Continuous identity authentication method based on incremental learning and meta-learning and related device

Abstract

A continuous identity authentication method based on incremental learning and meta-learning and related devices, comprising: visualizing sensor data and determining whether the sensor data has a drift phenomenon; constructing an authentication framework based on meta-learning and incremental learning, including an offline registration phase and an online authentication phase; obtaining features of each sensor time dimension and their correlation features between different dimensions according to the offline registration; learning from the obtained data according to the online authentication phase, and performing online model updating; and testing and verifying the updated model to obtain the model capability. The identity authentication technology framework MetaAuth based on touch screen behavior solves the long-term touch screen identity authentication problem, and an online updating mechanism AMUM based on meta-learning is designed for the incremental learning problem in the long-term continuous identity authentication scenario, so as to improve the stability of the model.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and specifically to a continuous identity authentication method and related apparatus based on incremental learning and meta-learning. Background Technology

[0002] In today's society, mobile phones and tablets have become the most basic tools for social interaction, entertainment, communication, and e-commerce, and are an indispensable part of modern life. Currently, 83.72% of people worldwide own smartphones, and with the continuous development of technology, the computing and storage capabilities of smartphones are constantly being expanded, leading to the development and use of more and more apps, such as those related to social networking, online shopping, and financial lending. Statistics show that nearly 60% of internet users aged 16 to 64 have shopped online, with annual online shopping spending currently around 3.8 trillion yuan, an average annual growth rate of 18%, demonstrating that an increasing number of payment transactions are being conducted on mobile devices.

[0003] The high dependence on smartphones has led to widespread concern about the privacy of the data stored on these devices. Normally, this data is protected and accessible only to the user. However, data shows that nearly one-third of users have had their phones stolen, and one-tenth of these users believe their privacy has been violated to varying degrees. Therefore, mobile phone privacy leaks, caused by various reasons, have become a major problem troubling all users and mobile phone manufacturers.

[0004] To overcome this challenge and address mobile phone privacy issues, protecting private mobile data, numerous identity authentication problems have been addressed. Mobile phone identity authentication primarily refers to verifying a user's identity using a short, identifiable identifier on the mobile device. This allows the user to either continue access or deny access. Over the years, research has identified three main authentication methods: knowledge-based authentication, biometric authentication, and behavioral authentication. Knowledge-based authentication is a traditional and widely accepted method, with representative examples including numeric passwords and pattern unlocking. These methods are popular due to their simplicity and low learning curve. However, research indicates that complex passwords significantly increase the user's memory burden, while simple passwords like "abc" or "12345678" may lack security. Furthermore, frequent mobile phone use allows users to potentially use screen smudges or skin oils to decipher passwords, highlighting the significant vulnerabilities of traditional methods. Summary of the Invention

[0005] The purpose of this invention is to provide a continuous identity authentication method and related apparatus based on incremental learning and meta-learning, so as to solve the security problems and high memory costs of traditional methods.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Continuous identity authentication methods based on incremental learning and meta-learning include: Collect sensor data from the phone during use, visualize the sensor data, and determine if there is any drift in the sensor data. If drift occurs, construct the MetaAuth authentication framework, which includes an offline registration phase and an online authentication phase, based on meta-learning and incremental learning. A meta-learning model is established, which acquires the time dimension features of each sensor and the correlation features between their different dimensions based on offline registration. Based on the online authentication stage, the model learns from the acquired data and updates the model online. The updated model was tested and verified to obtain its capabilities.

[0007] Furthermore, sensor data collected during phone use is visualized to determine if there is any data drift. Sensor data from users' mobile phones during a random period of time is collected. This period can be divided into early, middle, and late stages. Feature extraction is then performed on the data. After feature extraction, the features are visualized in a two-dimensional space using the t-SNE method, which shows the similarity and changes in features between users. When the feature data between users changes and the distance between them increases, a drift phenomenon occurs.

[0008] Further, the offline registration phase: Data collection: Data is collected by selecting user touch screen behavior. Data collected by the sensor when the screen is on is used as positive samples, and data from other users is used as negative samples for training. Data preprocessing stage: First, a fixed-length sliding window is used to divide the data into a series of time periods that will repeat, which are used as the basic unit of authentication. Then, each time period is further divided into non-overlapping small time periods, which are called time sources. Local features are then extracted from the time sources. Continuous data acquisition part: Build a dual-channel deep learning model 2-GCTN model to extract the features of each sensor in the time dimension and the correlation features between their different dimensions.

[0009] Further, the online verification stage: In this stage, an improved meta-learning method, AMUM, is used for online model updates. It consists of two parts: 1. a drift detector; 2. an improved MAML update strategy. During the detection process, the detection time window can be changed according to changes in time and data, and a drift signal will be output based on the prediction results within the window. If the probability of a line error exceeds the set threshold, a drift judgment will be given.

[0010] Further details of the process: First, experimental data was acquired, including three-dimensional data from sensors such as accelerometers, gyroscopes, and magnetometers. Establish evaluation metrics: Use metrics to measure the accuracy of identity authentication in various scenarios. Metrics include: Equal error rate: The false acceptance rate is equal to the false rejection rate. The equal error rate can be affected by changing the threshold of the predicted score. False acceptance rate: defined as the ratio of the number of falsely accepted illegal samples to the total number of all illegal test samples; a high false acceptance rate indicates a high rate of false separation of intruder samples; False rejection rate: defined as the number of valid samples that are falsely rejected out of the total number of valid test samples; a high false rejection rate indicates poor identification of legitimate user samples. Authentication accuracy variance: a measure of variance that represents the dispersion of authentication results over a long period of time; it is calculated by averaging the squares of the differences between the equal error rate and the average of the equal error rates for each authentication; low variance reflects the stable predictive ability of the authentication model. Memory usage: Indicates the amount of memory used for model training and updates; Time consumption: This refers to the time spent training and updating the model.

[0011] Further experimental verification: The authentication framework 2-GCTN is evaluated using M users in the dataset. The model is trained and incrementally tested individually for each user. Each user's mobile phone is designated as the legitimate user, while the remaining M-1 users are designated as illegitimate users. For positive samples, sensor data is selected from each user in chronological order and divided into 20 parts, each corresponding to one of the 20 incremental learning tasks. Negative samples are randomly selected from the remaining M-1 users. During training, each training and testing cycle is defined as a task. The next data block is first used to test the model before training and updating it.

[0012] Further, specific steps: Sort all processed data by time and divide it into identical data blocks T; Each data block will be a user's positive sample dataset, with T-1 tasks, each task containing data D from two periods. i and D i+1 ; Use D i Update the model using the training set, and then use D i+1 Test its capabilities; Repeat the above steps for each user.

[0013] Furthermore, a continuous identity authentication system based on incremental learning and meta-learning includes: The data processing and judgment module is used to collect sensor data during the use of the mobile phone, visualize the sensor data, and determine whether there is any drift in the sensor data. The framework building module is used to construct the MetaAuth authentication framework, which includes an offline registration phase and an online authentication phase, based on meta-learning and incremental learning, in the event of drifting. The model building module is used to build a meta-learning model. It obtains the features of each sensor in the time dimension and the correlation features between their different dimensions based on offline registration. Based on the online authentication stage, it learns from the acquired data and performs online model updates. The validation module is used to test and validate the updated model to obtain its capabilities.

[0014] Furthermore, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a continuous authentication method based on incremental learning and meta-learning.

[0015] Furthermore, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a continuous identity authentication method based on incremental learning and meta-learning.

[0016] Compared with the prior art, the present invention has the following technical effects: This invention, based on deep learning and meta-learning, does not add any extra constraints during the data collection process, thus fully reflecting the user's actual mobile phone usage. Secondly, a dual-channel deep learning model, 2-GCTN, is designed to utilize data correlation information between multiple sensors. This model can acquire data correlation information between different dimensions and significantly improve the ability to represent user behavior under multi-sensor data. Finally, an incremental learning mechanism, AMUM, based on meta-learning technology, is proposed. This mechanism can use the knowledge learned by the model and new data for real-time updates, fully demonstrating the model's effectiveness. The MetaAuth identity authentication framework based on touchscreen behavior solves the challenges in the long-term touchscreen identity authentication field. Furthermore, the meta-learning-based online update mechanism AMUM is designed for incremental learning in scenarios involving long-term continuous identity authentication, thereby improving model stability. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the sensor data of the present invention.

[0018] Figure 2 This is a schematic diagram of the authentication framework based on meta-learning and incremental learning of the present invention.

[0019] Figure 3 This is a schematic diagram illustrating the long-term certification process of this invention.

[0020] Figure 4 The figure shows the impact of meta-learning mechanisms on long-term model validation.

[0021] Figure 5 This is an experimental diagram showing the use of adaptive-MAML as an incremental update mechanism in MetaAuth.

[0022] Figure 6A comparison chart showing the authentication capabilities of different models using sensor data. Detailed Implementation

[0023] The implementation methods and experimental steps of the invention described in this paper will be described in detail below: Step 1: First, visualize the sensor data. We randomly selected 5 users and collected their sensor data from the Taobao app. The total time was divided into 2 months, which can also be divided into early, middle and late stages.

[0024] The second step: After feature extraction, the features are visualized in two-dimensional space using the (t-SNE) method. This shows the similarity and variation of features among the five users. Figure 1 It's clear from the data that in the early stages, the feature data of users 3, 4, and 5 were quite similar, with user 5 being significantly different from user 1. In the mid-stage, the original features remained largely unchanged, but the distance between users 1 and 4 increased. In the later stages, the distance between users 3, 4, and 5 also widened considerably. Therefore, it can be seen that user behavior data will inevitably change over time, exhibiting a drift phenomenon.

[0025] Step 3: To address the performance degradation issue in long-term authentication, the MetaAuth authentication framework, based on meta-learning and incremental learning, was proposed. The overall framework is as follows: Figure 2 As shown, it contains two important components: an offline registration section and an online authentication section.

[0026] The offline registration process mainly consists of three steps: 1. Data collection phase: The main approach is to collect data through touch screen behavior. Data collected by the sensor when the screen is on is used as positive samples, while data from other users is used as negative samples for training.

[0027] 2. Data preprocessing stage: In order to effectively extract the features of time series data, a fixed-length sliding window is first used to divide the data into a series of time periods that will repeat, which are used as the basic unit for authentication. Then, each time period is further divided into non-overlapping smaller time periods, which are called time sources. Local features can then be extracted from the time sources.

[0028] 3. Continuous data acquisition component: This requires providing more accurate representation of time-series data scenarios. Therefore, a novel 2-GCTN model was designed, capable of extracting features from each sensor's time dimension and the correlation features between their different dimensions.

[0029] During the online authentication phase: The authenticator can continuously learn from new data and improve previous methods for online model updates. The ADWIN method can be used as a drift detector to reduce resource consumption on mobile devices and improve model updates. Next, a meta-learning model will be designed to learn the model's initialization parameters and the learning rate used during training.

[0030] This stage uses the improved meta-learning method AMUM for online model updates. It mainly consists of two parts: 1. A drift detector; 2. An improved MAML update strategy. Drift detection technology can be used to improve the efficiency and rationality of model updates.

[0031] During the detection process, the detection time window can be adjusted according to changes in time and data. A drift signal is also output based on the prediction results within the window. If the probability of a line misalignment exceeds a set threshold, a drift judgment is given. Furthermore, the MAML model can improve its efficiency by continuously optimizing the initial parameters of the base model during task learning.

[0032] Specific experimental procedure: Step 1: First, acquire experimental data, mainly including three-dimensional data from sensors such as accelerometers, gyroscopes, and magnetometers.

[0033] Step 2: Establish evaluation metrics: The accuracy of identity authentication in various scenarios can be measured using the following metrics, the main ones being: 1. Equal Error Rate (EER): The false acceptance rate (FAR) is equal to the false rejection rate (FRR). The equal error rate can be affected by changing the threshold of the predicted score.

[0034] 2. False Acceptance Rate (FAR): Defined as the number of falsely accepted illegal samples out of the total number of illegal test samples. A higher FAR indicates a higher false separation rate of intruder samples.

[0035] 3. False Rejection Rate (FRR): Defined as the ratio of the number of valid samples falsely rejected to the total number of valid test samples. A higher FRR indicates poorer identification of legitimate user samples.

[0036] 4. Certification Accuracy Variance: A measure of variance representing the dispersion of certification results across different time periods over a long period. It is calculated by averaging the squares of the differences between each certification's EER and the mean EER. Lower variance reflects the stable predictive ability of the certification model.

[0037] 5. Memory Usage: Indicates the amount of memory used for model training and updates. Lower memory usage reflects less consumption of smartphone resources.

[0038] 6. Time Consumption: This indicates the time spent training and updating the model. Less training time means faster model updates.

[0039] Step 3: Conduct the experiment. In the experiment, the 2-GCTN authentication framework was evaluated using 40 users from the dataset. The model was trained and incrementally tested individually for each user. Each user's mobile phone was designated as the legitimate user, while the remaining 39 users were designated as illegitimate users. For positive samples, sensor data was selected from each user in chronological order and divided into 20 parts, each corresponding to one of the 20 incremental learning tasks. Negative samples were randomly selected from the remaining 39 users. During training, each training and testing phase can be defined as a task. Therefore, the next data block is first used to test the model before training and updating it.

[0040] Specific steps 1. Sort all processed data according to time and divide it into identical data blocks T.

[0041] Each data block will be a positive sample dataset from the user. Therefore, there will be T-1 tasks, each task Taski containing data from two periods, Di and Di+1.

[0042] 2. Update the model using Di as the training set, and then test its capabilities using Di+1.

[0043] 3. Repeat steps 1 and 2 for each user.

[0044] In this experiment, a systematic experiment was conducted to analyze the model's capabilities, effectiveness, and long-term certification ability.

[0045] First, we analyze the issue of model accuracy degradation under long-term user usage. For example... Figure 3 As shown, during the long-term authentication process, LSTM, SVM, one-SVM, and 2-GCTN models were used, and the accuracy of all four models decreased. The main reason for this is that the increased environmental complexity led to changes in user behavior when using mobile phones, which in turn caused changes in data distribution. Secondly, while LSTM and 2-GCTN models, which utilize time-dependent characteristics, initially showed good performance, their accuracy decreased significantly compared to the remaining models as the data distribution changed.

[0046] Then, the impact of the meta-learning mechanism on the long-term validation of the model was analyzed. Experimental results are as follows: Figure 4 As shown MetaAuth represents model updates performed using the AMUM mechanism proposed in this paper. The figure shows that when there is no... When using the Adaptive-MAML mechanism, the accuracy of the long-term continuous authentication model decreases, and the model update time increases. This is because the meta-learning mechanism can obtain better initialization parameters, allowing the model to learn more effectively. Therefore, it can be seen that the improved meta-learning mechanism can improve the accuracy and efficiency of the model in long-term identity authentication.

[0047] Next, we discuss the effectiveness of the drift detection mechanism AMUM. The new incremental update mechanism AMUM primarily uses the ADWIN method to detect drift in new batches of data, and then selects whether to update the model based on the detection results. This allows the MetaAuth structure to not only update the model at a reasonable frequency but also become more efficient. We compare the model capabilities of AMUM using traditional AMUM in combination with 2-GCTN, namely Retrain-2-GCTN and MetaAuth. In MetaAuth, adaptive-MAML is used as the incremental update mechanism, updating the model only with new data. Experimental results are as follows... Figure 5 As shown in the diagram, the experiment demonstrates that MetaAuth has a very low EER, indicating high average authentication accuracy and good model stability. This is because the model can utilize information from historical data through meta-learning, allowing it to better learn user behavioral characteristics. Furthermore, MetaAuth has a relatively short update time and shows a significant improvement in accumulator accuracy. Therefore, it can be concluded that MetaAuth is more practical for real-world mobile scenarios.

[0048] Finally, we discuss long-term authentication capabilities. Experiments compared the authentication capabilities of different models using sensor data. For example... Figure 6 As shown, compared to conventional methods, the proposed model achieves better accuracy (i.e., the smallest EER) under long-term incremental updates, while MetaAuth also exhibits the best stability. Furthermore, the violin plot of memory usage clearly shows that MetaAuth has the lowest memory usage, indicating minimal resource consumption during the authentication process. Finally, it can also be seen that MetaAuth has the shortest training time among all models, demonstrating its real-time performance and making it the most suitable model for real-world scenarios.

[0049] In another embodiment of the present invention, a continuous identity authentication system based on incremental learning and meta-learning is provided, which can be used to implement the above-mentioned continuous identity authentication method based on incremental learning and meta-learning. Specifically, the system includes: The data processing and judgment module is used to collect sensor data during the use of the mobile phone, visualize the sensor data, and determine whether there is any drift in the sensor data. The framework building module is used to construct the MetaAuth authentication framework, which includes an offline registration phase and an online authentication phase, based on meta-learning and incremental learning, in the event of drifting. The model building module is used to build a meta-learning model. It obtains the features of each sensor in the time dimension and the correlation features between their different dimensions based on offline registration. Based on the online authentication stage, it learns from the acquired data and performs online model updates. The validation module is used to test and validate the updated model to obtain its capabilities.

[0050] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0051] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a continuous authentication method based on incremental learning and meta-learning.

[0052] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the continuous authentication method based on incremental learning and meta-learning in the above embodiments.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A continuous identity authentication method based on incremental learning and meta-learning, characterized in that, include: Collect sensor data from the phone during use, visualize the sensor data, and determine if there is any drift in the sensor data. If drift occurs, construct the MetaAuth authentication framework, which includes an offline registration phase and an online authentication phase, based on meta-learning and incremental learning. A meta-learning model is established, which acquires the time dimension features of each sensor and the correlation features between their different dimensions based on offline registration. Based on the online authentication stage, the model learns from the acquired data and updates the model online. The updated model is tested and verified to obtain its capabilities; Offline registration phase: Data collection: Data is collected by selecting user touch screen behavior. Data collected by the sensor when the screen is on is used as positive samples, and data from other users is used as negative samples for training. Data preprocessing stage: First, a fixed-length sliding window is used to divide the data into a series of time periods that will repeat, which are used as the basic unit of authentication. Then, each time period is further divided into non-overlapping small time periods, which are called time sources. Local features are then extracted from the time sources. Continuous data acquisition part: Build a dual-channel deep learning model 2-GCTN model to extract the features of each sensor in the time dimension and the correlation features between their different dimensions; Online verification stage: In this stage, the improved meta-learning method AMUM is used for online model updates, which consists of two parts:

1. a drift detector; 2. an improved MAML update strategy. During the detection process, the detection time window can be changed according to the changes in time and data, and a drift signal will be output based on the prediction results within the window. If the probability of a line error exceeds the set threshold, a judgment that a drift has occurred will be given. Specific process: First, experimental data was acquired, including three-dimensional data from sensors such as accelerometers, gyroscopes, and magnetometers. Establish evaluation metrics: Use metrics to measure the accuracy of identity authentication in various scenarios. Metrics include: Equal error rate: The false acceptance rate is equal to the false rejection rate. The equal error rate can be affected by changing the threshold of the predicted score. False acceptance rate: defined as the ratio of the number of falsely accepted illegal samples to the total number of all illegal test samples; a high false acceptance rate indicates a high rate of false separation of intruder samples; False rejection rate: defined as the number of valid samples that are falsely rejected out of the total number of valid test samples; a high false rejection rate indicates poor identification of legitimate user samples. Authentication accuracy variance: a measure of variance that represents the dispersion of authentication results over a long period of time; it is calculated by averaging the squares of the differences between the equal error rate and the average of the equal error rates for each authentication; low variance reflects the stable predictive ability of the authentication model. Memory usage: Indicates the amount of memory used for model training and updates; Time consumption: This refers to the time spent training and updating the model.

2. The continuous identity authentication method based on incremental learning and meta-learning according to claim 1, characterized in that, Collect sensor data from the phone during use, visualize the sensor data, and determine if there is any data drift. Sensor data from users' mobile phones during a random period of time is collected. This period can be divided into early, middle, and late stages. Feature extraction is then performed on the data. After feature extraction, the features are visualized in a two-dimensional space using the t-SNE method, which shows the similarity and changes in features between users. When the feature data between users changes and the distance between them increases, a drift phenomenon occurs.

3. The continuous identity authentication method based on incremental learning and meta-learning according to claim 1, characterized in that, Experimental verification: The authentication framework 2-GCTN is evaluated using M users in the dataset. The model is trained and incrementally tested for each user individually. Each user's corresponding mobile phone is set as the legitimate user, while the other M-1 mobile phones are set as illegitimate users. For the positive sample data, sensor data is selected from each user in chronological order and the data is divided into 20 parts, each corresponding to 20 incremental learning tasks. Negative samples are randomly selected from the remaining M-1 users. During the training process, each training and testing session is defined as a task. The data block of the next cycle is first used to test the model, and then the model is trained and updated.

4. The continuous identity authentication method based on incremental learning and meta-learning according to claim 3, characterized in that, Specific steps: Sort all processed data by time and divide it into identical data blocks T; Each data block will be a user's positive sample dataset, with T-1 tasks, each task containing data D from two periods. i and D i+1 ; Use D i Update the model using the training set, and then use D i+1 Test its capabilities; Repeat the above steps for each user.

5. A continuous identity authentication system based on incremental learning and meta-learning, characterized in that, include: The data processing and judgment module is used to collect sensor data during the use of the mobile phone, visualize the sensor data, and determine whether there is any drift in the sensor data. The framework building module is used to construct the MetaAuth authentication framework, which includes an offline registration phase and an online authentication phase, based on meta-learning and incremental learning, in the event of drifting. The model building module is used to build a meta-learning model. It obtains the features of each sensor in the time dimension and the correlation features between their different dimensions based on offline registration. Based on the online authentication stage, it learns from the acquired data and performs online model updates. The validation module is used to test and validate the updated model to determine its capabilities. Offline registration phase: Data collection: Data is collected by selecting user touch screen behavior. Data collected by the sensor when the screen is on is used as positive samples, and data from other users is used as negative samples for training. Data preprocessing stage: First, a fixed-length sliding window is used to divide the data into a series of time periods that will repeat, which are used as the basic unit of authentication. Then, each time period is further divided into non-overlapping small time periods, which are called time sources. Local features are then extracted from the time sources. Continuous data acquisition part: Build a dual-channel deep learning model 2-GCTN model to extract the features of each sensor in the time dimension and the correlation features between their different dimensions; Online verification stage: In this stage, the improved meta-learning method AMUM is used for online model updates, which consists of two parts:

1. a drift detector; 2. an improved MAML update strategy. During the detection process, the detection time window can be changed according to the changes in time and data, and a drift signal will be output based on the prediction results within the window. If the probability of a line error exceeds the set threshold, a judgment that a drift has occurred will be given. Specific process: First, experimental data was acquired, including three-dimensional data from sensors such as accelerometers, gyroscopes, and magnetometers. Establish evaluation metrics: Use metrics to measure the accuracy of identity authentication in various scenarios. Metrics include: Equal error rate: The false acceptance rate is equal to the false rejection rate. The equal error rate can be affected by changing the threshold of the predicted score. False acceptance rate: defined as the ratio of the number of falsely accepted illegal samples to the total number of all illegal test samples; a high false acceptance rate indicates a high rate of false separation of intruder samples; False rejection rate: defined as the number of valid samples that are falsely rejected out of the total number of valid test samples; a high false rejection rate indicates poor identification of legitimate user samples. Authentication accuracy variance: a measure of variance that represents the dispersion of authentication results over a long period of time; it is calculated by averaging the squares of the differences between the equal error rate and the average of the equal error rates for each authentication; low variance reflects the stable predictive ability of the authentication model. Memory usage: Indicates the amount of memory used for model training and updates; Time consumption: This refers to the time spent training and updating the model.

6. A computer 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 computer program, it implements the steps of the continuous identity authentication method based on incremental learning and meta-learning as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the continuous identity authentication method based on incremental learning and meta-learning as described in any one of claims 1 to 4.

Images