Smart campus access control identity accurate verification method and system based on multi-mode biological recognition

Abstract

The invention discloses a smart campus access control identity accurate verification method and system based on multi-mode biological recognition. The method comprises the following steps: collecting dynamic biological characteristic flow and campus space-time trajectory data through a distributed terminal; performing three-dimensional space-time alignment preprocessing to generate a space-time associated biological characteristic matrix; constructing a campus twinborn multi-modal model output structured characteristic spectrum of knowledge distillation; introducing a campus identity confidence conduction mechanism to calculate a dynamic confidence weight; the feature chain of block chain evidence storage is combined with a homomorphic encryption algorithm to complete localized comparison, and a verification result containing a confidence interval is generated; according to the result, differential response is executed, multi-level verification is triggered, and potential risk advanced early warning is achieved; the system comprises a multi-modal data acquisition module, a feature preprocessing module, a twinborn model construction module, a confidence coefficient conduction module, and a block chain comparison and response and early warning module. Through multi-modal fusion and campus scene adaptation, efficiency and privacy protection are balanced, campus security closed-loop management is constructed, and the method is suitable for various access control scenes of smart campuses.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent campus safety management, and particularly relates to an intelligent campus access control identity accurate verification method and system based on multi-modal biometric recognition. BACKGROUND

[0002] With the deepening of the construction of intelligent campus, the campus access control as the core link of safety management is faced with the problems of large personnel mobility, complex identity verification scenes (such as peak class time, temporary visitor entry into the campus, etc.), and traditional verification methods being easily used by impostors. The existing single modal biometric recognition technology (such as fingerprint, face recognition) is difficult to meet the needs of high accuracy and high robustness of identity verification due to environmental interference (such as changes in light, fingerprint wear) and scene limitations (such as mask obstruction). At the same time, the campus population includes students, faculty, visitors and other different types, and their behavior patterns (such as entry frequency, activity area) differ significantly, and the traditional access control system lacks a targeted dynamic verification mechanism. In addition, the campus safety management needs to balance verification efficiency and privacy protection, and the existing centralized data processing mode has the risk of data leakage. Therefore, developing an accurate verification technology that integrates multi-modal biometric features and campus scene characteristics has become the key to solving the pain points of intelligent campus access control.

[0003] Traditional campus access control technology mainly falls into three categories: one is a non-biometric recognition scheme based on IC cards and passwords, which has the advantages of low cost and easy deployment, but has security risks such as card loss and password leakage, and cannot cope with the risk of impostors; the second is a single modal biometric recognition scheme, such as face recognition, which captures facial features through a camera for comparison, and has the advantages of non-contact and high convenience, but the accuracy drops sharply in scenes such as backlight and obstruction; fingerprint recognition is easily affected by finger moisture and wear, and has insufficient stability; the third is a simple multi-modal fusion scheme, such as "fingerprint + password" combination, which improves security, but does not solve the problems of poor complementarity between modalities and fixed weights, and still has the risk of misjudgment in complex scenes. Overall, traditional technology is difficult to balance accuracy, robustness and scene adaptability, and lacks the ability to manage different types of campus populations.

[0004] The existing patent (CN119251945A) discloses a multi-modal biometric access control system, which collects fingerprints, faces, and iris images through a biometric module, filters effective images through k-means clustering, extracts features using LeNet network, and completes identity verification through weighted fusion of multi-modal features. Its advantages are: introducing multi-modal fusion improves recognition stability, and filtering effective images through clustering algorithm reduces noise interference. However, there are significant defects: first, the feature fusion uses fixed weights, without considering the reliability differences of different modalities in different scenarios, such as the credibility of facial features in backlight scenarios should be reduced; second, it does not combine the characteristics of the campus scene, and cannot distinguish the behavior patterns of students, faculty and other crowds; third, it uses centralized data processing, which has the risk of privacy data leakage, and does not involve an abnormal behavior warning mechanism, making it difficult to meet the dynamic needs of campus safety management.

[0005] The existing patent (CN118968665A) discloses an intelligent access control management method based on multi-modal recognition and Internet of Things technology, which collects visual, audio, and thermal data through distributed devices, aligns them through dynamic time warping, uses tensor decomposition and graph convolution network for deep fusion, and realizes identity verification through behavior pattern analysis and environmental context verification. Its advantages are: introducing Internet of Things sensing network improves scene adaptability, high-order feature fusion enhances recognition robustness, and supports dynamic access control. However, there are still deficiencies: first, it is not optimized for campus scenarios, such as not using course schedules, classroom check-ins, and other campus-specific data to assist verification; second, multi-modal fusion does not introduce a crowd differentiation model, and a unified verification strategy is used for students, visitors, and other groups, making it difficult to balance efficiency and accuracy; third, abnormal detection relies on general algorithms such as isolation forest, without considering campus behavior rules (such as permission verification for entering laboratories at night), making the warning less targeted.

[0006] Therefore, to solve the above technical problems, the present application proposes a smart campus access control identity verification method and system based on multi-modal biometric recognition. SUMMARY

[0007] Based on the above technical problems, the present application discloses a smart campus access control identity verification method and system based on multi-modal biometric recognition, wherein the smart campus access control identity verification method based on multi-modal biometric recognition comprises:

[0008] S1, collecting dynamic biometric feature streams and campus space-time trajectory data of the person to be verified through a distributed multi-modal collection terminal linked by a campus Internet of Things sensing network;

[0009] S2, performing three-dimensional space-time alignment preprocessing on the dynamic biometric feature stream, extracting key frame sequences, and then generating cross-modal complementary features through an improved generative adversarial network to obtain a space-time correlated biometric feature matrix;

[0010] S3, construct a campus twin multi-modal model of knowledge distillation, anchor the campus scene exclusive feature points of each modality through an attention mechanism, and output a structured feature atlas;

[0011] S4, introduce a campus identity confidence propagation mechanism to convert campus spatiotemporal trajectory data into a behavior feature vector, construct an association atlas of biological features and behavior features through an improved graph neural network, calculate dynamic confidence weights of each modality feature, and realize deep coupling of the feature layer;

[0012] S5, compare the campus identity feature chain stored by the blockchain, call the feature template fragments of the corresponding security domain through a smart contract, combine a homomorphic encryption similarity calculation algorithm optimized based on the campus scene, complete the local comparison at the edge node, and generate a verification result containing a confidence interval;

[0013] S6, execute a differentiated access control response according to the verification result, trigger a multi-level verification channel for abnormal results, write the feature evolution trajectory of the verification process into a campus feature evolution database, and realize early warning of potential risks through a time series anomaly detection algorithm.

[0014] Preferably, the dynamic biological feature stream and the campus spatiotemporal trajectory data of the person to be verified are collected, specifically: the dynamic biological feature stream includes a facial micro-expression sequence, a live fingerprint pressure change curve, and an iris dynamic texture video stream; and the campus spatiotemporal trajectory data includes a movement heat map in an electronic fence in a campus for a period of time, a classroom check-in associated position, and a teaching activity matching degree.

[0015] Preferably, the dynamic biological feature stream is subjected to three-dimensional spatiotemporal alignment preprocessing, specifically: the time dimension frame sequence of the dynamic biological feature stream is extracted , the light stability coefficients of each frame are calculated through a campus light adaptive dynamic frame screening algorithm , wherein , are scene adaptation coefficients, represents the average light intensity of the th frame, represents the contrast entropy value of the th frame, and the key frame sequence is screened out , wherein is a campus scene light threshold, a cross-modal feature association matrix is constructed, and complementary features are generated through an improved generative adversarial network, wherein the loss function of the generator is , is a discriminator, is a single modality feature, is a multi-modal feature, , is a modal balance coefficient, represents a feature similarity, and the key frame sequence is subjected to spatial coordinate normalization processing to obtain a biological feature matrix associated with space and time , wherein the matrix element is a space-time calibration coefficient, is a time weight factor of the th frame.

[0016] Preferably, when constructing the campus twin multi-modal model based on knowledge distillation, specifically: by using a federal learning framework, the teacher network T aggregates historical feature data of each campus, wherein the loss function of the teacher network is: , wherein , are respective weight coefficients, is a classification loss function, is a feature distribution difference loss function of each campus, and a dynamic forgetting factor model is constructed for full-time students, faculty and visitors respectively, and the forgetting factor is: , wherein is an initial forgetting factor, is a forgetting decay coefficient, is the time difference from the last feature update, and when anchoring the campus scene exclusive feature points of each modality through the attention mechanism, the attention weight of the th modality is: , and are learnable attention parameters, is a feature mapping of the th modality, and after the features of each modality are weighted by attention, they are input into the feature fusion layer to obtain a structured feature map : , represents a feature splicing operation, is the total number of modalities.

[0017] Preferably, when introducing a campus identity confidence transmission mechanism, the campus space-time trajectory data is converted into a behavior feature vector , wherein is the space-time trajectory feature of the th time window, and when constructing the association graph of biological features and behavior features through an improved graph neural network, the element of the adjacency matrix is , wherein is the th biological feature vector, For the first A behavioral feature vector, For activation function, These are learnable parameters.

[0018] Preferably, the dynamic confidence weights of each modal feature are calculated based on the PageRank values ​​of the graph nodes, using an iterative formula. Calculate the node importance score, where For PageRank vectors, This is the adjacency matrix of the association graph. The damping coefficient is... The initial importance distribution is given; the dynamic confidence weights for each modality feature are... ,in The weight matrix is ​​a learnable matrix, forming a feature fusion formula. , For the first Dynamic confidence weights for each modal feature This represents the total number of modes fused.

[0019] Preferably, when comparing campus identity feature chains stored on blockchain, the feature template fragments of the corresponding security domain are invoked through smart contracts. ,in This represents the security domain index, which, combined with a homomorphic encryption similarity calculation algorithm optimized for campus scenarios, calculates the features to be verified at the edge nodes. With template segmentation The similarity score is calculated using the following formula: ,in It is a homomorphic encryption function. For the encryption domain similarity calculation, generate a verification result containing a confidence interval, with the upper and lower limits of the confidence interval being... ,in , For historical comparison scoring, The mean, The dynamic calibration factor output by the campus safety incident tree model. This represents the number of historical comparison score samples used when calculating the confidence interval.

[0020] Preferably, the multi-level verification channel includes remote verification by counselors and cross-verification of the timetable; when the lower limit of the confidence interval of the verification result is lower than the preset basic threshold but higher than the suspicious threshold, a second-level verification based on the campus social relationship graph is automatically triggered, and a verification anomaly reminder containing a spatiotemporal trajectory heatmap is pushed through the remote verification by counselors, and the matching degree of class time is verified; when the lower limit of the confidence interval is lower than the suspicious threshold, a third-level verification is initiated, triggering the teaching management system to retrieve the class attendance records of the preset time period and perform continuity verification with the current location, and push a risk warning containing historical behavior profiles to the security department.

[0021] Preferably, when using a temporal anomaly detection algorithm to achieve early warning of potential risks, the specific steps are: constructing a temporal state space of feature evolution trajectories. ,in Indicates the first Each feature dimension at time... Given the state value, calculate the Mahalanobis distance between the current state and the historical normal pattern, using the following formula: ,in and These are the mean vector and covariance matrix of the historical characteristic distribution, respectively. When continuous Each time step satisfies When an early warning is triggered, among which The standard deviation of the Mahalanobis distance. This is the early warning threshold coefficient that is dynamically adjusted based on the frequency of campus safety incidents.

[0022] The smart campus access control system based on multimodal biometrics includes a multimodal data acquisition module, a feature preprocessing module, a twin model construction module, a confidence transmission module, a blockchain comparison module, and a response and early warning module.

[0023] The multimodal data acquisition module collects dynamic biometric data and campus spatiotemporal trajectory data through distributed terminals and links with the campus Internet of Things sensing network.

[0024] The input of the feature preprocessing module is connected to the multimodal data acquisition module, which is used to perform three-dimensional spatiotemporal alignment preprocessing on the dynamic biofeature stream and generate a spatiotemporally correlated biofeature matrix.

[0025] The input of the twin model construction module is connected to the feature preprocessing module, which is used to construct a knowledge distillation campus twin multimodal model and output a structured feature map.

[0026] The confidence transfer module is connected to the multimodal data acquisition module and the twin model construction module respectively, and is used to convert campus spatiotemporal trajectory data into behavioral feature vectors, construct a correlation map between biological features and behavioral features, and calculate dynamic confidence weights.

[0027] The input of the blockchain comparison module is connected to the confidence transmission module, which is used to call the security domain feature template fragment and complete the localized comparison through the homomorphic encryption similarity algorithm to generate a verification result containing a confidence interval;

[0028] The input of the response and early warning module is connected to the blockchain comparison module, which is used to execute differentiated access control responses and trigger multi-level verification channels based on the verification results, and to achieve early warning of potential risks through time-series anomaly detection.

[0029] Compared with the prior art, the technical solution of this application has the following technical effects:

[0030] This invention significantly improves the recognition accuracy of multimodal features by fusing dynamic biometric data streams with campus spatiotemporal trajectory data and combining this with an improved generative adversarial network to eliminate complex lighting interference. Dynamic forgetting factor sub-models are constructed for different groups such as students and faculty, and attention mechanisms are used to anchor campus-specific feature points, making the verification results more closely match the behavioral characteristics of the campus population. This effectively solves the problem of misjudgment caused by single modality under occlusion and environmental interference, raising the accuracy of identity verification to a new level.

[0031] This invention introduces a campus identity confidence transmission mechanism. It constructs a correlation graph between biometric and behavioral features using an improved graph neural network, and dynamically adjusts the weights of each modality based on PageRank values, achieving deep fusion of the feature layers. This dynamic coupling mechanism can adaptively optimize feature weights according to real-time scenarios (such as class times or sensitive areas), preserving the complementarity of multimodal data while avoiding interference from invalid features, significantly improving verification reliability in complex scenarios.

[0032] This invention employs a blockchain-based campus identity feature chain for evidence storage, combined with a homomorphic encryption algorithm optimized for campus scenarios. Localized comparison is performed at edge nodes, preventing the risk of centralized leakage of feature data while reducing data transmission latency. The smart contract's design of using security domain template sharding ensures access control isolation across different campuses and security levels, guaranteeing privacy and meeting the real-time response requirements of campus access control.

[0033] This invention performs secondary verification of abnormal results through multi-level verification channels (such as remote verification by counselors and cross-verification of course schedules), and combines time-series anomaly detection algorithms to analyze the feature evolution trajectory, realizing closed-loop management from identity verification to risk warning. The system can proactively identify potential abnormal behaviors (such as unauthorized entry into the laboratory) and trigger the early warning mechanism in advance, providing proactive defense capabilities for campus safety management and effectively reducing the probability of safety incidents.

[0034] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0035] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0037] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0038] Figure 1 A flowchart illustrating a method for accurate identity verification in smart campus access control based on multimodal biometrics;

[0039] Figure 2 A schematic diagram of the campus twin multimodal model structure for knowledge distillation;

[0040] Figure 3 This is an architecture diagram of a smart campus access control system for accurate identity verification based on multimodal biometrics.

[0041] Figure 4 A comparison chart of identity verification accuracy rates in different scenarios;

[0042] Figure 5 A comparison chart of verification efficiency and response time;

[0043] Figure 6 A statistical chart showing the number of unauthorized cross-campus access denials and the pass rate of valid permission verification requests;

[0044] Figure 7 A comparison chart of anomaly handling efficiency for multi-level verification channels;

[0045] Figure 8 This chart compares the number of early warnings for time-series anomalies with the number of campus safety incidents.

[0046] Figure 9A comparison chart of feature matching degree and decay rate (3 months). Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0048] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0049] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0050] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0051] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0052] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0053] Example 1

[0054] This embodiment mainly describes a method for accurate identity verification in smart campus access control based on multimodal biometrics, such as... Figure 1 As shown, specifically:

[0055] S1. Collect dynamic biometric data and campus spatiotemporal trajectory data of the personnel to be verified through a distributed multimodal acquisition terminal linked by the campus Internet of Things sensing network.

[0056] S2. Perform three-dimensional spatiotemporal alignment preprocessing on the dynamic biometric stream, extract key frame sequences, and then generate cross-modal complementary features through an improved adversarial generative network to obtain a spatiotemporally correlated biometric matrix.

[0057] S3. Construct a knowledge distillation campus twin multimodal model, anchor the campus scene-specific feature points of each modality through the attention mechanism, and output a structured feature map;

[0058] S4. Introduce a campus identity confidence transmission mechanism to transform campus spatiotemporal trajectory data into behavioral feature vectors. Construct a correlation map between biometrics and behavioral features through an improved graph neural network, calculate the dynamic confidence weights of each modality feature, and achieve deep coupling of the feature layer.

[0059] S5. The campus identity feature chain stored on the blockchain is used for comparison. The feature template fragment of the corresponding security domain is called through the smart contract. Combined with the homomorphic encryption similarity calculation algorithm optimized for the campus scenario, the localized comparison is completed at the edge node, and the verification result containing the confidence interval is generated.

[0060] S6. Execute differentiated access control responses based on verification results, trigger multi-level verification channels for abnormal results, write the characteristic evolution trajectory of the verification process into the campus characteristic evolution database, and achieve early warning of potential risks through time-series anomaly detection algorithms.

[0061] Furthermore, the collection of dynamic biometric data and campus spatiotemporal trajectory data of the personnel to be verified specifically includes: the dynamic biometric data includes facial micro-expression sequences, live fingerprint pressure change curves, and iris dynamic texture video streams; the campus spatiotemporal trajectory data includes a daily mobile heat map within the campus electronic fence over a recent period, classroom attendance-related locations, and matching degree of teaching activities.

[0062] Furthermore, the three-dimensional spatiotemporal alignment preprocessing of the dynamic biometric stream specifically involves: extracting the temporal dimension frame sequence of the dynamic biometric stream. The illumination stability coefficient of each frame is calculated using a dynamic frame selection algorithm that adapts to campus illumination. ,in , These are the scene adaptation coefficients, Indicates the first The average illumination intensity of the frame, Indicates the first Frame contrast entropy value, filtered out keyframe sequence ,in To determine the illumination threshold for campus scenes, a cross-modal feature correlation matrix is ​​constructed. Complementary features are generated through an improved adversarial generative network, where the generator... The loss function is , For discriminator, It is a single modal feature. For multimodal features, , These are the modal equilibrium coefficients. Representing feature similarity for keyframe sequences Spatial coordinate normalization is performed to obtain a spatiotemporally correlated biofeature matrix. , where matrix elements For spatiotemporal calibration coefficients, For the first The time weighting factor of a frame.

[0063] Furthermore, such as Figure 2 As shown, the construction of the campus twin multimodal model based on knowledge distillation specifically involves: aggregating historical feature data from each campus using a federated learning framework through a teacher network T, wherein the loss function of the teacher network is: ,That , These are the corresponding weighting coefficients. For classification loss function, To determine the loss function based on the differences in characteristic distributions across campuses, dynamic forgetting factor models are constructed for three groups: full-time students, faculty, and visitors. The forgetting factor is: ,in As the initial forgetting factor, This is the forgetting decay coefficient. Given the time difference since the most recent feature update, when anchoring campus scene-specific feature points for each modality using an attention mechanism, the calculation is performed for the... The attention weights for each modality are: , and For learnable attention parameters, For the first The feature maps of each modality are processed, and the attention-weighted features of each modality are then input into the feature fusion layer to obtain a structured feature map. for: , This indicates a feature concatenation operation. This represents the total number of modes.

[0064] Furthermore, when introducing a campus identity confidence transmission mechanism, campus spatiotemporal trajectory data is transformed into behavioral feature vectors through an LSTM network. ,in For the first When constructing a correlation graph between biological and behavioral features using an improved graph neural network based on the spatiotemporal trajectory features of each time window, an adjacency matrix is ​​defined. The elements are ,in For the first A biological feature vector, For the first A behavioral feature vector, For activation function, These are learnable parameters.

[0065] Furthermore, dynamic confidence weights for each modal feature are calculated based on the PageRank values ​​of the graph nodes, using an iterative formula. Calculate the node importance score, where For PageRank vectors, This is the adjacency matrix of the association graph. The damping coefficient is... The initial importance distribution is given; the dynamic confidence weights for each modality feature are... ,in The weight matrix is ​​a learnable matrix, forming a feature fusion formula. , For the first Dynamic confidence weights for each modal feature This represents the total number of modes fused.

[0066] Furthermore, when comparing campus identity feature chains stored on blockchain, a smart contract is used to call the feature template fragments of the corresponding security domain. ,in This represents the security domain index, which, combined with a homomorphic encryption similarity calculation algorithm optimized for campus scenarios, calculates the features to be verified at the edge nodes. With template segmentation The similarity score is calculated using the following formula: ,in It is a homomorphic encryption function. For the encryption domain similarity calculation, generate a verification result containing a confidence interval, with the upper and lower limits of the confidence interval being... ,in , For historical comparison scoring, The mean, The dynamic calibration factor output by the campus safety incident tree model. This represents the number of historical comparison score samples used when calculating the confidence interval.

[0067] Furthermore, the multi-level verification channel includes remote verification by counselors and cross-verification of the course schedule. When the lower limit of the confidence interval of the verification result is lower than the preset basic threshold but higher than the suspicious threshold, a secondary verification based on the campus social relationship graph is automatically triggered. The counselor remote verification pushes a verification anomaly reminder containing a spatiotemporal trajectory heatmap and performs class time matching verification. When the lower limit of the confidence interval is lower than the suspicious threshold, a tertiary verification is initiated. The teaching management system retrieves the class attendance records for the preset time period and performs continuity verification with the current location, and pushes a risk warning containing historical behavior profiles to the security department.

[0068] Furthermore, when using temporal anomaly detection algorithms to achieve early warning of potential risks, the specific steps are as follows: constructing a temporal state space of feature evolution trajectories. ,in Indicates the first Each feature dimension at time... Given the state value, calculate the Mahalanobis distance between the current state and the historical normal pattern, using the following formula: ,in and These are the mean vector and covariance matrix of the historical characteristic distribution, respectively. When continuous Each time step satisfies When an early warning is triggered, among which The standard deviation of the Mahalanobis distance. This is the early warning threshold coefficient that is dynamically adjusted based on the frequency of campus safety incidents.

[0069] This implementation details how the proposed method significantly improves the accuracy and robustness of access control verification through multimodal fusion and adaptation to campus scenarios. The deep coupling of dynamic biometrics and behavioral data solves the misjudgment problem of traditional single-modal recognition in complex environments. The combination of blockchain and edge computing enables real-time verification while ensuring privacy. Multi-level verification channels and anomaly warning mechanisms construct a closed-loop management system for campus security, providing an efficient and secure identity verification solution for smart campuses.

[0070] Example 2

[0071] This embodiment, based on Embodiment 1, describes in detail a smart campus access control system for accurate identity verification based on multimodal biometrics, such as... Figure 3 As shown, it specifically includes a multimodal data acquisition module, a feature preprocessing module, a twin model construction module, a confidence transmission module, a blockchain comparison module, and a response and early warning module;

[0072] The multimodal data acquisition module's acquisition terminals are deployed at various campus access points, teaching buildings, libraries, and other locations. These terminals include high-definition cameras, fingerprint scanners, and iris scanners, all equipped with waterproof, dustproof, and strong light-resistant features to adapt to the complex campus environment. Data acquisition involves real-time acquisition of dynamic biometric data streams of individuals to be verified, such as facial micro-expression sequences, live fingerprint pressure change curves, and iris dynamic texture video streams. Simultaneously, it collects spatiotemporal trajectory data within the campus's electronic fence over the past 30 days, including mobile heat maps within the electronic fence, classroom attendance records, and matching information with teaching activities. IoT integration involves deep integration with the campus IoT sensing network to obtain real-time location information and movement trajectories of individuals to be verified, providing a more comprehensive reference for subsequent identity verification.

[0073] The feature preprocessing module includes dynamic frame filtering: a campus lighting-adaptive dynamic frame filtering algorithm is used to extract key frame sequences from the collected dynamic biometric stream, effectively reducing feature interference caused by factors such as lighting changes; cross-modal feature generation: cross-modal complementary features are generated through an improved generative adversarial network to enhance the complementarity between different biometric features and further improve the accuracy and reliability of features; biometric matrix construction: the processed feature data are integrated to construct a spatiotemporally correlated biometric matrix, providing basic data for subsequent model construction and comparison.

[0074] The twin model construction module includes: a teacher network: using a federated learning framework to aggregate historical feature data from each campus to build a global feature model, fully utilizing the data resources of each campus to improve the model's generalization ability; a student network: for specific campus groups, such as full-time students, retired faculty and staff, and temporary construction workers, a dynamic forgetting factor sub-model is constructed, anchoring campus-specific feature points for each modality through an attention mechanism, making the model more suitable for the characteristics of different campus groups; and a structured feature map output: integrating the feature information of the teacher and student networks to output a structured feature map, providing rich feature representations for subsequent confidence propagation and comparison.

[0075] The confidence transmission module transforms behavioral feature vectors: it converts campus spatiotemporal trajectory data into behavioral feature vectors, such as the frequency of entering and exiting the campus, frequently visited areas, and corresponding time periods, providing supplementary information for identity verification from a behavioral perspective; it constructs a correlation graph: it builds a correlation graph between biometrics and behavioral features through an improved graph neural network, and mines the potential relationships between biometrics and behavioral features; it calculates dynamic confidence weights: it calculates the dynamic confidence weights of each modality feature based on the PageRank value of the graph nodes, realizing deep coupling of the feature layer, so that features of different modalities play a more reasonable role in the verification process.

[0076] Feature template sharding invocation of the blockchain comparison module: Feature template shards of the corresponding security domain are invoked through smart contracts to ensure the security and privacy of feature data; Homomorphic encryption similarity calculation: Homomorphic encryption similarity calculation algorithm optimized for campus scenarios is combined to complete localized comparison at edge nodes, reducing security risks during data transmission and improving comparison efficiency; Verification result generation: Verification results containing confidence intervals are generated, where the confidence interval threshold is dynamically calibrated by the campus security event tree model, making the verification results more reliable and targeted.

[0077] Differentiated access control response in the response and early warning module: Executes different access control operations based on the verification results, such as allowing access, prohibiting access, and prompting for further verification, to meet access control management needs under different verification results; Multi-level verification channel triggering: Triggers multi-level verification channels for abnormal results, such as remote verification by counselors and cross-verification of course schedules, ensuring the accuracy of identity verification through multiple verification methods; Advanced early warning of potential risks: Feeds back the verification data and results to the campus security management system, and uses a time-series anomaly detection algorithm to provide advanced early warning of potential risks, identifying potential security hazards in advance.

[0078] Data storage in the university data center: It stores basic data of faculty and staff, basic data of students, campus card number data, university organizational structure and other data to provide data support for the entire system; Data synchronization: It synchronizes data with the multimodal data acquisition module, feature preprocessing module and other modules to ensure data consistency and real-time performance, and to ensure smooth data flow between various modules of the system.

[0079] API interfaces for third-party business systems: Provides API business function interfaces to interact with third-party business systems in universities, such as smart security systems, smart attendance systems, academic affairs systems, and student affairs systems, to achieve data sharing and collaborative work, thereby improving the overall efficiency of campus management; Device integration: Integrates with various facial recognition devices and platforms to achieve data sharing and collaborative work, further improving the functionality of the campus access control and identity verification system.

[0080] This embodiment details how the collaborative work of modules such as multimodal data acquisition, feature preprocessing, twin model construction, confidence propagation, blockchain comparison, and response and early warning enables accurate verification of identity for smart campus access control. It effectively solves the difficult problems existing in the prior art and has high accuracy, security, and reliability.

[0081] Based on Embodiment 1 or 2, this embodiment describes in detail the implementation effect of accurate identity verification for smart campus access control based on multimodal biometrics, specifically as follows:

[0082] like Figure 4As shown, the verification accuracy rate under different scenarios is improved by dynamic fusion of multimodal features to enhance the verification accuracy of complex scenarios. Figure 4 The system's verification performance in various campus scenarios is presented, covering eight typical scenarios including morning rush hour at the entrance of teaching buildings, dormitory areas on rainy days, libraries with masks covering the faces, and laboratories at night. At the entrance of teaching buildings during the morning rush hour (7:30-8:30), the daily verification volume reaches 12,000 times. The system uses three-dimensional spatiotemporal alignment preprocessing (claim 3) to select lighting-stable frames and combines the dynamic complementarity of facial micro-expressions and fingerprint pressure curves, achieving a stable accuracy of 99.4%. The misjudgment rate of facial features due to changes in lighting is only 0.3%, which is 94.8% lower than that of traditional single face recognition (misjudgment rate 5.8%).

[0083] In a rainy dormitory environment, fingerprint sensors are susceptible to moisture damage, causing the pass rate of traditional fingerprint recognition to drop to 78.2%. This system, by using dynamic iris texture video streams (increasing the proportion to 45%) and trajectory data for verification, maintained an accuracy rate of 98.1%. Within 30 days, only 12 instances of secondary verification due to blurred fingerprints occurred, all of which were quickly corrected using iris features. At the library entrance, where masks are present (approximately 3200 people wear masks daily), the system automatically increased the weighting of iris and fingerprint data, achieving a verification accuracy rate of 97.6%, only 1.8 percentage points lower than in unmasked scenarios. In contrast, traditional facial recognition only achieved 62.3% accuracy in this scenario, requiring manual intervention in 31.5% of cases.

[0084] In high-security areas such as laboratories at night, the system adopts stricter standards for verification between 11:00 PM and 6:00 AM. By comparing iris texture with activity trajectories over the past 7 days, it intercepted 17 unauthorized attempts within 30 days, including 12 cases of students entering illegally and 5 cases of unauthorized personnel from outside the school impersonating others. The interception success rate was 100%, which is 23.7 percentage points higher than the traditional access control system (interception success rate of 76.3%).

[0085] Regarding the optimization of efficiency and security of dynamic confidence propagation and blockchain evidence storage, such as Figure 5 As shown, in terms of verification efficiency, localized comparison at edge nodes controls the verification time to 0.6-0.8 seconds per transaction. During peak hours (3 peak periods per day, totaling 4 hours), an average of 1.8 transactions per second is processed, a 157% improvement compared to centralized cloud processing (0.7 transactions per second). No congestion due to system delays occurred within 30 days. The average response time for feature template sharding in blockchain-based evidence storage is 0.2 seconds, a 77.8% reduction compared to traditional database queries (0.9 seconds). The template call success rate in the teaching area security domain (s=1) is 100%, while the response time in the laboratory security domain (s=3), due to its higher encryption level, is 0.3 seconds, still meeting real-time requirements.

[0086] In terms of data security, homomorphic encryption algorithms keep signature data encrypted throughout transmission. Simulated attack tests over 30 days showed that signature-based encryption is 8 times more difficult to crack than traditional AES encryption, and no data breaches occurred. Figure 6 As shown in (a), the smart contract's access control precision for feature templates reaches 100%. Each campus can only call templates within its own region; cross-campus calls require triple permission verification. Within 30 days, 126 unauthorized access requests were rejected. Regarding the statistics of unauthorized access rejections across campuses, requests from campuses A to B, B to C, and C to A were rejected 38, 42, and 46 times respectively, with a total of 126 rejections across regions. This clearly reflects the smart contract's strict permission control capabilities. Figure 6 As shown in (b), the pass rate of permission verification for legitimate requests is clearly seen. The pass rates for requests from campus A to B, B to C, C to A, and the total cross-regional requests are 99.2%, 98.8%, and 99.5%, respectively, indicating that while strictly blocking unauthorized behavior, the efficiency of legitimate requests is not significantly affected.

[0087] The adjustment of dynamic confidence weights makes resource allocation more reasonable. For example, during class time (9:00-12:00), facial features account for 62% of the weight (efficient and fast), while during non-class time, iris features account for 58% of the weight (higher security). The resource waste rate (the proportion of invalid feature calculations) within 30 days is reduced from 18.3% in the traditional system to 3.2%, and the peak server load is reduced by 42%.

[0088] For closed-loop early warning and multi-level verification to strengthen campus safety management, such as Figure 7 The data shown illustrates the security management data from three months of system operation. The multi-level verification channels significantly improved anomaly handling efficiency. Level 2 verification (remote verification by counselors) had an average processing time of 48 seconds, handling 156 minor anomalies within three months. These included 89 cases of decreased facial matching due to hairstyle changes and 47 cases of recognition deviations due to fingerprint wear. All were verified through cross-verification of class schedules, representing a 79.7% reduction compared to traditional manual verification (average 3.2 minutes). Level 3 verification (security department intervention) handled 32 high-risk incidents, averaging 5.2 minutes, a 59.4% reduction compared to the traditional process (12.8 minutes). Eight of these cases were verified as identity theft, achieving a 100% interception rate.

[0089] like Figure 8As shown, the system's early warning effect in detecting temporal anomalies is outstanding. The system's accuracy in identifying anomalies in feature evolution trajectories reached 92.3%, and it issued early warnings for 43 potential risks within three months: 17 cases of students repeatedly loitering near laboratories during unauthorized hours (average warning 1.8 hours in advance), 12 cases of visitors deviating from reported routes (average warning 23 minutes in advance), and 14 cases of faculty and staff showing a continuous decline in feature matching (warning updated 3 days in advance). After implementation, the average monthly number of campus security incidents decreased from 9 before system deployment to 3, a reduction of 66.7%. Among these, laboratory equipment theft incidents reached zero, a 100% decrease compared to before deployment (average 2 incidents per month).

[0090] like Figure 9 As shown, the dynamic update mechanism of the feature evolution database continuously improves the effectiveness of templates. Within three months, feature update reminders were automatically triggered 213 times (187 for students and 26 for faculty). After the updates, the average matching accuracy rose from 93.2% to 99.5%, and the feature decay rate was controlled below 0.2% per month, a reduction of 88.9% compared to the static template library (1.8% decay per month). For example, when student Zhang's facial features changed due to sun exposure during the summer, the system automatically reminded him to update in early September. After the update, the verification accuracy remained at 99.7% for 30 consecutive days.

[0091] This embodiment describes the access control system in detail. By dynamically integrating biometric features and campus behavior data, it significantly improves the accuracy of identity verification in complex scenarios, effectively addresses issues such as light interference and occlusion, and optimizes efficiency and security through blockchain evidence storage and edge computing. Multi-level verification and time-series early warning form a security closed loop, which not only ensures campus safety but also adapts to the behavioral patterns of teachers and students, achieving precise, efficient, and intelligent access control management.

[0092] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for accurate identity verification in smart campus access control based on multimodal biometrics, characterized in that, include: S1. Collect dynamic biometric data and campus spatiotemporal trajectory data of the personnel to be verified through a distributed multimodal acquisition terminal linked by the campus Internet of Things sensing network. S2. Perform three-dimensional spatiotemporal alignment preprocessing on the dynamic biometric stream, extract key frame sequences, and then generate cross-modal complementary features through an improved adversarial generative network to obtain a spatiotemporally correlated biometric matrix. S3. Construct a knowledge distillation campus twin multimodal model, anchor the campus scene-specific feature points of each modality through the attention mechanism, and output a structured feature map; S4. Introduce a campus identity confidence transmission mechanism to transform campus spatiotemporal trajectory data into behavioral feature vectors. Construct a correlation map between biometrics and behavioral features through an improved graph neural network, calculate the dynamic confidence weights of each modality feature, and achieve deep coupling of the feature layer. S5. The campus identity feature chain stored on the blockchain is used for comparison. The feature template fragment of the corresponding security domain is called through the smart contract. Combined with the homomorphic encryption similarity calculation algorithm optimized for the campus scenario, the localized comparison is completed at the edge node, and the verification result containing the confidence interval is generated. S6. Execute differentiated access control responses based on verification results, trigger multi-level verification channels for abnormal results, write the characteristic evolution trajectory of the verification process into the campus characteristic evolution database, and achieve early warning of potential risks through time-series anomaly detection algorithms.

2. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1, characterized in that, The collection of dynamic biometric data and campus spatiotemporal trajectory data of the personnel to be verified specifically includes: the dynamic biometric data includes facial micro-expression sequences, live fingerprint pressure change curves, and iris dynamic texture video streams; the campus spatiotemporal trajectory data includes a daily mobile heat map within the campus electronic fence over a recent period, classroom attendance-related locations, and matching degree of teaching activities.

3. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1, characterized in that, The three-dimensional spatiotemporal alignment preprocessing of the dynamic biometric stream specifically involves: extracting the temporal dimension frame sequence of the dynamic biometric stream. The illumination stability coefficient of each frame is calculated using a dynamic frame selection algorithm that adapts to campus illumination. ,in , These are the scene adaptation coefficients, Indicates the first The average illumination intensity of the frame, Indicates the first Frame contrast entropy value, filtered out keyframe sequence ,in To determine the illumination threshold for campus scenes, a cross-modal feature correlation matrix is ​​constructed. Complementary features are generated through an improved adversarial generative network, where the generator... The loss function is , For discriminator, It is a single modal feature. For multimodal features, , These are the modal equilibrium coefficients. Representing feature similarity for keyframe sequences Spatial coordinate normalization is performed to obtain a spatiotemporally correlated biofeature matrix. , where matrix elements For spatiotemporal calibration coefficients, For the first The time weighting factor of a frame.

4. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1, characterized in that, The construction of the campus twin multimodal model based on knowledge distillation specifically involves: aggregating historical feature data from each campus using a federated learning framework through a teacher network T, wherein the loss function of the teacher network is: ,That , These are the corresponding weighting coefficients. For classification loss function, To determine the loss function based on the differences in characteristic distributions across campuses, dynamic forgetting factor models are constructed for three groups: full-time students, faculty, and visitors. The forgetting factor is: ,in As the initial forgetting factor, This is the forgetting decay coefficient. Given the time difference since the most recent feature update, when anchoring campus scene-specific feature points for each modality using an attention mechanism, the calculation is performed for the... The attention weights for each modality are: , and For learnable attention parameters, For the first The feature maps of each modality are processed, and the attention-weighted features of each modality are then input into the feature fusion layer to obtain a structured feature map. for: , This indicates a feature concatenation operation. This represents the total number of modes.

5. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1, characterized in that, When introducing a campus identity confidence transmission mechanism, campus spatiotemporal trajectory data is transformed into behavioral feature vectors through an LSTM network. ,in For the first When constructing a correlation graph between biological and behavioral features using an improved graph neural network based on the spatiotemporal trajectory features of each time window, an adjacency matrix is ​​defined. The elements are ,in For the first A biological feature vector, For the first A behavioral feature vector, For activation function, These are learnable parameters.

6. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1 or 5, characterized in that, Dynamic confidence weights for each modal feature are calculated using PageRank values ​​based on graph nodes, and then iteratively formulated. Calculate the node importance score, where For PageRank vectors, This is the adjacency matrix of the association graph. The damping coefficient is... The initial importance distribution is given; the dynamic confidence weights for each modality feature are... ,in The weight matrix is ​​a learnable matrix, forming a feature fusion formula. , For the first Dynamic confidence weights for each modal feature This represents the total number of modes fused.

7. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1, characterized in that, When comparing campus identity feature chains stored on blockchain, the feature template fragments of the corresponding security domain are invoked through smart contracts. ,in This represents the security domain index, which, combined with a homomorphic encryption similarity calculation algorithm optimized for campus scenarios, calculates the features to be verified at the edge nodes. With template segmentation The similarity score is calculated using the following formula: ,in It is a homomorphic encryption function. For the encryption domain similarity calculation, generate a verification result containing a confidence interval, with the upper and lower limits of the confidence interval being... ,in , For historical comparison scoring, The mean, The dynamic calibration factor output by the campus safety incident tree model. This represents the number of historical comparison score samples used when calculating the confidence interval.

8. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1, characterized in that, The multi-level verification channels include remote verification by counselors and cross-verification of the course schedule. When the lower limit of the confidence interval of the verification result is lower than the preset basic threshold but higher than the suspicious threshold, a secondary verification based on the campus social relationship graph is automatically triggered. The counselor remote verification pushes a verification anomaly reminder containing a spatiotemporal trajectory heatmap and performs class time matching verification. When the lower limit of the confidence interval is lower than the suspicious threshold, a tertiary verification is initiated. The teaching management system retrieves the class attendance records for the preset time period and performs continuity verification with the current location, and pushes a risk warning containing historical behavior profiles to the security department.

9. The method for accurate identity verification of smart campus access control based on multimodal biometrics according to claim 1 or 8, characterized in that, When using temporal anomaly detection algorithms to achieve early warning of potential risks, the specific steps are: constructing a temporal state space of feature evolution trajectories. ,in Indicates the first Each feature dimension at time... Given the state value, calculate the Mahalanobis distance between the current state and the historical normal pattern, using the following formula: ,in and These are the mean vector and covariance matrix of the historical characteristic distribution, respectively. When continuous Each time step satisfies When an early warning is triggered, among which The standard deviation of the Mahalanobis distance. This is the early warning threshold coefficient that is dynamically adjusted based on the frequency of campus safety incidents.

10. A smart campus access control system for accurate identity verification based on multimodal biometrics, applicable to claims 1-9, characterized in that, It includes a multimodal data acquisition module, a feature preprocessing module, a twin model construction module, a confidence level transmission module, a blockchain comparison module, and a response and early warning module; The multimodal data acquisition module collects dynamic biometric data and campus spatiotemporal trajectory data through distributed terminals and links with the campus Internet of Things sensing network. The input of the feature preprocessing module is connected to the multimodal data acquisition module, which is used to perform three-dimensional spatiotemporal alignment preprocessing on the dynamic biofeature stream and generate a spatiotemporally correlated biofeature matrix. The input of the twin model construction module is connected to the feature preprocessing module, which is used to construct a knowledge distillation campus twin multimodal model and output a structured feature map. The confidence transfer module is connected to the multimodal data acquisition module and the twin model construction module respectively, and is used to convert campus spatiotemporal trajectory data into behavioral feature vectors, construct a correlation map between biological features and behavioral features, and calculate dynamic confidence weights. The input of the blockchain comparison module is connected to the confidence transmission module, which is used to call the security domain feature template fragment and complete the localized comparison through the homomorphic encryption similarity algorithm to generate a verification result containing a confidence interval; The input of the response and early warning module is connected to the blockchain comparison module, which is used to execute differentiated access control responses and trigger multi-level verification channels based on the verification results, and to achieve early warning of potential risks through time-series anomaly detection.

Images