Cross-time persistent identity authentication method and related apparatus
By constructing a continuous identity authentication method across time and utilizing interactive behavior data and feature extraction models, the stability problem of the authentication model over time is solved, achieving continuous identity authentication for device operators, improving the stability and robustness of authentication, and reducing the burden of repeated user registration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2026-06-17
- Publication Date
- 2026-07-14
AI Technical Summary
In existing technologies, the authentication model formed during the initial registration phase is difficult to maintain in the long term, resulting in a continuous decrease in authentication stability over time, and failing to provide continuous and stable identity authentication during device use.
A cross-time continuous identity authentication method is adopted. By acquiring current interaction behavior data, constructing current interaction vectors in various dimensions, calling dimensional and global feature extraction models to extract features, and combining them with periodic model training, the domain discriminability is reduced, and the method adapts to changes in user behavior and devices.
It enables continuous authentication of device operators across time in complex and open scenarios, improves the stability and robustness of features, reduces the burden of repeated user registration, and is suitable for long-term authentication in real-world environments.
Smart Images
Figure CN122389016A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing and relates to a method and related apparatus for continuous identity authentication across time. Background Technology
[0002] With the rapid proliferation of sensitive mobile applications such as mobile payments, instant messaging, and privacy data storage, smartphones and other mobile devices have become core terminals carrying important personal information and critical operations. How to continuously and reliably verify the identity of the current user throughout the entire use of such devices has become a crucial issue in mobile terminal security. Traditional authentication methods such as fingerprint and facial recognition typically perform one-time verification only when the device is unlocked or accessed for a specific application, making it difficult to provide continuous protection during subsequent use. When the device is temporarily transferred, a session is switched, or there is unauthorized brief contact, existing one-time verification mechanisms often fail to detect changes in identity in a timely manner, thus creating security vulnerabilities.
[0003] Currently, persistent identity authentication based on behavioral biometrics is gaining increasing attention. These methods analyze continuous interaction features such as touchscreen interactions and device movement during device use to achieve low-interference and implicit persistent identity recognition. However, most of these solutions still follow a direct deployment approach after registration. This involves collecting a set of labeled registration data in the initial stage to train an authentication model, assuming the model will remain effective for a considerable period. This implicitly assumes that user behavior is relatively stable over time. However, in real-world usage environments, user habits, usage scenarios, device posture, and physical condition constantly change over time, making it difficult for the authentication model formed in the initial registration stage to remain effective in the long term. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art, where the authentication model formed in the initial registration stage is difficult to maintain for a long time, resulting in a continuous decrease in authentication stability over time, and to provide a cross-time continuous identity authentication method and related apparatus.
[0005] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for continuous identity authentication across time, comprising: acquiring current user interaction behavior data, and constructing current user interaction vectors of various dimensions based on the current user interaction behavior data; calling a preset dimensional feature extraction model for each dimension to extract interaction features of the current user interaction vectors of each dimension; fusing the interaction features of the current user interaction vectors of each dimension to obtain a current fusion result, and calling a preset global feature extraction model to extract features of the current fusion result to obtain current user identity authentication features; calling a preset identity authentication model based on the current user identity authentication features to obtain an identity authentication result; recording the current user interaction behavior data, and periodically training the preset dimensional feature extraction model and the preset global feature extraction model in combination with the recorded user interaction behavior data; wherein, the dimensional feature extraction model is trained with the goal of reducing the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the user interaction vector, and the global feature extraction model is trained with the goal of reducing the domain discriminability of the registration identity authentication features and the user identity authentication features; wherein, the interaction features of the registration interaction vector and the registration identity authentication features are obtained based on the registration interaction behavior data, and the interaction features of the user interaction vector and the user identity authentication features are obtained based on the user interaction behavior data.
[0006] Optionally, the registration interaction data and the usage interaction data both include touch screen interaction data and inertial sensor data; wherein, the touch screen interaction data includes touch point coordinates and touch point timestamps; and the inertial sensor data includes the three-axis data and timestamps of the inertial sensor.
[0007] Optionally, the step of constructing current usage interaction vectors in each dimension based on current usage interaction behavior data includes: dividing the current usage interaction behavior data on the time axis by a preset window length and sliding step size to obtain several current usage interaction behavior window data at the window level; for each current usage interaction behavior window data, dividing it on the time axis according to the change range of touch point coordinates or time interval threshold to obtain several data units of each current usage interaction behavior window data; based on the several data units of each current usage interaction behavior window data, constructing the intra-unit features and inter-unit features of each data unit of each current usage interaction behavior window data, and performing dimensional segmentation in conjunction with touch screen interaction data modes and inertial sensor data modes to obtain the current interaction vectors of touch screen units, touch screen units, inertial sensor units, and inertial sensor units for each current usage interaction behavior window data.
[0008] Optionally, both the dimensional feature extraction model and the global feature extraction model are constructed using two interconnected fully connected neural networks; the dimensional feature extraction models corresponding to the dimensions within and between touchscreen units in the touchscreen interaction data mode share the network parameters of the first fully connected neural network; and the dimensional feature extraction models corresponding to the dimensions within and between inertial sensor units in the inertial sensor data mode share the network parameters of the first fully connected neural network.
[0009] Optionally, the preset dimensional feature extraction model is obtained through the following training method: constructing a dimensional feature extraction model and a local domain discriminator with a two-layer perceptron structure; acquiring registered interaction behavior data and several usage interaction behavior data to obtain a training dataset; optimizing the local domain discriminator based on the training dataset with the goal of improving the domain recognition accuracy of the local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data, to obtain an optimized local domain discriminator; and reducing the domain recognition accuracy of the optimized local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data. The objective is to optimize the dimensional feature extraction model to obtain a preset dimensional feature extraction model. The preset global feature extraction model is obtained through the following training method: constructing a global feature extraction model and a global domain discriminator with a two-layer perceptron structure; optimizing the global domain discriminator based on the training dataset with the objective of improving the domain recognition accuracy of the global domain discriminator for registration authentication features and usage authentication features to obtain an optimized global domain discriminator; and optimizing the global feature extraction model with the objective of reducing the domain recognition accuracy of the optimized global domain discriminator for registration authentication features and usage authentication features to obtain a preset global feature extraction model.
[0010] Optionally, the step of fusing the interaction features of the currently used interaction vectors in each dimension to obtain the current fusion result includes: for each dimension, obtaining the distribution difference between the interaction features of the currently used interaction vectors and the interaction features of the registered interaction vectors as the drift metric for each dimension; generating weight coefficients for each dimension based on the drift metrics of each dimension and normalizing them to obtain normalized weight coefficients for each dimension; and weighting and fusing the interaction features of the currently used interaction vectors in each dimension based on the normalized weight coefficients of each dimension to obtain the current fusion result.
[0011] Optionally, the periodic training of the preset dimensional feature extraction model and the preset global feature extraction model, which involves recording the current user interaction behavior data and combining it with the already recorded user interaction behavior data, includes: when the amount of user interaction behavior data that has not participated in the training in the already recorded user interaction behavior data reaches a preset data amount threshold, or when the time since the last training is greater than a preset time threshold, training the preset dimensional feature extraction model and the preset global feature extraction model based on the user interaction behavior data that has not participated in the training, and updating the preset dimensional feature extraction model and the preset global feature extraction model using the trained dimensional feature extraction model and the preset global feature extraction model.
[0012] In a second aspect, the present invention provides a cross-time continuous identity authentication system, comprising: a data processing module, configured to acquire current usage interaction behavior data and construct current usage interaction vectors of various dimensions based on the current usage interaction behavior data; a local feature module, configured to call a preset dimensional feature extraction model for each dimension to extract interaction features of the current usage interaction vectors of each dimension; a global feature module, configured to fuse the interaction features of the current usage interaction vectors of each dimension to obtain a current fusion result, and call a preset global feature extraction model to extract features of the current fusion result to obtain current usage identity authentication features; and an identity authentication module, configured to call a preset identity authentication model based on the current usage identity authentication features to obtain... The system includes an identity authentication result module and a model update module. The model update module records current user interaction data and performs periodic training on a pre-defined dimensional feature extraction model and a pre-defined global feature extraction model, based on the recorded user interaction data. The dimensional feature extraction model is trained to reduce the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the usage interaction vector. The global feature extraction model is trained to reduce the domain discriminability of the registration identity authentication features and the usage identity authentication features. The interaction features of the registration interaction vector and the registration identity authentication features are obtained based on the registration interaction data, and the interaction features of the usage interaction vector and the usage identity authentication features are obtained based on the usage interaction data.
[0013] In a third aspect, the present invention provides a computer device including 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 the cross-time persistent authentication method.
[0014] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the cross-time persistent authentication method.
[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention presents a cross-time continuous identity authentication method. Identity authentication is based on interaction behavior data. During authentication, the method first acquires currently used interaction behavior data and constructs current interaction vectors for each dimension. Then, it calls a pre-defined dimensional feature extraction model for each dimension to extract the interaction features of the current interaction vectors. Next, the interaction features of the current interaction vectors for each dimension are fused to obtain the current fusion result. A pre-defined global feature extraction model is then called to extract the features of the current fusion result, resulting in the current identity authentication features. Finally, a pre-defined identity authentication model is called to obtain the identity authentication result. Because the dimensional feature extraction model is trained with the goal of reducing the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the usage interaction vector, it can more finely characterize the changing patterns of different types of behavior features in cross-time scenarios. Simultaneously, the global feature extraction model is trained with the goal of reducing the domain discriminability of the registration identity authentication features and the usage identity authentication features, which can effectively reduce the distribution shift caused by time variations and device differences, improving the stability and robustness of features while preserving identity discrimination information. Furthermore, the method of this invention also records current user interaction data and combines this data with pre-set dimensional feature extraction models and pre-set global feature extraction models for periodic training. This eliminates the need for users to provide new labeled data in subsequent stages, avoiding the additional burden of repeated user re-registration required in existing technologies. It not only reduces manual intervention costs but also minimizes interference with normal user processes, making it more suitable for long-term, continuous identity authentication scenarios in real-world environments. Through the above design, the method of this invention achieves continuous identity authentication of device operators across time in complex open scenarios and effectively alleviates the problem of identity authentication performance degradation caused by the passage of time and device changes. Attached Figure Description
[0016] Figure 1 This is a flowchart of the cross-time continuous identity authentication method according to an embodiment of the present invention.
[0017] Figure 2 This is a block diagram of the cross-time continuous identity authentication system according to an embodiment of the present invention. Detailed Implementation
[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0020] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 In one embodiment of the present invention, a method for continuous identity authentication across time is provided, which can effectively improve the stability of continuous identity authentication for users in open scenarios under conditions of cross time and cross devices, and reduce the usage burden caused by repeated requests for users to re-register due to performance degradation in the prior art.
[0021] Specifically, the cross-time persistent identity authentication method of the present invention includes the following steps: S1: Obtain the current user interaction behavior data and construct the current user interaction vector for each dimension based on the current user interaction behavior data.
[0022] S2: Call the preset dimensional feature extraction model for each dimension to extract the interaction features of the currently used interaction vector for each dimension.
[0023] S3: Fuse the interaction features of the current interaction vectors in each dimension to obtain the current fusion result, and call the preset global feature extraction model to extract the features of the current fusion result to obtain the current identity authentication features.
[0024] S4: Based on the currently used identity authentication features, call the preset identity authentication model to obtain the identity authentication result.
[0025] S5: Record the current user interaction data, and combine the recorded user interaction data to perform periodic training of the preset dimensional feature extraction model and the preset global feature extraction model.
[0026] The dimensional feature extraction model is trained with the goal of reducing the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the usage interaction vector, while the global feature extraction model is trained with the goal of reducing the domain discriminability of the registration identity authentication features and the usage identity authentication features. The interaction features of the registration interaction vector and the registration identity authentication features are obtained based on the registration interaction behavior data, while the interaction features of the usage interaction vector and the usage identity authentication features are obtained based on the usage interaction behavior data.
[0027] Interpretive, registration interaction data specifically refers to the behavioral data generated by the user's interaction with the device during the registration phase based on registration needs, serving as the registration domain. Usage interaction data specifically refers to the behavioral data generated by the user's interaction with the device during the post-registration usage phase, serving as the usage domain. Based on this, domain discriminability specifically refers to determining whether the corresponding data originates from the registration domain or the usage domain.
[0028] Explanatoryly, the process of obtaining the interaction features and registration authentication features of the registration interaction vector is the same as that of obtaining the interaction features and usage authentication features of the usage interaction vector, except that the data object being processed is changed from usage interaction behavior data to registration interaction behavior data.
[0029] This invention presents a cross-time continuous identity authentication method. Identity authentication is based on interaction behavior data. During authentication, the method first acquires currently used interaction behavior data and constructs current interaction vectors for each dimension. Then, it calls a pre-defined dimensional feature extraction model for each dimension to extract the interaction features of the current interaction vectors. Next, the interaction features of the current interaction vectors for each dimension are fused to obtain the current fusion result. A pre-defined global feature extraction model is then called to extract the features of the current fusion result, resulting in the current identity authentication features. Finally, a pre-defined identity authentication model is called to obtain the identity authentication result. Because the dimensional feature extraction model is trained with the goal of reducing the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the usage interaction vector, it can more finely characterize the changing patterns of different types of behavior features in cross-time scenarios. Simultaneously, the global feature extraction model is trained with the goal of reducing the domain discriminability of the registration identity authentication features and the usage identity authentication features, which can effectively reduce the distribution shift caused by time variations and device differences, improving the stability and robustness of features while preserving identity discrimination information.
[0030] Furthermore, the method of the present invention also records the current user interaction behavior data, and combines the recorded user interaction behavior data to perform periodic training of a preset dimensional feature extraction model and a preset global feature extraction model. This eliminates the need for users to provide new labeled data in subsequent stages, avoiding the additional burden caused by users having to repeatedly re-register in the prior art. It not only reduces the cost of manual participation, but also reduces interference with the normal user process, making it more suitable for long-term and continuous identity authentication scenarios in real-world environments.
[0031] Through the above design, the cross-time continuous identity authentication method of the present invention can realize cross-time continuous identity authentication of device operators in complex open scenarios, and effectively alleviate the problem of identity authentication performance degradation caused by the passage of time and changes in equipment.
[0032] In one possible implementation, both the registration interaction data and the usage interaction data include touchscreen interaction data and inertial sensor data; wherein, the touchscreen interaction data includes touch point coordinates and touch point timestamps; and the inertial sensor data includes triaxial data and timestamps from the inertial sensor.
[0033] In terms of data acquisition, users freely use the device (e.g., a smartphone) in an open, unrestricted environment. The device system collects touchscreen interaction data and inertial sensor data from its built-in inertial sensors in a low-interference manner. Touchscreen interaction data typically includes touch timestamps and touch coordinates (usually two-dimensional coordinates), while inertial sensors typically include accelerometers and gyroscopes.
[0034] For example, touchscreen interaction data may also include additional information such as touch event type (e.g., press, move, and lift), touch pressure, and touch area to enhance behavioral representation capabilities.
[0035] In one possible implementation, constructing current interaction vectors for each dimension based on current interaction behavior data includes: dividing the current interaction behavior data on the time axis by a preset window length and sliding step size to obtain several current interaction behavior window data at the window level; for each current interaction behavior window data, dividing it on the time axis according to the range of change of touch point coordinates or time interval thresholds to obtain several data units for each current interaction behavior window data; based on the several data units of each current interaction behavior window data, constructing the intra-unit features and inter-unit features of each data unit of each current interaction behavior window data, and performing dimensional segmentation by combining the touch screen interaction data mode and the inertial sensor data mode to obtain the current interaction vector within the touch screen unit dimension, the current interaction vector between the touch screen units dimension, the current interaction vector within the inertial sensor unit dimension, and the current interaction vector between the inertial sensor units dimension for each current interaction behavior window data.
[0036] Specifically, before processing the current user interaction data, it can be preprocessed, such as time synchronization, resampling, normalization, and correction of outliers and missing values.
[0037] Interpretive methods use a fixed window length and sliding step size to segment continuous current usage interaction data along the time axis, resulting in several window-level segments, i.e., several current usage interaction window data. Then, the data within each current usage interaction window data is aggregated and encoded to form a window-level structured behavior representation, which is then used for the construction of current usage interaction vectors and feature extraction in subsequent dimensions.
[0038] Specifically, by adopting a sliding window aggregation method, the instantaneous fluctuations of a single interaction and the impact of environmental noise can be reduced, the stability of behavioral representation in cross-time scenarios can be improved, and thus the accuracy of identity authentication can be guaranteed.
[0039] Explaining the process, for each currently used interactive window data, the process begins by using touchscreen interaction data as a baseline and dividing the timeline based on the range of touch point coordinate changes or time interval thresholds. For example, dividing based on the range of touch point coordinate changes could involve dividing the entire screen into several ranges, treating consecutive touch point coordinates within the same range as a single data unit; dividing based on time interval thresholds would involve treating touch point coordinates at regular time intervals as a single data unit. After dividing the touchscreen interaction data, based on time correspondence, touchscreen interaction data and inertial sensor data within the same time period are treated as the same data unit.
[0040] Interpretively, after constructing the data units, the process involves building the current interaction vectors within and between units, using these units as the basic unit. This is based on both touchscreen interaction data and inertial sensor data. Ultimately, through the pairing of modal and information hierarchy dimensions, four dimensions are formed: touchscreen unit-level dimensions, touchscreen unit-level dimensions, inertial sensor unit-level dimensions, and inertial sensor unit-level dimensions, fully exploring the potential characteristics of the data. Specifically, the touchscreen unit-level dimensions describe the trajectory shape, speed changes, and fine-grained temporal patterns within a single touch / swipe; the touchscreen unit-level dimensions describe the time intervals, rhythmic changes, and operation sequence patterns between adjacent interactions; the inertial sensor unit-level dimensions describe the local motion state within a single interaction window; and the inertial sensor unit-level dimensions describe the attitude change trends and long-term motion patterns across interactions / windows. Based on this processing, structured organization of behavioral information from different sources and at different levels can be achieved, providing a foundation for subsequent processing.
[0041] For example, when constructing the current interaction vector within a cell dimension, for touchscreen interaction data, the main considerations are trajectory shape and motion statistics, such as trajectory length, duration, average / maximum velocity, rate of change of direction, acceleration change, and curvature; for inertial sensor data, the main considerations are statistical characteristics, such as mean, variance, peak value, energy, and frequency domain amplitude. Similarly, when constructing the current interaction vector between cells dimension, since the current interaction vector between cells mainly represents the interaction rhythm and cross-interaction changes, for touchscreen interaction data, the main considerations are cell spacing, displacement of the start and end points of the touch trajectory between cells, and the velocity / direction change trend of the touch trajectory between cells; for inertial sensor data, the main considerations are the difference or rate of change of the three-axis data of the corresponding inertial sensor data of adjacent data cells.
[0042] In one possible implementation, both the dimensional feature extraction model and the global feature extraction model are constructed using two interconnected fully connected neural networks; the dimensional feature extraction models corresponding to the dimensions within and between touchscreen units in the touchscreen interaction data mode share the network parameters of the first fully connected neural network; and the dimensional feature extraction models corresponding to the dimensions within and between inertial sensor units in the inertial sensor data mode share the network parameters of the first fully connected neural network.
[0043] Interpretive, a dimensional feature extraction model is constructed for each dimension using a two-layer fully connected neural network. The first fully connected neural network layer is a fully connected layer with batch normalization and non-linear activation functions, used for initial representation learning of the input. The second fully connected neural network layer is a fully connected layer with batch normalization and non-linear activation functions, used for further compression and output of interactive features.
[0044] For example, to facilitate the fusion of interaction features using interaction vectors across different dimensions, the output dimension of the dimensional feature extraction model for each dimension can be set to the same value. Simultaneously, to enhance information interaction within the same modality and reduce model complexity, a shared layer can be set up under the same data type modality. For instance, for touchscreen interaction data modality, the network parameters of the first fully connected neural network layer of the dimensional feature extraction model corresponding to dimensions within and between touchscreen units are shared, while the network parameters of the second fully connected neural network layer remain independent. For inertial sensor data modality, the network parameters of the first fully connected neural network layer of the dimensional feature extraction model corresponding to dimensions within and between inertial sensor units are shared, while the network parameters of the second fully connected neural network layer remain independent. Through this sharing method, a consistent low-level representation can be formed within the same modality, while preserving the differences in high-level semantics between interactions; while maintaining network independence between different modalities to preserve complementary modal information.
[0045] In one possible implementation, the preset dimensional feature extraction model is obtained through the following training method: constructing a dimensional feature extraction model and a local domain discriminator with a two-layer perceptron structure; acquiring registered interaction behavior data and several usage interaction behavior data to obtain a training dataset; optimizing the local domain discriminator based on the training dataset with the goal of improving the domain recognition accuracy of the local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data, to obtain an optimized local domain discriminator; and optimizing the dimensional feature extraction model with the goal of reducing the domain recognition accuracy of the optimized local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data, to obtain the preset dimensional feature extraction model.
[0046] The main idea behind training the dimensional feature extraction model is adversarial feature adjustment, which involves two phases. The first phase trains the local domain discriminator, aiming to improve its domain recognition accuracy for interaction features from both the registration and usage domains. This involves optimizing the local domain discriminator's performance to effectively distinguish whether interaction features belong to the registration or usage domain. The second phase trains the dimensional feature extraction model by introducing gradient inversion / adversarial optimization, such as through gradient inversion layers or equivalent backward gradient processing. This aims to reduce the accuracy of the local domain discriminator in recognizing interaction features from both the registration and usage domains, making it difficult for the discriminator to differentiate between them. This continuously pulls the interaction features towards a representation space less sensitive to time and device changes, achieving fine-grained distribution alignment within each dimension and reducing the risk of overall authentication failure due to severe drift in a single dimension. In the first stage, the model parameters of the fixed-dimensional feature extraction model are fixed, while in the second stage, the model parameters of the local domain discriminator are fixed and optimized.
[0047] For example, the local domain discriminator adopts a lightweight two-layer perceptron structure. The first layer performs a non-linear mapping on the input. Specifically, the structure can be a linear layer followed by a non-linear activation function. The second layer outputs a scalar and obtains the domain probability through activation functions such as Sigmoid, which is used to indicate whether the interaction feature is more likely to come from the data of the registration domain or the data of the usage domain.
[0048] In one possible implementation, the preset global feature extraction model is obtained through the following training method: constructing a global feature extraction model and a global domain discriminator with a two-layer perceptron structure; optimizing the global domain discriminator based on the training dataset with the goal of improving the domain recognition accuracy of the global domain discriminator for registered identity authentication features and used identity authentication features, to obtain an optimized global domain discriminator; and optimizing the global feature extraction model with the goal of reducing the domain recognition accuracy of the optimized global domain discriminator for registered identity authentication features and used identity authentication features, to obtain the preset global feature extraction model.
[0049] Explainingly, the training process of the global feature extraction model is similar to that of the dimensional feature extraction model, with the main difference being the input and output data. The training process for the dimensional feature extraction model can be referenced above. Specifically, in the first stage, the global domain discriminator is optimized to improve its accuracy in recognizing the domains of registration and usage authentication features, enabling it to learn and characterize the overall domain differences in the fusion result. In the second stage, the global feature extraction model is optimized to reduce its accuracy in recognizing the domains of registration and usage authentication features, gradually weakening the domain feature information related to time and / or device changes in the output features. This leads to a more consistent overall distribution of the registration and usage domains at the fusion feature representation level, achieving full domain alignment. This training method allows the global feature extraction model to further suppress residual cross-time / cross-device overall offsets, thereby improving the stability of the usage authentication features during the deployment phase.
[0050] For example, the structure of the global domain discriminator can be set to be the same as that of the local domain discriminator, with the first layer performing nonlinear mapping and the second layer outputting domain probabilities.
[0051] For example, when calling a preset identity authentication model to obtain the identity authentication result based on the currently used identity authentication features, multiple window-level identity authentication results obtained by the same user within a continuous time period can be aggregated, such as the mean or weighted mean, to obtain the user's final identity authentication result.
[0052] In one possible implementation, fusing the interaction features of the currently used interaction vectors in each dimension to obtain the current fusion result includes: for each dimension, obtaining the distribution difference between the interaction features of the currently used interaction vectors and the interaction features of the registered interaction vectors, as a drift metric for each dimension; generating weight coefficients for each dimension based on the drift metrics and normalizing them to obtain normalized weight coefficients for each dimension; and weighting and fusing the interaction features of the currently used interaction vectors in each dimension based on the normalized weight coefficients for each dimension to obtain the current fusion result.
[0053] Explanatoryly, in this implementation, during fusion, the interaction features of the currently used interaction vectors in each dimension are adaptively weighted and fused according to the degree of drift of the interaction features of the currently used interaction vectors in each dimension.
[0054] Specifically, for each dimension, the interaction feature representations of the registration domain and the usage domain on that dimension are first collected, and the distribution difference between the two is calculated as the drift metric for that dimension. The distribution difference can be calculated using the maximum mean difference (MMD), which measures the distance between the set of interaction features of the registration domain and the set of interaction features of the usage domain in the feature space using a preset kernel function, thus obtaining a drift metric characterizing the degree of drift across time and / or across devices for that dimension. Subsequently, corresponding weight coefficients are generated based on the drift metrics for each dimension. For example, the drift metrics can be input into a preset weight generation model to obtain weight scores for each dimension, and the weight scores are normalized to obtain weight coefficients. The purpose is to give dimensions with smaller drift metrics larger weight coefficients. Finally, attention-weighted fusion of the interaction features of each dimension is performed according to the weight coefficients. Specifically, the interaction feature of each dimension is multiplied by its corresponding weight coefficient to obtain a weighted feature, and all weighted features are summed to obtain the fusion result, thereby achieving adaptive weighted fusion based on the degree of drift. This drift-aware fusion method allows feature blocks with higher stability and stronger consistency across time to occupy higher weights in the final representation, thereby improving the robustness of the fusion.
[0055] In one possible implementation, the periodic training of the preset dimensional feature extraction model and the preset global feature extraction model, which involves recording the current user interaction behavior data and combining it with the already recorded user interaction behavior data, includes: when the amount of user interaction behavior data that was not used in the training reaches a preset data amount threshold, or when the time since the last training is greater than a preset time threshold, training the preset dimensional feature extraction model and the preset global feature extraction model based on the user interaction behavior data that was not used in the training, and updating the preset dimensional feature extraction model and the preset global feature extraction model using the trained dimensional feature extraction model and the preset global feature extraction model.
[0056] The training of the pre-defined dimensional feature extraction model and the pre-defined global feature extraction model can refer to the training process of the dimensional feature extraction model and the global feature extraction model described above. Specifically, during training, the model is trained by combining user interaction behavior data and registration interaction behavior data that were not used in the training. This allows the model to gradually reduce the feature distribution difference between the registration domain and the user domain without using domain identity labels, thereby achieving adaptation to drift across time and / or devices.
[0057] For example, during periodic training, to avoid over-adaptation leading to a decline in identity discrimination ability, some parameters of the dimensional feature extraction model and the global feature extraction model can be frozen, or fine-tuned with a smaller learning rate, thereby improving cross-time robustness while maintaining identity discrimination.
[0058] Based on this data accumulation and adaptive update process, the relevant models are continuously adjusted as user behavior evolves, and a consistently high level of identity authentication performance is maintained without requiring users to register repeatedly.
[0059] When the interpretable, pre-defined dimensional feature extraction model and the pre-defined global feature extraction model are deployed for the first time, interactive behavior data is first acquired and user identity labels are used as supervision signals to train the dimensional feature extraction model and the global feature extraction model. This enables the dimensional feature extraction model and the global feature extraction model to learn feature representations with identity discrimination capabilities and authentication decision boundaries, and the trained model parameters are saved for pre-defined purposes.
[0060] Explanatory, pre-defined identity authentication models can adopt existing mature identity authentication model structures and be trained using mature training methods, or they can directly use existing pre-trained identity authentication models.
[0061] In summary, the cross-time persistent authentication method of this invention is essentially a cross-time persistent authentication method based on unlabeled behavioral data, capable of continuously authenticating the current operator of a device in open, unrestricted scenarios. This method establishes a basic model using a small amount of labeled behavioral data from the initial user registration phase, and further adaptively updates the model using unlabeled behavioral data continuously generated during subsequent natural user usage, eliminating the need for users to re-register after authentication performance degrades. Compared to existing technical solutions that rely on a fixed model after a single registration, the method of this invention can more effectively adapt to the distribution shift caused by the evolution of user behavior over time and changes in device.
[0062] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.
[0063] See Figure 2 In another embodiment of the present invention, a time-based persistent identity authentication system is provided, which can be used to implement the above-mentioned time-based persistent identity authentication method. Specifically, the time-based persistent identity authentication system includes a data processing module, a local feature module, a global feature module, an identity authentication module, and a model update module.
[0064] The system comprises the following modules: a data processing module to acquire current user interaction behavior data and construct current user interaction vectors for each dimension; a local feature module to call preset dimensional feature extraction models for each dimension to extract interaction features from the current user interaction vectors; a global feature module to fuse the interaction features from the current user interaction vectors for each dimension to obtain the current fusion result, and call a preset global feature extraction model to extract features from the fusion result to obtain the current user authentication features; an authentication module to call a preset authentication model based on the current user authentication features to obtain the authentication result; and a model update module. This is used to record current user interaction data and to periodically train a preset dimensional feature extraction model and a preset global feature extraction model based on the recorded user interaction data. The dimensional feature extraction model is trained with the goal of reducing the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the usage interaction vector, while the global feature extraction model is trained with the goal of reducing the domain discriminability of the registration authentication features and the usage authentication features. The interaction features of the registration interaction vector and the registration authentication features are obtained based on the registration interaction data, and the interaction features of the usage interaction vector and the usage authentication features are obtained based on the usage interaction data.
[0065] In one possible implementation, both the registration interaction data and the usage interaction data include touchscreen interaction data and inertial sensor data; wherein, the touchscreen interaction data includes touch point coordinates and touch point timestamps; and the inertial sensor data includes triaxial data and timestamps from the inertial sensor.
[0066] In one possible implementation, constructing current interaction vectors for each dimension based on current interaction behavior data includes: dividing the current interaction behavior data on the time axis by a preset window length and sliding step size to obtain several current interaction behavior window data at the window level; for each current interaction behavior window data, dividing it on the time axis according to the range of change of touch point coordinates or time interval thresholds to obtain several data units for each current interaction behavior window data; based on the several data units of each current interaction behavior window data, constructing the intra-unit features and inter-unit features of each data unit of each current interaction behavior window data, and performing dimensional segmentation by combining the touch screen interaction data mode and the inertial sensor data mode to obtain the current interaction vector within the touch screen unit dimension, the current interaction vector between the touch screen units dimension, the current interaction vector within the inertial sensor unit dimension, and the current interaction vector between the inertial sensor units dimension for each current interaction behavior window data.
[0067] In one possible implementation, both the dimensional feature extraction model and the global feature extraction model are constructed using two interconnected fully connected neural networks; the dimensional feature extraction models corresponding to the dimensions within and between touchscreen units in the touchscreen interaction data mode share the network parameters of the first fully connected neural network; and the dimensional feature extraction models corresponding to the dimensions within and between inertial sensor units in the inertial sensor data mode share the network parameters of the first fully connected neural network.
[0068] In one possible implementation, the preset dimensional feature extraction model is obtained through the following training method: constructing a dimensional feature extraction model and a local domain discriminator with a two-layer perceptron structure; acquiring registered interaction behavior data and several usage interaction behavior data to obtain a training dataset; optimizing the local domain discriminator based on the training dataset with the goal of improving the domain recognition accuracy of the local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data, to obtain an optimized local domain discriminator; and reducing the domain recognition accuracy of the optimized local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data. The dimensional feature extraction model is optimized with the goal of improving accuracy to obtain a preset dimensional feature extraction model. The preset global feature extraction model is obtained through the following training method: constructing a global feature extraction model and a global domain discriminator with a two-layer perceptron structure; optimizing the global domain discriminator with the goal of improving the domain recognition accuracy of the global domain discriminator for registration authentication features and usage authentication features based on the training dataset to obtain an optimized global domain discriminator; and optimizing the global feature extraction model with the goal of reducing the domain recognition accuracy of the optimized global domain discriminator for registration authentication features and usage authentication features to obtain a preset global feature extraction model.
[0069] In one possible implementation, fusing the interaction features of the currently used interaction vectors in each dimension to obtain the current fusion result includes: for each dimension, obtaining the distribution difference between the interaction features of the currently used interaction vectors and the interaction features of the registered interaction vectors, as a drift metric for each dimension; generating weight coefficients for each dimension based on the drift metrics and normalizing them to obtain normalized weight coefficients for each dimension; and weighting and fusing the interaction features of the currently used interaction vectors in each dimension based on the normalized weight coefficients for each dimension to obtain the current fusion result.
[0070] In one possible implementation, the periodic training of the preset dimensional feature extraction model and the preset global feature extraction model, which involves recording the current user interaction behavior data and combining it with the already recorded user interaction behavior data, includes: when the amount of user interaction behavior data that was not used in the training reaches a preset data amount threshold, or when the time since the last training is greater than a preset time threshold, training the preset dimensional feature extraction model and the preset global feature extraction model based on the user interaction behavior data that was not used in the training, and updating the preset dimensional feature extraction model and the preset global feature extraction model using the trained dimensional feature extraction model and the preset global feature extraction model.
[0071] All relevant content of each step involved in the aforementioned embodiments of the cross-time persistent identity authentication method can be referenced from the functional description of the corresponding functional module of the cross-time persistent identity authentication system in the embodiments of the present invention, and will not be repeated here.
[0072] 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.
[0073] 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, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), 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 function. The processor described in this embodiment of the present invention can be used for the operation of a continuous authentication method across time.
[0074] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium, which is a memory device in a computer device for storing 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, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a 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 cross-time persistent authentication method in the above embodiments.
[0075] 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.
[0076] 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.
[0077] 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 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0078] 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.
[0079] 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 method for continuous authentication across time, characterized in that, include: Obtain current user interaction data and construct current user interaction vectors for each dimension based on the current user interaction data; Call the preset dimensional feature extraction model for each dimension to extract the interaction features of the currently used interaction vector for each dimension; The interaction features of the current interaction vectors in each dimension are fused to obtain the current fusion result, and the preset global feature extraction model is called to extract the features of the current fusion result to obtain the current identity authentication features. Based on the current identity authentication features used, the preset identity authentication model is invoked to obtain the identity authentication result; Record the current user interaction data, and combine the recorded user interaction data to perform periodic training of the preset dimensional feature extraction model and the preset global feature extraction model; The dimensional feature extraction model is trained with the goal of reducing the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the usage interaction vector, while the global feature extraction model is trained with the goal of reducing the domain discriminability of the registration identity authentication features and the usage identity authentication features. The interaction features of the registration interaction vector and the registration identity authentication features are obtained based on the registration interaction behavior data, while the interaction features of the usage interaction vector and the usage identity authentication features are obtained based on the usage interaction behavior data.
2. The cross-time persistent identity authentication method according to claim 1, characterized in that, The registration interaction data and usage interaction data both include touchscreen interaction data and inertial sensor data; The touchscreen interaction data includes touch point coordinates and touch point timestamps; Inertial sensor data includes triaxial data and timestamps from the inertial sensor.
3. The cross-time persistent identity authentication method according to claim 2, characterized in that, The construction of current user interaction vectors in various dimensions based on current user interaction behavior data includes: The current interactive behavior data is divided on the timeline by a preset window length and sliding step size to obtain several window-level current interactive behavior window data. For each currently used interactive behavior window data, the time axis is divided according to the range of change of touch point coordinates or time interval threshold to obtain several data units of each currently used interactive behavior window data; Based on several data units of each currently used interactive behavior window data, the intra-unit features and inter-unit features of each data unit of each currently used interactive behavior window data are constructed. Dimensional segmentation is then performed by combining the touch screen interaction data mode and the inertial sensor data mode to obtain the current interaction vector within the touch screen unit dimension, the current interaction vector between touch screen units dimension, the current interaction vector within the inertial sensor unit dimension, and the current interaction vector between inertial sensor units dimension of each currently used interactive behavior window data.
4. The cross-time persistent identity authentication method according to claim 3, characterized in that, Both the dimensional feature extraction model and the global feature extraction model are constructed using two interconnected fully connected neural networks. The dimensional feature extraction models corresponding to the dimensions within and between touchscreen units in the touchscreen interactive data mode share the network parameters of the first layer of fully connected neural network. The dimensional feature extraction model corresponding to the intra-inertial sensor unit dimension and inter-inertial sensor unit dimension in the inertial sensor data mode shares the network parameters of the first layer fully connected neural network.
5. The cross-time persistent identity authentication method according to claim 1, characterized in that, The preset dimensional feature extraction model is obtained through the following training method: A dimensional feature extraction model and a local domain discriminator with a two-layer perceptron structure are constructed; registered interaction behavior data and several usage interaction behavior data are acquired to obtain a training dataset; based on the training dataset, the local domain discriminator is optimized with the goal of improving the domain recognition accuracy of the local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data, resulting in an optimized local domain discriminator; and the dimensional feature extraction model is optimized with the goal of reducing the domain recognition accuracy of the optimized local domain discriminator for the interaction features of the registered interaction behavior data and the interaction features of the usage interaction behavior data, resulting in a preset dimensional feature extraction model. The preset global feature extraction model is obtained through the following training method: A global feature extraction model and a two-layer perceptron structure are constructed for the global domain discriminator. Based on the training dataset, the global domain discriminator is optimized to improve the domain recognition accuracy of the global domain discriminator for registered identity authentication features and used identity authentication features, resulting in an optimized global domain discriminator. The global feature extraction model is then optimized to reduce the domain recognition accuracy of the optimized global domain discriminator for registered identity authentication features and used identity authentication features, resulting in a pre-defined global feature extraction model.
6. The cross-time persistent authentication method according to claim 1, characterized in that, The process of fusing the interaction features of each dimension using the current interaction vector to obtain the current fusion result includes: For each dimension, the distribution difference between the interaction features of the currently used interaction vector and the interaction features of the registered interaction vector is obtained as a drift measure for each dimension. The weight coefficients for each dimension are generated based on the drift metric of each dimension and then normalized to obtain the normalized weight coefficients for each dimension. The interaction features of the current interaction vectors for each dimension are then weighted and fused based on the normalized weight coefficients of each dimension to obtain the current fusion result.
7. The cross-time persistent identity authentication method according to claim 1, characterized in that, The periodic training of the recorded user interaction data and the pre-set dimensional feature extraction model and the pre-set global feature extraction model, combined with the recorded user interaction data, includes: When the amount of user interaction data that was not used in training reaches a preset data volume threshold, or when the time since the last training exceeds a preset time threshold, the preset dimensional feature extraction model and the preset global feature extraction model are trained based on the user interaction data that was not used in training. The trained dimensional feature extraction model and the preset global feature extraction model are then used to update the preset dimensional feature extraction model and the preset global feature extraction model.
8. A persistent authentication system across time, characterized in that, include: The data processing module is used to acquire current user interaction behavior data and construct current user interaction vectors in various dimensions based on the current user interaction behavior data. The local feature module is used to call the preset dimensional feature extraction models for each dimension and extract the interaction features of the currently used interaction vector for each dimension. The global feature module is used to fuse the interaction features of the current interaction vectors in each dimension to obtain the current fusion result, and call the preset global feature extraction model to extract the features of the current fusion result to obtain the current identity authentication features. The identity authentication module is used to call a preset identity authentication model to obtain the identity authentication result based on the currently used identity authentication features; The model update module is used to record the current user interaction behavior data and combine the recorded user interaction behavior data to perform periodic training of the preset dimensional feature extraction model and the preset global feature extraction model. The dimensional feature extraction model is trained with the goal of reducing the domain discriminability of the interaction features of the registration interaction vector and the interaction features of the usage interaction vector, while the global feature extraction model is trained with the goal of reducing the domain discriminability of the registration identity authentication features and the usage identity authentication features. The interaction features of the registration interaction vector and the registration identity authentication features are obtained based on the registration interaction behavior data, while the interaction features of the usage interaction vector and the usage identity authentication features are obtained based on the usage interaction behavior data.
9. 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 time-based persistent authentication method as described in any one of claims 1 to 7.
10. 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 time-based persistent authentication method as described in any one of claims 1 to 7.