, improving the model's expressive ability and generalization ability.

[0117] Query Vector, Key Vector, Value Vector: First, the input is transformed into query, key, and value through three linear layers.

[0118] Attention Score: Then, the dot product between the query vector and the key vector is calculated to obtain the attention score. That is, calculate the correlation between the query vector and the key vector:

[0119] ;

[0120] In the formula, represents the query vector, represents the -th element value of the key vector, represents the correlation between the query vector and the -th element value of the key vector.

[0121] Attention weight: Apply the Softmax function to the attention scores to obtain the attention weights. Introduce a calculation method similar to SoftMax to perform numerical conversion on the scores in the first stage. On the one hand, normalization can be performed to organize the original calculated scores into a probability distribution where the sum of the weights of all elements is 1; on the other hand, the weight of important elements can be made more prominent through the internal mechanism of SoftMax. That is, generally, the following formula is used:

[0122] ;

[0123] In the formula, represents the weight of the -th element in the value vector, represents the SoftMax function, is the dimension of the key vector, represents the correlation between the query vector and the -th element value of the key vector.

[0124] Weighted sum: Use the attention weights to perform a weighted sum on the values to obtain a weighted feature representation.

[0125] Aggregation: Finally, sum the weighted feature representations in the sequence length dimension to aggregate information, which helps to integrate the global information in the sequence into the representation of each time step. That is, perform a weighted sum on Value according to the weight coefficients to obtain the final attention value:

[0126] ;

[0127] In the formula, is the attention value, represents the -th element value of the value vector.

[0128] 7. Fully Connected Layer Finally, map the output of the self-attention layer to the probability distribution of the target categories through a fully connected layer. The input size of this fully connected layer is input_channels, and the output size is num_classes, that is, the number of categories. The output of the temporal convolutional network is a tensor with a shape of (batch_size, num_classes), representing the probability that each sample belongs to each category.

[0129] (4) Obtain the feature data of the mobile device user to be authenticated as the input of the optimized and updated owner authentication model, and obtain the authentication result of whether the mobile device user to be authenticated is the owner based on the output of the owner authentication model.

[0130] Obtain the characteristic data of the user of the mobile device to be authenticated as the input of the owner authentication model. During the verification process, in this embodiment, the existing IOS development library AFNetworking and the existing ApacheHttpClient library of the JAVA application are used to cover mobile devices with different operating systems, so as to simplify network request operations when interacting with the server. The data sent by the client is processed by a server built based on Django. A C++ data preparation module is used to distinguish between two motion states and filter out data that is invalid for the fingerprint user mode. LibXtract is a function library for audio feature extraction. The library implements a variety of statistical feature extraction functions, such as variance, mean deviation, skewness, and kurtosis, etc. These functions can share calculation results to improve efficiency, reduce repeated calculations, improve processing speed, and at the same time increase the flexibility of the library, allowing developers to freely combine different feature extraction functions according to needs. Considering the generality of the extraction functions in the library, the system uses LibXtract to extract features. In the case of local identity authentication, authentication is performed using Libsvm for iOS and AndroidLibsvm for Android. In the case of server-side identity authentication, the collected samples are uploaded to the cloud server for judgment, and the cloud server returns the judged similarity. If the similarity is lower than the set threshold, it can be considered that the actual user does not match the owner's identity. When the output of the owner authentication model is the similarity between the user of the mobile device to be authenticated and the owner, if the similarity is greater than the threshold, the authentication result is that the identity matches the owner and the authentication is successful; otherwise, the authentication result is that the identity does not match the owner and the authentication fails.

[0131] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0132] The above-described embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A mobile device owner authentication method based on denoising and optimizing the time-domain convolutional network, characterized in that The mobile device owner authentication method based on denoising and optimizing the temporal convolutional network includes: Obtain a training sample set, where the training sample set contains multiple training samples, and each training sample covers the feature data used by the mobile device during the owner authentication process and the label corresponding to the feature data; Denoise the training sample set by using the joint denoising method, where the joint denoising method includes generating signal, signal, signal and signal for the training sample set by using ensemble empirical mode decomposition, and extracting the detailed parts of signal and signal by using the criterion. At the same time, use the Fast-ICA algorithm to perform noise reduction and unmixing on the detailed parts of signal, signal, signal and signal to obtain component, component, component and component. Apply wavelet threshold denoising to the component to obtain the denoised training sample set; Use the denoised training sample set to optimize and update the optimized temporal convolutional network to obtain an owner authentication model, where the optimized temporal convolutional network includes a self-attention layer; Obtain the feature data of the mobile device user to be authenticated as the input of the optimized and updated owner authentication model, and obtain the authentication result of whether the mobile device user to be authenticated is the owner based on the output of the owner authentication model; The structure of the optimized temporal convolutional network includes an input layer, a temporal convolutional layer, a self-attention layer, and a fully connected layer connected in sequence; the structure of the temporal convolutional layer includes multiple residual blocks, and a convolutional layer and a ReLU activation function are added after every two residual blocks; the residual block includes a first branch and a second branch, the first branch sequentially includes a convolutional layer, a ReLU activation function, and a convolutional layer, the second branch includes a downsampling layer, and the outputs of the first branch and the second branch are added as the output of the residual block; the calculation process of the attention value of the self-attention layer is as follows: Calculate the correlation between the query vector and the key vector: ； In the formula, represents the query vector, represents the -th element value of the key vector, represents the correlation between the query vector and the -th element value of the key vector; Calculate the weight of the value vector: ； In the formula, represents the weight of the -th element in the value vector, represents the SoftMax function, is the dimension of the key vector, represents the correlation between the query vector and the -th element value of the key vector; Calculate the attention value: ； In the formula, is the attention value, represents the -th element value of the value vector.

2. The mobile device owner authentication method based on denoising and optimizing time-domain convolutional network according to claim 1, characterized in that, The obtaining of the training sample set includes: When the mobile device is in the owner authentication process, collect the three-axis data of the acceleration sensor, the three-axis data of the gyroscope, and the three-axis data of the gravity sensor; Add a label to each collected data point, and the label includes the user number, sensor type, axis type, and timestamp when the data point is collected.

3. The mobile device owner authentication method based on denoising and optimizing the time-domain convolutional network according to claim 2, wherein, The denoising of the training sample set using the joint denoising method includes: Take the three-axis data of the acceleration sensor, the three-axis data of the gyroscope, and the three-axis data of the gravity sensor in the training sample set to obtain a total of nine-axis time-series data, and use the joint denoising method to denoise the time-series data of each axis.

4. The mobile device owner authentication method based on denoising and optimizing the time-domain convolutional network according to claim 1, wherein, The owner authentication model runs on the mobile device for local owner authentication, or the owner authentication model is deployed on the server side to perform cloud owner authentication based on the feature data uploaded by the mobile device.

5. The mobile device owner authentication method based on denoising and optimizing the time-domain convolutional network according to claim 1, characterized in that, The output of the owner authentication model is the similarity between the mobile device user to be authenticated and the owner. If the similarity is greater than the threshold, the authentication result is that the identity matches the owner and the authentication is successful; otherwise, the authentication result is that the identity does not match the owner and the authentication fails.

Images