A behavior-adaptive continuous authentication method and system
By using component attribution analysis and parameter search adaptive methods, the continuous authentication model of mobile devices is deconstructed and optimized, solving the problem of decreased authentication accuracy caused by changes in user behavior patterns, and achieving efficient and privacy-preserving adaptive authentication.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV HANGZHOU RES INST
- Filing Date
- 2025-12-24
- Publication Date
- 2026-05-26
AI Technical Summary
Existing continuous authentication technologies on mobile devices suffer from decreased authentication accuracy due to changes in user behavior over time. Furthermore, existing solutions are computationally expensive or violate privacy principles, and cannot effectively adapt to changes in user behavior patterns.
By employing component attribution analysis and parameter search adaptive methods, the basic continuous identity authentication model is deconstructed into multiple computational units. Through ablation experiments and Bayesian optimization search, an adaptive authentication model is generated to adapt to changes in user behavior, improve authentication accuracy, and reduce computational overhead.
It achieves high-precision continuous identity authentication on mobile devices, protects user privacy, reduces computational overhead, and is suitable for resource-constrained edge devices.
Smart Images
Figure CN121389091B_ABST
Abstract
Description
Technical Field
[0001] This invention pertains to mobile device security and biometric identification, specifically relating to a behavior-adaptive continuous identity authentication method and system. Background Technology
[0002] Existing continuous authentication technologies passively monitor users' touch gestures, key dynamics, and movement patterns to continuously verify user identity after device unlocking, providing more seamless and robust security than traditional one-time authentication (such as passwords, fingerprints, and facial recognition). Current mainstream CA (Continuous Authentication) systems are typically built on deep learning models (such as CNNs, LSTMs, and Transformers), which perform well on initial training datasets. However, users' biometric characteristics evolve naturally over time, causing CA models pre-trained on historical data to experience a significant drop in authentication accuracy when faced with future user behavior data. Existing solutions to this problem are not suitable for mobile devices, for example:
[0003] 1. Model retraining / fine-tuning: This requires updating the entire model weights through backpropagation on new data, which is computationally expensive and impractical for mobile devices.
[0004] 2. Continual / Incremental Learning: This usually requires uploading new user data to a central server for model updates, which violates the privacy principle that biometric data must be kept locally; or it requires retaining historical data locally, which places a heavy storage burden on resource-constrained mobile devices.
[0005] Therefore, there is an urgent need for a continuous authentication technology that can adapt to changes in user behavior patterns locally on mobile devices, without retraining, and with low computational overhead. Summary of the Invention
[0006] To address the shortcomings of existing technologies and maintain the accuracy of continuous identity authentication on the edge device, this invention adopts the following technical solution:
[0007] A behavior-adaptive continuous authentication method includes the following steps:
[0008] Step 1: Component Attribution Analysis; The basic continuous identity authentication model is deconstructed into multiple computational units. Taking the computational unit as a component, the ablation vector is designed to determine whether the parameters of the component are set to zero during the inference process. Ablation experiments are conducted on the components based on the user behavior dataset to obtain a set of components with optimal performance.
[0009] Step 2: Adaptive Parameter Search; Based on the current user's behavior data, apply scaling factors to the parameters of key components to maximize the accuracy of identity authentication. Perform optimization search to obtain the optimal scaling factor, and apply the optimal scaling factor to the key components of the basic continuous identity authentication model to generate an adaptive continuous identity authentication model for the current user's behavior, and perform continuous identity authentication for the current user.
[0010] Furthermore, step one includes the following steps:
[0011] Step 1.1: Component division;
[0012] The pre-trained basic continuous identity authentication model is deconstructed into an identity authentication reasoning graph composed of multiple computational units. The computational units in the graph are divided into components, and the granularity of the components is adjustable.
[0013] Step 1.2: Component ablation experiment;
[0014] Generate a set of random ablation vectors, each vector determining which components are ablated during inference, i.e., component parameters are set to zero; run the authentication model built from these partially ablated components on a historical user behavior dataset and obtain the performance loss.
[0015] Step 1.3: Ranking by importance;
[0016] Linear regression analysis was used to calculate the impact weight of each component on the performance loss of the identity authentication model; the larger the negative weight value, the greater the positive contribution of the component to the model performance; all components were sorted in ascending order according to their weights, and the key components with the greatest impact on the model performance were selected.
[0017] Step 1.4: Distribute model and component information;
[0018] The basic continuous identity authentication model and the information of the key components are distributed and deployed to the edge devices.
[0019] Furthermore, step two includes the following steps:
[0020] Step 2.1: Initialize parameter adjustments;
[0021] On mobile devices, the parameters of the selected key components are adjusted based on the user's latest local behavior data.
[0022] Step 2.2: Bayesian optimization search;
[0023] Using Gaussian process regression as a surrogate model, with expected improvement as the acquisition function and maximizing authentication accuracy as the optimization objective, this algorithm guides the parameter adjustment of key components and can find the global optimum with the fewest evaluations.
[0024] Step 2.3: Iterative optimization;
[0025] Repeat the process of adjusting parameters, evaluating the accuracy of the identity authentication model on local data, updating the proxy model and the collection function until the preset number of iterations or convergence conditions are reached.
[0026] Step 2.4: Generate an adaptive model;
[0027] Obtain a set of optimal scaling factors θ* from the iterative optimization output, apply them to the key components of the basic continuous identity authentication model, and generate an adaptive authentication model optimized for the current user's current behavior pattern.
[0028] Furthermore, in step 2.1, the parameter adjustment strategy is to apply a uniform scaling factor θ (e.g., p) to all parameters within the component. j = p j * (1 + θ j The initial value of the scaling factor θ is sampled from a uniform distribution of [-1, 1].
[0029] A behavior-adaptive continuous identity authentication system includes a component attribution analysis module and a parameter search adaptive module;
[0030] The component attribution analysis module deconstructs the basic continuous identity authentication model into multiple computational units. Taking the computational units as components, it determines whether the parameters of the component are set to zero during the inference process by designing ablation vectors. Based on the user behavior dataset, it conducts ablation experiments on the components to obtain a set of components with optimal performance.
[0031] The parameter search adaptive module applies scaling factors to the parameters of key components based on the current user's behavior data. With the optimization goal of maximizing the accuracy of identity authentication, it performs an optimization search to obtain the optimal scaling factor. The optimal scaling factor is then used in the key components of the basic continuous identity authentication model to generate an adaptive continuous identity authentication model for the current user's behavior, and performs continuous identity authentication for the current user.
[0032] Furthermore, the component attribution analysis module includes a component segmentation module, a component ablation experiment module, a sorting module, and a distribution module.
[0033] The component partitioning module deconstructs the pre-trained basic continuous identity authentication model into an identity authentication reasoning graph composed of multiple computing units, and partitions the computing units in the graph into components, with adjustable component granularity;
[0034] The component ablation experiment module generates a set of random ablation vectors, each vector determining which components are ablated during inference, i.e., the component parameters are set to zero; the identity authentication model built by these partially ablated components is run on a historical user behavior dataset to obtain the performance loss;
[0035] The sorting module calculates the impact weight of each component on the performance loss of the identity authentication model through linear regression analysis; the larger the negative weight value, the greater the positive contribution of the component to the model performance; all components are sorted in ascending order according to the weight, and the key components with the greatest impact on the model performance are selected.
[0036] The distribution module distributes and deploys the basic continuous identity authentication model and the information of the key components to the end-side device.
[0037] Furthermore, the parameter search adaptive module includes an initialization parameter adjustment module, an optimization search module, and an adaptive model generation module;
[0038] The initialization parameter adjustment module adjusts the initialization parameters of the selected key components based on the user's latest local behavior data.
[0039] The optimization search module uses Gaussian process regression as the surrogate model, desired improvement as the acquisition function, and maximizing authentication accuracy as the optimization objective to guide the parameter adjustment of key components. This algorithm can find the global optimum with the fewest evaluations. Iterative optimization involves repeatedly performing parameter adjustment, evaluating the accuracy of the identity authentication model on local data, and updating the surrogate model and acquisition function until the preset number of iterations or convergence conditions are reached.
[0040] The adaptive model generation module obtains a set of optimal scaling factors θ* from the iterative optimization output, applies them to the key components of the basic continuous identity authentication model, and generates an adaptive authentication model optimized for the current user's current behavior pattern.
[0041] An adaptive continuous identity authentication system includes a server. The server deconstructs a basic continuous identity authentication model into multiple computational units, using these units as components. By designing ablation vectors, it determines whether the parameters of a component are set to zero during inference. Based on a user behavior dataset, it conducts ablation experiments on the components to obtain a set of key components with optimal performance. The terminal then acquires these key components and applies scaling factors to their parameters, aiming to maximize authentication accuracy. An optimization search is performed to obtain the optimal scaling factor, which is then applied to the key components of the basic continuous identity authentication model to generate an adaptive continuous identity authentication model for the current user behavior. The terminal then performs continuous identity authentication for the current user.
[0042] An adaptive continuous identity authentication system includes a terminal. The terminal obtains key components from a server. These key components are derived by deconstructing a basic continuous identity authentication model and using the deconstructed computational units as components. An ablation vector is designed to determine whether the parameters of the components are set to zero during inference. Ablation experiments are conducted on the components based on a user behavior dataset to identify the largest set of optimal-performing key components. The terminal applies scaling factors to the parameters of the key components to maximize the identity authentication accuracy as the optimization objective. An optimization search is performed to obtain the optimal scaling factor, which is then applied to the key components of the basic continuous identity authentication model to generate an adaptive continuous identity authentication model for the current user's behavior, enabling continuous identity authentication for the current user.
[0043] The advantages and beneficial effects of this invention are as follows:
[0044] This invention proposes a behavior-adaptive continuous identity authentication method and system, which improves the authentication accuracy of continuous identity authentication for users, and also improves computational efficiency and reduces computational overhead, making it very suitable for mobile devices. The entire adaptive process of this invention can be completed locally on the user's device without uploading any new biological behavior data to the server, thus protecting user privacy. At the same time, it does not depend on a specific model architecture and has been verified to be effective on various mainstream CA models such as CNN, LSTM, and Transformer, demonstrating strong versatility. Attached Figure Description
[0045] Figure 1 This is an architecture diagram of the system in an embodiment of the present invention.
[0046] Figure 2 This is a flowchart of the method in an embodiment of the present invention.
[0047] Figure 3 This is a UMAP dimensionality reduction visualization comparison chart of user behavior features at different time periods in an embodiment of the present invention.
[0048] Figure 4 This is a graph showing the decrease in authentication accuracy of the basic CA model over time for different users in this embodiment of the invention (H-MOG).
[0049] Figure 5 This is a detailed diagram illustrating the accuracy improvement achieved for different users on the H-MOG dataset in this embodiment of the invention.
[0050] Figure 6 This is a detailed diagram illustrating the accuracy improvement achieved for different users on the HuMIdb dataset in this embodiment of the invention.
[0051] Figure 7 This is a visualization comparison chart of positive and negative samples of the pre-adaptive model using UMAP in an embodiment of the present invention.
[0052] Figure 8 This is a visualization comparison chart of positive and negative samples of the adaptive model using UMAP in an embodiment of the present invention. Detailed Implementation
[0053] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0054] like Figure 1 As shown, this invention systematically identifies and quantifies a key but long-neglected challenge for the first time—"temporal inconsistency of identity authentication behavior patterns"—and proposes a behavior-adaptive continuous identity authentication system. This system includes component attribution analysis on the server and parameter search on the mobile device, performing continuous identity authentication based on user behavior and biometrics. It aims to address the problem of decreased accuracy in continuous authentication on mobile devices caused by the drift of user behavior patterns over time. Without model retraining, uploading user privacy data, or consuming large amounts of local storage, it achieves local adaptation of the model to maintain high-precision continuous identity authentication. This invention proposes a new paradigm of "behavior adaptation" by treating the pre-trained continuous identity authentication (CA) model as a "white box" and innovatively combining component attribution analysis and parameter search adaptation as two core modules to achieve automatic, efficient, and localized model adaptation.
[0055] The component attribution analysis module, on the server side, deconstructs the basic continuous identity authentication (CA) model. base Generate a set of random ablation vectors for multiple computational units. An ablation model is constructed, using computational units as components, along with a historical user behavior dataset X. B Input the ablation model, and use the ablation vector Determine which components are "ablated" during inference (parameters set to zero), and adjust the ablation vectors based on the evaluation results of each ablation model. Perform linear regression analysis to obtain the relationship between each component and model M. base The impact weight W of the performance loss is used to select the improvement model M. base The k key components of performance.
[0056] The parameter search adaptive module, on the mobile device, applies scaling factors to the parameters of k selected key components based on the current user's behavior data. With the optimization objective of maximizing authentication accuracy, it performs an optimization search to obtain the optimal scaling factor, which is then used in the basic continuous authentication (CA) model M. base The system uses k key components to generate an adaptive continuous authentication model for the current user's behavior, which is then used for continuous authentication of the current user.
[0057] like Figure 2 As shown, specifically, the method of the present invention includes the following steps:
[0058] Step 1: Component Attribution Analysis (COMAtt), which includes the following steps:
[0059] Step 1.1: Component division;
[0060] The pre-trained basic continuous identity authentication CA model M base It is deconstructed into an authentication inference graph consisting of multiple computational units (called "components" C). The granularity of the components is adjustable; for example, for Transformer models, it can be a Q / K / V projection matrix, a single attention head, or a combination of multiple heads; for CNN models, it can be a single convolutional kernel, convolutional kernels grouped by output channels, or an entire network layer.
[0061] Step 1.2: Component ablation experiment;
[0062] Generate a set of random "ablation vectors", each vector determining which components are "ablated" during inference (parameters are set to zero); run these partially ablated models on a historical dataset and record their performance loss.
[0063] Step 1.3: Ranking by importance;
[0064] By using linear regression analysis, the weight W of each component on the model performance loss is calculated; the larger the negative weight value, the greater the positive contribution of the component to the model performance; all components are sorted in ascending order according to their weights, and the top k key components (Top-k) with the greatest impact on model performance are selected.
[0065] Step 1.4: Distribute model and component information;
[0066] The basic continuous identity authentication CA model M base Along with information about the first k key components mentioned above (such as the component's location identifier in the model), it is deployed to the user's mobile device.
[0067] Step 2: Adaptive Parameter Search (ParaSearch), which includes the following steps:
[0068] Step 2.1: Initialization;
[0069] On mobile devices, based on the user's latest local behavior data X new Positive samples of local behavioral data, including XP new The reference anchor point XA of the positive samplenew (Template), negative samples XN, initialize and adjust the parameters of the selected k key components; the adjustment strategy is to apply a uniform scaling factor θ (e.g., p) to all parameters within the component. j = p j * (1 +θ j The initial value of θ is sampled from a uniform distribution in [-1, 1].
[0070] Step 2.2: Bayesian optimization search;
[0071] Using Gaussian Process Regression as a surrogate model, Expected Improvement (EI) as the acquisition function, and maximizing authentication accuracy as the optimization objective, this algorithm can find the global optimum with the fewest number of evaluations.
[0072] Step 2.3: Iterative optimization;
[0073] Repeat the process of adjusting parameters, evaluating model accuracy on local data, and updating the surrogate model and acquisition function until the preset number of iterations or convergence conditions are reached.
[0074] Step 2.4: Generate an adaptive model;
[0075] The algorithm ultimately outputs a set of optimal scaling factors θ*, which are then applied to the basic continuous authentication (CA) model M. base Using k key components, an adaptive authentication model M is generated that is optimized for the user's current behavior pattern. adapted .
[0076] In this embodiment of the invention, the UMAP (Uniform Manifold Approximation and Projection) dimensionality reduction visualization of user behavior features at different time periods (the first two days and the last two days) is shown, such as... Figure 3 The UMAP mapping between old and new data shown in the figure reveals a drift phenomenon in user behavior patterns. On the H-MOG dataset, the authentication accuracy of the basic continuous identity authentication (CA) model decreases over time for different users, as illustrated in the figure. Figure 4 As shown. To address the issue of decreased accuracy in continuous user identification due to the drift of user behavior features, this invention's method is implemented for different users based on the H-MOG and HuMIdb datasets, achieving an accuracy improvement as shown below. Figure 5 , Figure 6As shown, the authentication accuracy was improved by 3.50% and 2.26% respectively, effectively solving the performance degradation problem caused by behavioral pattern drift and significantly improving authentication accuracy. UMAP was used to visualize and compare positive and negative samples of the models before and after adaptation, as shown below. Figure 7 , Figure 8 The old model feature embedding UMAP heatmap and the new model feature embedding UMAP heatmap shown demonstrate that the inter-class discrimination of the adaptive model is significantly improved. Furthermore, the adaptive process of this invention only takes 12.04 seconds (H-MOG) and 6.08 seconds (HuMIdb) on mobile devices, which is 17.75 times more efficient than traditional fine-tuning methods. It has extremely low computational overhead and is very suitable for edge devices.
[0077] This invention is particularly suitable for edge devices with limited computing resources and storage space, such as smartphones and tablets, and can be applied to sensitive service scenarios that require continuous protection of user identity security, such as mobile banking, healthcare, and social software.
[0078] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A behavior-adaptive continuous identity authentication method, characterized in that... Includes the following steps: Step 1: Component Attribution Analysis; The basic continuous authentication model is deconstructed into multiple computational units. Each computational unit is treated as a component. Ablation vectors are designed to determine whether the component's parameters are set to zero during inference. Ablation experiments are conducted on the components based on a user behavior dataset to obtain a set of key components with optimal performance. This includes the following steps: Step 1.1: Component division; The pre-trained basic continuous identity authentication model is deconstructed into an identity authentication reasoning graph composed of multiple computational units. The computational units in the graph are divided into components, and the granularity of the components is adjustable. Step 1.2: Component ablation experiment; Generate a set of random ablation vectors, each vector determining which components are ablated during inference, i.e., component parameters are set to zero; run the authentication model built from these partially ablated components on a historical user behavior dataset and obtain the performance loss. Step 1.3: Ranking by importance; By using linear regression analysis, the impact weight of each component on the performance loss of the identity authentication model is calculated; the components are then ranked according to their weights, and the key components with the greatest impact on model performance are selected. Step 1.4: Distribute model and component information; The basic continuous identity authentication model and the information of the key components will be distributed and deployed. Step Two: Adaptive Parameter Search; Based on the current user's behavior data, scaling factors are applied to the parameters of key components to maximize authentication accuracy. An optimization search is performed to obtain the optimal scaling factor, which is then used in the key components of the basic continuous authentication model to generate an adaptive continuous authentication model for the current user's behavior, enabling continuous authentication for the current user. This includes the following steps: Step 2.1: Initialize parameter adjustments; Based on the latest local user behavior data, the parameters of the selected key components are adjusted for initialization. Step 2.2: Optimize the search; Gaussian process regression is used as a surrogate model, with expected improvement as the acquisition function and maximizing authentication accuracy as the optimization objective to guide the parameter adjustment of key components; Step 2.3: Iterative optimization; Repeat the process of adjusting parameters, evaluating the accuracy of the identity authentication model, updating the proxy model and the collection function until the preset number of iterations or convergence conditions are reached. Step 2.4: Generate an adaptive model; Obtain a set of optimal scaling factors from the iterative optimization output, apply them to key components of the basic continuous authentication model, and generate an adaptive authentication model optimized for the current user's current behavior pattern.
2. The behavior-adaptive continuous identity authentication method according to claim 1, characterized in that: In step 2.1, the parameter adjustment strategy is to apply a uniform scaling factor to all parameters within the component, with the initial value of the scaling factor sampled from a uniform distribution of [-1, 1].
3. A behavior-adaptive continuous identity authentication system, comprising a component attribution analysis module and a parameter search adaptive module, characterized in that: The component attribution analysis module deconstructs the basic continuous identity authentication model into multiple computational units. Taking the computational units as components, it determines whether the parameters of the components are set to zero during the inference process by designing ablation vectors. Based on the user behavior dataset, it conducts ablation experiments on the components to obtain a set of key components with optimal performance. The component attribution analysis module includes a component partitioning module, a component ablation experiment module, a sorting module, and a distribution module. The component partitioning module deconstructs the pre-trained basic continuous identity authentication model into an identity authentication reasoning graph composed of multiple computing units, and partitions the computing units in the graph into components, with adjustable component granularity; The component ablation experiment module generates a set of random ablation vectors, each vector determining which components are ablated during inference, i.e., the component parameters are set to zero; the identity authentication model built by these partially ablated components is run on a historical user behavior dataset to obtain the performance loss; The sorting module calculates the weight of each component's impact on the performance loss of the identity authentication model through linear regression analysis. The components are sorted according to their weights, and the key components that have the greatest impact on model performance are selected. The distribution module distributes and deploys the basic continuous identity authentication model and the key components. The parameter search adaptive module applies scaling factors to the parameters of key components based on the current user's behavior data. With the optimization goal of maximizing the accuracy of identity authentication, it performs an optimization search to obtain the optimal scaling factor. The optimal scaling factor is then used for the key components of the basic continuous identity authentication model to generate an adaptive continuous identity authentication model for the current user's behavior, and performs continuous identity authentication for the current user. The parameter search adaptive module includes an initialization parameter adjustment module, an optimization search module, and an adaptive model generation module. The initialization parameter adjustment module adjusts the initialization parameters of the selected key components based on the user's latest local behavior data. The optimization search module uses Gaussian process regression as the surrogate model, desired improvement as the acquisition function, and maximizing authentication accuracy as the optimization objective to guide the parameter adjustment of key components; iterative optimization repeatedly executes the process of parameter adjustment, evaluating the accuracy of the identity authentication model, and updating the surrogate model and acquisition function until the preset number of iterations or convergence conditions are reached. The adaptive model generation module obtains a set of optimal scaling factors from the iterative optimization output, applies them to key components of the basic continuous identity authentication model, and generates an adaptive authentication model optimized for the current user's current behavior pattern.
4. A behavior-adaptive continuous identity authentication system according to claim 3, comprising a server, characterized in that: The server deconstructs the basic continuous authentication model into multiple computational units. Using these computational units as components, it designs ablation vectors to determine whether the parameters of a component are set to zero during inference. Based on a user behavior dataset, it conducts ablation experiments on the components to obtain a set of key components with optimal performance. The terminal then acquires these key components and applies scaling factors to their parameters, maximizing authentication accuracy as the optimization objective. An optimization search is performed to obtain the optimal scaling factor, which is then applied to the key components of the basic continuous authentication model to generate an adaptive continuous authentication model for the current user's behavior. This model is then used for continuous authentication of the current user on the terminal.
5. A behavior-adaptive continuous identity authentication system according to claim 3, comprising a terminal, characterized in that: The terminal obtains key components from the server. These key components are deconstructed from the basic continuous authentication model, using the deconstructed computational units as components. By designing ablation vectors, the terminal determines whether the parameters of the components are set to zero during inference. Ablation experiments are conducted on the components based on a user behavior dataset to identify the largest set of key components with optimal performance. The terminal applies scaling factors to the parameters of the key components to maximize authentication accuracy as the optimization objective. An optimization search is performed to obtain the optimal scaling factor, which is then applied to the key components of the basic continuous authentication model to generate an adaptive continuous authentication model for the current user's behavior, enabling continuous authentication for the current user.