Multi-modal biological recognition access verification method and system and storage medium

Abstract

The invention relates to the technical field of information security, and discloses a multi-modal biological recognition access verification method and system and a storage medium, and the method comprises the steps: collecting multi-modal biological information through terminal equipment, and carrying out the quality evaluation of each type of modal information, and obtaining a corresponding quality score; based on the quality score, the terminal equipment determines a target oracle node type, constructs a differentiated verification request, sends the differentiated verification request to a corresponding oracle node, and forwards the differentiated verification request to a back-end service system; and the back-end service system constructs the feature matching result, the quality score and the routing information into a context feature vector, inputs the context feature vector into a fusion decision model to generate a final verification conclusion, and finally executes authorization operation according to the conclusion. According to the invention, the verification accuracy and the operation efficiency of the system in a complex environment are effectively improved through the fusion decision of the adaptive routing decision and the context awareness, and the safety and the traceability of the whole link are enhanced at the same time.

Description

Technical Field

[0001] This invention relates to the field of information security technology, specifically to a multimodal biometric access verification method, system, and storage medium. Background Technology

[0002] Multimodal biometrics, due to its higher security and accuracy compared to unimodal methods, has been widely used in identity authentication, access control, and financial payments. This technology typically combines two or more biometric features, such as facial features, fingerprints, and iris scans, to improve the reliability of verification by fusing information from multiple dimensions.

[0003] In a standard verification process, the terminal device collects various biometric data, extracts their features, and sends them to a backend server for comparison. After obtaining the comparison scores for each modality, the server uses a pre-defined fusion strategy, such as weighted summation, to calculate the final comprehensive score and makes an authorization or rejection decision accordingly.

[0004] However, existing technologies generally employ fixed processing flows and fusion logic. This static mechanism lacks dynamic awareness of the quality of information at the verification front end. For example, when users perform verification in poor lighting conditions or with dirty fingers, the quality of the collected biometric information will significantly decrease. However, existing systems still process this low-quality information according to standard procedures and incorporate its comparison results into fixed fusion calculations. This may lead to erroneous rejections due to the uncertainty of low-quality information, reducing user experience and system usability.

[0005] Furthermore, this fixed processing mode fails to fully utilize the benefits of high-quality information. Even under favorable acquisition conditions and with clear and reliable information across all modalities, the system still executes the entire verification process designed to handle complex situations, resulting in unnecessary computational resource and time overhead. Therefore, the processing logic of the existing method is relatively rigid, unable to dynamically adjust its security strategies and resource allocation based on the quality of real-time acquired information. Consequently, the system's adaptability, efficiency, and robustness in dealing with complex and ever-changing real-world application environments need improvement. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a multimodal biometric access verification method, system, and storage medium, which solves the problem that existing technologies, which employ fixed fusion strategies, lack the ability to perceive the quality of information collected at the front end, thus making it difficult to balance verification reliability and system efficiency.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a multimodal biometric access verification method. This method performs real-time quality assessment of the collected multimodal biometric information locally on the terminal device, and adaptively selects different verification paths and constructs differentiated verification requests based on the assessment results. Subsequently, the backend service system combines the verification path, information quality and matching results to make context-aware fusion decisions, thereby realizing a verification process that balances security, efficiency and accuracy.

[0008] According to one embodiment of the present invention, the method includes:

[0009] S1: Multimodal biological information is collected by terminal devices;

[0010] S2: The terminal device performs a quality assessment on each modality of information in the multimodal biological information to obtain a corresponding quality score;

[0011] S3: The terminal device determines the target oracle node type based on the quality score;

[0012] S4: The terminal device constructs and sends a verification request to the oracle node of the target oracle node type according to the target oracle node type;

[0013] S5: The oracle node forwards the verification request to the backend service system, which performs matching based on the verification request and generates a final verification conclusion.

[0014] S6: Based on the final verification conclusion, perform the corresponding authorization operation.

[0015] In some implementations, the quality assessment process is completed by quantifying and weighting preset quality parameters. For example, facial quality score. and fingerprint quality score The calculation can be expressed by the following formula:

[0016] ;

[0017] ;

[0018] in, These are the normalized quality parameter values ​​for illumination, pose, and sharpness; These are the preset weights for the corresponding parameters.

[0019] In some implementations, the process of determining the target oracle node type is based on a comparison of a quality score with a preset threshold. A high-quality threshold is set... and low quality threshold When the quality scores of both modal information are higher than When the scene is deemed to be of high quality, a standard node is selected; when the quality score of one modality is higher than... And another type is lower When the scenario is deemed a high-quality, one-sided scenario, a high-security node is selected; when the quality scores of both modalities are lower than [a certain value], [the scenario is considered a high-quality, one-sided scenario]. When a scenario is identified as having two low-quality aspects, an audit node is selected. This process enables the verification path to dynamically match the reliability of the front-end information.

[0020] In some implementations, the construction of the verification request differs. For standard nodes, the verification request primarily contains a feature hash value; for high-security nodes, the verification request, in addition to the feature hash value, also carries the original quality score to allow the backend to make a more comprehensive decision.

[0021] In some implementations, to ensure data security, the terminal device extracts biometric vectors from the original biometric information locally and generates irreversible feature hash values ​​using salted hashing. This process can be represented as follows: ,in For biological feature vectors, This is a random salt value.

[0022] In some implementations, the process by which the backend service system generates the final verification conclusion includes: matching the feature hash value in the request with a pre-registered hash template to obtain a comparison confidence score; and then constructing a context feature vector together with the score, the quality score parsed from the request, and the routing information representing the verification path. ,For example

[0023] ;

[0024] Finally, the vector is input into a pre-trained fusion decision model, such as a support vector machine, gradient boosting decision tree model, or feedforward neural network, to output the final verification conclusion.

[0025] A second aspect of the present invention provides a multimodal biometric access verification system, the system comprising:

[0026] The acquisition module is used to acquire multimodal biological information;

[0027] The evaluation module is used to evaluate the quality of each modality of information in the multimodal biological information and obtain the corresponding quality score;

[0028] The decision module is used to determine the target oracle node type based on the quality score, and to construct a verification request based on the target oracle node type;

[0029] The sending module is used to send the verification request to the oracle node of the target oracle node type;

[0030] The forwarding module is configured to receive the verification request and forward it.

[0031] The verification module is configured to receive the forwarded verification request, perform matching based on the verification request, and generate a final verification conclusion.

[0032] The execution module is used to perform the corresponding authorization operation based on the final verification conclusion.

[0033] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect of the present invention.

[0034] In summary, this application includes at least one of the following beneficial technical effects:

[0035] 1. This invention performs real-time quality assessment of collected multimodal biological information locally on the terminal device and adaptively selects different verification paths based on the assessment results, matching appropriate processing resources to verification scenarios of different quality levels. This method avoids unnecessary deep processing of high-quality information, improving the overall operating efficiency and response speed of the system while ensuring security.

[0036] 2. This invention, through a backend fusion decision-making mechanism, not only relies on the matching results of biometric features but also comprehensively considers the quality of the original information uploaded from the front end and the verification path traversed by the request. This context-aware, multi-dimensional adjudication method can effectively reduce verification misjudgments caused by poor acquisition of single-modal information, thereby significantly improving the verification accuracy and robustness of the system in complex environments such as changes in lighting and poor posture.

[0037] 3. This invention ensures that the original biometric information does not leave the terminal device and that the data in the transmission link is not parsed by performing feature extraction and salted hashing locally on the terminal, combined with secure relay forwarding by oracle nodes. Simultaneously, by mandating auditing and logging of low-quality verification events, a secure closed loop is constructed from information collection to final authorization, strengthening data security and the traceability of critical events throughout the entire chain. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method in this application;

[0039] Figure 2 This is a system architecture diagram for this application. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example:

[0042] Please see the appendix Figure 1 This invention provides a multimodal biometric access verification method, the execution entity of which can be a verification system comprising terminal devices, an oracle cluster, and a backend service system. The method may include the following steps:

[0043] First, S1 is executed, in which the terminal device deployed at the user interaction front end collects at least two kinds of biometric information of the user in parallel through its integrated multimodal acquisition module 110, such as fingerprint image and facial image, and performs quality assessment and encryption processing on the biometric information locally through the preprocessing module 120.

[0044] Next, in step S2, the routing decision module 140 of the terminal device executes the preset routing decision logic based on the quality assessment results of each modality information output by the preprocessing module 120 to determine the target oracle node type of this verification request and construct a differentiated verification request containing encrypted biometric information.

[0045] Subsequently, in step S3, the communication module 130 of the terminal device sends the constructed verification request to the oracle cluster 200. The oracle node 210 of the specified type in the cluster receives the request and forwards it as a secure relay to the secure database 310 within the backend service system 300 for comparison.

[0046] Then, S4 is executed. The fusion decision module 320 in the backend service system 300 obtains the database comparison results and integrates the original information quality data from the terminal device and context information such as the routing path of the request. It is then input into a pre-trained fusion decision model to generate a final verification conclusion.

[0047] Finally, S5 is executed. The terminal device receives the final verification conclusion and sends instructions to the associated physical execution unit (such as the electronic door lock 400) to complete the authorization or denial of access operation. At the same time, the entire process information of this verification is generated into a log record to form an audit closed loop.

[0048] Based on the above steps, the overall workflow of the method provided in this embodiment of the invention can be summarized as follows: When verification is triggered, the terminal device not only collects the user's multimodal biometric information, but also performs quality quantification locally, and uses this quantification result as the core basis to determine the transmission path of the verification information stream in the distributed oracle network. This design ensures that the verification request carries metadata about its inherent reliability from the very beginning, and the backend decision-making system can make a comprehensive judgment based on multi-dimensional information such as biometric matching degree, original information quality, and transmission path security level. This process deeply couples the front-end perception, the mid-end transmission, and the back-end decision-making to form a dynamic and adaptive security verification system.

[0049] This invention provides a multimodal biometric access verification system, which can be applied to a verification system including terminal devices, oracle clusters, and backend service systems. The method may include the following steps:

[0050] S1. Collect and preprocess multimodal biometric information. The multimodal acquisition module 110 on the terminal device collects various biometric information of the user, such as fingerprints and facial features, and the preprocessing module 120 performs a quantitative evaluation of the quality of the collected information locally, while extracting features and generating corresponding salted hash values.

[0051] S2, Execute adaptive routing decision. The routing decision module 140 in the terminal device executes the preset routing judgment logic based on the quality scores of each modality information obtained in S1 to determine the target node type of this verification request in the oracle cluster 200, and constructs a verification request containing encrypted feature information and quality metadata accordingly.

[0052] S3, relay the request and perform a secure comparison. The communication module 130 of the terminal device sends the verification request to the oracle cluster 200. After receiving the request, the corresponding oracle node 210 in the cluster forwards it to the secure database 310 in the backend service system 300. The secure database 310 performs a comparison operation on the encrypted hash value.

[0053] S4. Perform context-aware fusion decision-making. The fusion decision module 320 in the backend service system 300 integrates the comparison results from the security database 310, the original information quality data from the verification request, and context information such as the transmission path type of the request, and inputs this multi-dimensional information into a fusion decision model to generate a final verification conclusion.

[0054] S5, Execute the verification conclusion and record the log. The terminal device receives the final verification conclusion generated in S4 and sends control commands to the associated physical execution unit 400 based on the conclusion. At the same time, the system generates a traceable audit log for the entire verification process.

[0055] In S1, multimodal biological information is collected and localized for preprocessing, which may include the following sub-steps.

[0056] First, the terminal device receives a verification trigger command. This command can originate from the user selecting a verification operation on the terminal device's interface, or from the terminal device detecting that the user is within a preset valid interaction range using its configured distance sensor, infrared sensor, or other means.

[0057] Next, the terminal device responds to the verification trigger command and activates its internally integrated multimodal acquisition module 110. The multimodal acquisition module 110 integrates at least two different types of biometric sensors to acquire user biometric characteristics from different dimensions.

[0058] In one embodiment, the sensor may be a combination of a visible light camera for acquiring facial images and an optical or capacitive fingerprint sensor for acquiring fingerprint images. In other embodiments, to cover a wider range of biometric combinations, the sensor integrated into the multimodal acquisition module 110 may also be a dedicated infrared camera for acquiring iris images, a microphone array for acquiring voiceprint information, or a near-infrared imaging device for acquiring finger vein or palm vein images. This invention does not limit the specific modal combination of biometric information; any scenario using two or more different biometrics for verification falls within the protection scope of this invention.

[0059] Different types of biosensors are configured to initiate their workflows in parallel or near-parallel to capture the user's biometrics within contiguous time windows. Parallel acquisition aims to reduce user wait times and ensure temporal consistency of the acquired modal information to reflect the user's true state at the moment of verification.

[0060] Then, the activated sensors collect the corresponding raw biometric information and transmit the collected data to the preprocessing module 120 of the terminal device. Raw biometric information refers to the data format directly output by the sensors without feature extraction. For example, a facial image can be one or more frames of digital image data, and a fingerprint image can be grayscale image data representing the ridges and valleys of a fingerprint.

[0061] After receiving the raw biological information, the preprocessing module 120 performs a real-time, localized quantitative assessment of the quality of each modality. This assessment process aims to provide objective data for subsequent adaptive routing decisions.

[0062] For facial images, the preprocessing module 120 evaluates a set of preset quality parameters, including lighting conditions, facial pose, and image sharpness. The values ​​of these parameters can be calculated using conventional image processing algorithms in the field, such as evaluating lighting by analyzing the image's brightness histogram and determining the pose deflection angle by detecting facial key points.

[0063] After obtaining the values ​​of each quality parameter, a weighted model is used to calculate the final facial quality score. Its calculation formula can be expressed as:

[0064] ;

[0065] in:

[0066] This represents the final facial image quality score;

[0067] These represent the normalized values ​​for illumination, pose, and sharpness parameters, respectively.

[0068] These are the preset weights for each parameter, and the sum of all weights is 1.

[0069] Similarly, for fingerprint images, the preprocessing module 120 evaluates parameters such as image clarity, effective pressing area, and the number of feature points contained therein.

[0070] Fingerprint quality score The calculation formula can be expressed as:

[0071] ;

[0072] in:

[0073] This represents the final fingerprint image quality score;

[0074] These represent the normalized sharpness, effective area, and number of feature points, respectively.

[0075] These are the preset weights for each parameter, and the sum of all weights is 1.

[0076] In a specific implementation, each mass parameter is normalized to a preset numerical range, such as [0,1], before being substituted into the formula for calculation, in order to eliminate the influence of different parameter dimensions.

[0077] Finally, the preprocessing module 120 outputs a set of quantified quality scores, such as... and This is for use by the subsequent routing decision module 140.

[0078] After completing the quality assessment, or concurrently, the preprocessing module 120 performs feature extraction and encryption on the raw biological information. This step ensures that the raw, sensitive biological information is not involved in subsequent transmission and comparison processes.

[0079] The preprocessing module 120 calls a pre-defined biometric extraction algorithm to extract structured feature vectors from the original fingerprint and facial images, respectively. Feature vectors are a set of data obtained after dimensionality reduction and abstraction of the original information, used to characterize the biological uniqueness of an individual. For example, the feature vector of a fingerprint... It can be a set of data describing the location and orientation of fingerprint ridge endpoints and bifurcation points; facial feature vectors It can be a high-dimensional floating-point vector extracted from a face image using a deep convolutional neural network.

[0080] The specific implementation of feature extraction, such as extracting facial feature vectors through CNN models or locating fingerprint detail feature points through image processing algorithms, are well-known technologies in this field and will not be elaborated here.

[0081] In obtaining feature vectors and Then, the preprocessing module 120 performs a salted hash operation to generate an irreversible hash digest for transmission.

[0082] In S1, after extracting the original biometric vector, the preprocessing module 120 performs salted hashing on the biometric vector to generate an irreversible feature hash value, thereby ensuring the security of the original biometric information. This process can be represented by the following formula:

[0083] ;

[0084] ;

[0085] in:

[0086] and These are the final generated fingerprint feature hash value and facial feature hash value, respectively;

[0087] This represents a cryptographically secure hash function, such as SHA-256, SHA-3, or SM3;

[0088] and This is the original biological feature vector extracted in the previous step;

[0089] and These are the salt values ​​corresponding to fingerprints and faces, respectively.

[0090] This indicates the operation of concatenating the byte sequence of the feature vector with the byte sequence of the salt value.

[0091] To ensure the feasibility and security of the verification process, this embodiment has specifically designed a management mechanism for the salt value. The salt value... and It is a unique random data string generated by the terminal device for the user when the user first registers their biometric information on the terminal device.

[0092] After registration, the backend service system only stores the generated hash template. and Correspondingly, salinity and The salt value is then securely stored in the local storage area of ​​the terminal device, such as the system's secure element or trusted execution environment, and the salt value itself is not transmitted to the backend service system over the network.

[0093] In each subsequent verification process, when the preprocessing module 120 extracts a new biometric vector... and Then, it will read the user's proprietary, saved salt value from the terminal device's local secure storage area. and Subsequently, the read salt value is concatenated with the newly extracted feature vector, and the hash value of this verification is calculated.

[0094] This salt value management method ensures that the salt value used to generate the hash template during registration is completely consistent with the salt value used to generate the hash value to be compared during verification, thus guaranteeing the correctness of the backend hash comparison operation. Furthermore, since the salt value does not leave the terminal device, this solution can effectively resist common security threats such as rainbow table attacks targeting hash databases, further enhancing the overall security of the system.

[0095] Finally, the preprocessing module 120 will generate the hash value. , and the aforementioned mass fraction , They are output together as input to the adaptive routing decision in S2.

[0096] In S2, the terminal device performs adaptive routing decisions. Specifically, the routing decision module 140 within the terminal device receives quality scores for each modality, such as fingerprint quality scores, output by the preprocessing module 120. and facial quality score .

[0097] The routing decision module 140 deploys a routing decision model. In this embodiment, this model is manifested as a set of conditional judgment rules based on preset thresholds. Its core judgment logic relies on two key numerical thresholds: a high-quality threshold and a low-quality threshold. and a low quality threshold . Information used to define whether a modality is reliable enough to serve as strong evidence; This is used to determine whether the information quality of a modality is too low, may be invalid, or may be at risk of being forged.

[0098] The specific logic for this determination is as follows:

[0099] When the quality scores of both modes are higher than the high quality threshold, the condition is met. At that time, the routing decision module 140 determined that this verification was a dual high-quality scenario and identified the target routing node as a standard node. .

[0100] The condition is met when the quality score of one mode is above the high quality threshold, while the quality score of another mode is below the low quality threshold. At that time, the routing decision module 140 determined it to be a one-sided high-quality scenario and identified the target routing node as a high-security node. .

[0101] When the quality scores of both modes are below the low quality threshold, the condition is met. At that time, the routing decision module 140 determined it to be a double low-quality scenario and identified the target routing node as the audit node. .

[0102] In other quality score distribution scenarios, such as when the quality scores of one or two modes are between the low quality threshold... With high quality threshold In this case, the routing decision module 140 defaults to classifying it as a standard scenario and determines the target routing node as a standard node. .

[0103] By executing the above decision logic, the routing decision module 140 finally outputs an identifier representing the type of the target oracle node. This identifier will be used to guide the construction and sending of the verification request in the next step.

[0104] After the routing decision module 140 determines the target oracle node type, it constructs a differentiated verification request corresponding to the determination result. The differentiation is reflected in the differences in the data structure and content of the request payload.

[0105] If the scenario is determined to be of high quality, the routing decision module 140 constructs a standard verification request payload. The load This may include: a unique device identifier A request timestamp and the feature hash values ​​of the two modalities. and .

[0106] If the scenario is determined to be a one-sided high-quality scenario, the routing decision module 140 constructs a high-security verification request payload. The load This may include: unique device identifier Request timestamp A hash value marked as primary evidence (Corresponding to high-quality modality), a hash value marked as supporting evidence. (corresponding to the low-mass mode), and the original mass fractions of the two modes. and .

[0107] If the scenario is determined to be a double low-quality scenario, the routing decision module 140 constructs an audit log request payload. The load This may include: unique device identifier Request timestamp Two feature hash values and and two low quality scores and .

[0108] After constructing the request payload, the communication module 130 of the terminal device selects and connects to a corresponding node in the oracle cluster 200 based on the target node type identifier output by the routing decision module 140. For example, if the target is a standard node... The communication module 130 then sends signals to one or more... One of the nodes sends the constructed request payload. The selection process can be based on a load balancing or random round-robin strategy. Subsequently, the communication module 130 sends the request payload out via a secure communication protocol.

[0109] In S3, an oracle node 210 of a specific type in the oracle cluster 200 receives verification requests sent by the end devices. The core function of oracle node 210 is to act as a secure relay between the end devices and the backend service system 300.

[0110] Specifically, when oracle node 210 receives a request, its primary function is to perform data forwarding. This node is configured not to parse sensitive data content in the request payload. Specifically, for biometric hash values ​​contained in the payload, such as... and Oracle node 210 does not perform any decryption or semantic analysis operations; it simply forwards the data as a block.

[0111] To ensure security, oracle node 210 does not persistently store request payloads containing sensitive information on its non-volatile storage media (such as hard disks). Any temporary storage of payload content is limited to volatile memory (such as RAM) during the execution of forwarding tasks.

[0112] Oracle node 210 forwards the received complete request payload to a designated network address within the backend service system 300 via a pre-established secure communication channel. This address corresponds to the receiving interface of the backend secure database 310. This forwarding process ensures the confidentiality and integrity of data throughout the link from the oracle node to the backend service system.

[0113] The security database 310 within the backend service system 300 receives a verification request forwarded by the oracle node 210.

[0114] Security Database 310 parses the request payload and extracts the device's unique identifier. and one or more biometric hash values, such as fingerprint hash values. and facial feature hash value .

[0115] in accordance with The security database 310 retrieves the user's pre-registered biometric hash template associated with the identifier from its storage space, and records it as the fingerprint hash template. and facial hash template .

[0116] Subsequently, the security database 310 performs a hash template matching operation. Specifically, this operation involves a strict, bit-by-bit comparison between the received feature hash value and the retrieved hash template. This comparison process can be represented as:

[0117] ;

[0118] ;

[0119] in, The function outputs a data structure containing the alignment results, for example:

[0120] ;

[0121] ;

[0122] and This is a Boolean value indicating whether the comparison is successful. and If they are exactly the same, then If true, then false; otherwise, then false. Similarly.

[0123] and The corresponding comparison confidence score is assigned. In one embodiment, when the comparison matches, the confidence score is set to a preset high value (e.g., 1.0); when the comparison does not match, it is set to a preset low value (e.g., 0.0).

[0124] After the comparison is completed, the security database 310 will generate the comparison results. and It is sent to the fusion decision module 320 in the backend service system 300 for use in the next step.

[0125] In S4, the fusion decision module 320 within the backend service system 300 performs context-aware fusion decision-making. This module first integrates information from different dimensions to construct a comprehensive context feature vector. This information is used by the subsequent fusion decision-making model for comprehensive judgment.

[0126] In order to provide the model with a set of effective and non-redundant decision criteria, this contextual feature vector It mainly consists of the following two types of core information:

[0127] Comparison confidence information: that is, the fingerprint comparison confidence score generated after performing a hash comparison using the security database 310. Confidence score compared with face This information reflects the degree of matching between the collected biometrics and the pre-registered template.

[0128] Raw quality score information: This refers to the fingerprint quality score evaluated by the terminal device and uploaded with the verification request. and facial quality score This information reflects the original reliability of the input source used in this verification.

[0129] Therefore, context feature vector It can be represented as:

[0130] ;

[0131] In this embodiment, the one-hot encoding representing the routing path (such as...) is no longer used. This is incorporated into the context feature vector. This is because the routing path itself is determined by a quality score. and Features derived from fixed rules have a definite functional relationship with quality scores. Introducing them again as features can cause information redundancy (i.e., multicollinearity), potentially interfering with the training process of the fusion decision model and the stability of decision weight allocation.

[0132] By inputting only continuous quality score values, fusion decision models (such as gradient boosting decision trees or neural networks) can learn autonomously during training and identify the nonlinear impact of different quality score ranges on the final decision, thereby establishing a more accurate and robust decision boundary without relying on a discretized routing label that has been preprocessed by the front end.

[0133] For scenarios requiring mandatory auditing (i.e., the path from which requests originate), the backend service system can identify the source path upon receiving a request. This path information is used to trigger a separate, mandatory logging program, rather than being used as input features in the computation of the fusion decision model. This design decouples the decision logic from the auditing strategy, ensuring the clarity of the system architecture and the rigor of the logic.

[0134] In S5, the system performs subsequent operations based on the final verification conclusion generated by the fusion decision module 320.

[0135] First, the terminal device receives the final verification conclusion returned by the backend service system 300 through its communication module 130.

[0136] Next, the terminal device analyzes the conclusion to determine the specific authorization level. For example, the analyzed conclusion could be high-confidence authorization, explicit denial, or low-confidence authorization.

[0137] Then, the terminal device generates a corresponding control command based on the parsed authorization level. If the conclusion is high-confidence authorization, a command to unlock or allow access is generated. If the conclusion is explicit denial, no authorization command is generated, and the user can be prompted with a verification failure message on its user interface. If the conclusion is low-confidence authorization, a command to perform limited-permission operations can be generated, or a two-factor authentication process can be triggered.

[0138] Finally, the terminal device sends the generated control command to an associated physical execution unit 400. The physical execution unit 400 is the device responsible for executing state changes in the physical world. In one embodiment, the physical execution unit 400 can be the drive circuit of an electronic door lock. In other embodiments, it can be the control system of a parking lot barrier gate, or the electronic lock cylinder of a safe. After receiving the control command, the physical execution unit 400 performs the corresponding physical action, such as driving the motor to retract the door lock's bolt, to complete the door opening operation.

[0139] During or after the physical operation, the logging module 330 in the backend service system 300 generates a complete event log record for this verification event.

[0140] This event log aims to solidify key information throughout the entire process, from information gathering to final decision-making, to achieve system traceability and auditability. The specific content of the log may include:

[0141] Unique device identifier from the authentication request With request timestamp ;

[0142] Raw quality scores of each modality uploaded by the terminal device , ;

[0143] Represents the routing information of the transmission path of the verification request;

[0144] Comparison results generated by Security Database 310 , ;

[0145] The context feature vector constructed by the fusion decision module 320 ;

[0146] And the final verification conclusions output by the fusion decision model.

[0147] Event log entries are sent and stored in a dedicated, tamper-proof log database. In scenarios with even higher requirements for data immutability, this storage medium can also be a distributed ledger system.

[0148] These log records are not only used for post-event manual auditing or security forensics, but also form the basis for system self-optimization. In a preferred embodiment, these log records with clearly labeled results can be collected periodically and used as incremental training data to iteratively optimize the fusion decision model, thereby forming a data-driven closed loop for system performance improvement.

[0149] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be covered within the scope of protection of this application.

Claims

1. A multi-modal biometric access verification method, characterized by, The method comprises the following steps: S1: collecting multi-modal biological information by a terminal device; S2: performing quality assessment on each modality of the multi-modal biological information by the terminal device to obtain a corresponding quality score; S3: determining a target oracle node type based on the quality score by the terminal device; S4: constructing and sending a verification request to an oracle node of the target oracle node type according to the target oracle node type by the terminal device; S5: forwarding the verification request to a backend service system by the oracle node, performing matching based on the verification request by the backend service system, and generating a final verification conclusion; S6: performing a corresponding authorization operation according to the final verification conclusion.

2. The method of claim 1, wherein, The quality assessment on each modality of the multi-modal biological information to obtain a corresponding quality score comprises: evaluating at least one preset quality parameter dimension from the following aspects: illumination condition, image posture, image sharpness, image definition, and effective pressing area; and calculating the quality score through a weighted model.

3. The method of claim 1, wherein, The determination of the target oracle node type based on the quality score comprises: presetting a high quality threshold and a low quality threshold; when the quality scores of at least two modalities are both higher than the high quality threshold, determining that the target oracle node type is a standard node; when the quality score of at least one modality is higher than the high quality threshold, and the quality score of at least another modality is lower than the low quality threshold, determining that the target oracle node type is a high-security node; when the quality scores of at least two modalities are both lower than the low quality threshold, determining that the target oracle node type is an audit node.

4. The method of claim 3, wherein, The construction and sending of the verification request according to the target oracle node type further comprises: when the target oracle node type is a standard node, the verification request contains feature hash values of at least two modalities; when the target oracle node type is a high-security node, the verification request contains the quality score in addition to the feature hash values.

5. The method of claim 4, wherein, Before determining the target oracle node type, the method further comprises: extracting a corresponding biological feature vector from the multi-modal biological information; performing salted hash processing on the biological feature vector to obtain the feature hash value.

6. The method of claim 1, wherein, The matching based on the verification request by the backend service system and the generation of the final verification conclusion comprise: matching the feature hash values in the verification request with pre-registered hash templates to obtain a comparison confidence score; constructing a context feature vector together from the comparison confidence score, the quality score carried in the verification request, and routing information representing the target oracle node type; inputting the context feature vector into a fusion decision model to generate the final verification conclusion.

7. The method of claim 6, wherein, The fusion decision model is a multi-class classifier obtained through offline training; the multi-class classifier comprises one of a support vector machine, a gradient boosting decision tree model, or a feedforward neural network.

8. The method of claim 3, wherein, When the target oracle node type is an audit node, the verification request further includes the quality score in addition to the feature hash value; and the backend service system, while generating the final verification conclusion based on the verification request, forcibly generates a high-risk event log record of the verification request and its key information and stores the high-risk event log record.

9. Multi-modal biometric access verification system according to the method of any of claims 1-8, characterized in that, The method comprises: a collection module configured to collect multi-modal biological information; an evaluation module configured to evaluate the quality of each modality of the multi-modal biological information to obtain a corresponding quality score; a decision module configured to determine a target oracle node type based on the quality score and construct a verification request according to the target oracle node type; a sending module configured to send the verification request to an oracle node of the target oracle node type; a forwarding module configured to receive and forward the verification request; a verification module configured to receive the forwarded verification request and generate a final verification conclusion based on the verification request; an execution module configured to perform a corresponding authorization operation according to the final verification conclusion.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the method of any one of claims 1-8.

Images