A bank service gate control system based on multi-modal identity verification

Abstract

The present application relates to the technical field of identity recognition access control, and particularly relates to a bank business library access control system based on multi-modal identity verification, comprising a set construction module, a sample acquisition module, a consensus statistics module, a threshold determination module, a decision passing module and a set updating module; by constructing a historical effective feature set to replace the traditional single centralized template, the identity authentication is changed into the consensus discrimination between the current sample and multiple historical samples, and a strict double threshold for entering the pool is set, thereby solving the template pollution problem driven by cross-modal confirmation in the prior art, effectively avoiding the cumulative risk of template chronic drift, ensuring that the historical effective feature set remains pure for a long time, thereby effectively avoiding the problems of sudden increase of high rejection rate and potential increase of security risk caused by template pollution during long-term operation of the system, and significantly improving the long-term operation stability and security of the access control system.

Description

Technical Field

[0001] This invention relates to the technical field of identity recognition access control, and in particular to a banking vault access control system based on multimodal identity verification. Background Technology

[0002] In access control systems of high-security locations such as banks and vaults, multimodal identity verification technology effectively improves the accuracy and resistance to attacks by integrating multiple biometric features such as face, fingerprint, finger vein, and iris scans, as well as auxiliary authentication factors such as employee badges and passwords. To adapt to the slow and natural changes in users' biometric features and improve the long-term user experience, some systems have introduced a template adaptive update mechanism. The conventional approach is as follows: after the multimodal fusion decision determines that the user's identity is legitimate, the corresponding modal baseline template stored in the background is updated using the successfully collected biometric samples through weighted averaging or other methods, so that the template gradually converges to the current features.

[0003] However, existing template update strategies typically use the final fusion decision as the sole condition for triggering an update. This presents a hidden and cumulative technical flaw in high-security environments. Specifically, in real-world environments such as banking databases, the quality of a single biometric modality often experiences temporary or persistent degradation due to infrequent, atypical factors. For example, the fingerprint collection window may have subtle, barely noticeable stains, or employees' hands may be dry and cracked in winter, or have abnormal skin conditions due to disinfectant residue. In such cases, even if the quality of that modality (such as fingerprint) is extremely low and difficult to authenticate independently, the multimodal fusion decision will still determine the user's identity as legitimate because other modalities (such as facial recognition) have extremely high confidence. In this situation, existing technologies indiscriminately use the low-quality or even erroneous samples collected in this instance to update the backend template. This unchecked update of the target modality, driven by strong confirmation from other modalities, constitutes a cross-modal confirmation-driven template contamination chain.

[0004] While a single low-quality update has a negligible impact on the template and won't immediately trigger system anomalies, over weeks or months of continuous accumulation, the background template drifts due to the constant absorption of low-quality samples. Its feature distribution gradually deviates from the user's true biometric state, or degenerates into a more ambiguous and generalized form. When a user attempts authentication in a normal state, the authentication fails due to significant differences from the contaminated template, leading to frequent rejections of legitimate users and a surge in false rejection rates. The template's discrimination boundary degrades due to contamination, and its feature uniqueness decreases, potentially increasing the risk of others with similar characteristics being incorrectly accepted, thus raising potential security risks. Currently, industry improvements mainly focus on improving the quality of single data collection or optimizing fusion algorithms, but these are insufficient to address the long-term, cumulative template contamination problem caused by the inherent decision-making logic of multimodal systems. Summary of the Invention

[0005] To address the technical problems existing in the background art, this invention proposes a bank business vault access control system based on multimodal identity verification, the specific solution of which is as follows: A banking vault access control system based on multimodal identity verification includes: A set of construction modules is used to pre-collect multiple initial biometric samples for at least one biometric modality of a target user, and to construct and store the historical effective feature set of the biometric modality; The sample acquisition module is used to acquire the current biometric sample of the target user that is currently awaiting authentication. The consensus statistics module is used to compare the current biometric sample with each historical sample in the historical valid feature set, and count the number of historical samples that meet the preset similarity conditions, which is used as the consensus quantity. The threshold determination module is used to obtain the historical authentication behavior sequence of the target user and dynamically determine the consensus threshold based on the regularity of the historical authentication behavior sequence, wherein the lower the regularity, the higher the consensus threshold. The decision-making module is used to determine whether the current authentication of the biometric modality is passed when the consensus count is not lower than the consensus threshold. The set update module is used to add the current biometric sample to the historical valid feature set only when the current authentication of the biometric modality is successful and the quality index of the current biometric sample meets the preset entry conditions; otherwise, it is prohibited from being added.

[0006] Furthermore, in the collection construction module, the biometric modality includes at least one of the following: Fingerprint modality, face modality, finger vein modality, palm vein modality, iris modality, and voiceprint modality.

[0007] Furthermore, the set construction module constructs and stores the historical valid feature set of the biometric modality, including: Based on the intra-class similarity distribution of the multiple initial biometric samples, a baseline consensus threshold for the biometric modality is determined, wherein the more concentrated the intra-class similarity distribution, the lower the baseline consensus threshold.

[0008] Furthermore, after constructing and storing the historical valid feature set, the method also includes: Cluster the historical samples in the historical valid feature set to obtain multiple feature prototypes, and establish a hierarchical feature index that includes a feature prototype layer and a cluster member layer.

[0009] Furthermore, in the consensus statistics module, the current biometric sample is compared with each historical sample in the historical valid feature set to determine the number of historical samples that meet the preset similarity conditions, including: Calculate the feature distance between the current biometric sample and each historical sample in the historical valid feature set; Historical samples whose feature distance is less than a preset distance threshold are determined to meet the preset similarity condition, and the number of historical samples that meet the preset similarity condition is counted as the consensus quantity.

[0010] Furthermore, in the threshold determination module, the consensus threshold is dynamically determined based on the regularity of the historical authentication behavior sequence, including: Symbolize each authentication result in the historical authentication behavior sequence to form a symbol sequence; Calculate the complexity of the symbol sequence, and use the complexity to characterize the degree of regularity; Based on the difference between the complexity and the preset baseline complexity, the consensus threshold is adjusted to obtain the consensus threshold. The higher the complexity, the greater the upward adjustment of the consensus threshold relative to the baseline consensus threshold.

[0011] Furthermore, in the set update module, the quality index is a composite quality score, which is obtained by weighted fusion of at least two of the following: image quality index, feature extraction confidence, and distance anomaly between the current biometric sample and the center vector of the historical effective feature set.

[0012] Furthermore, when adding the current biometric sample to the historical valid feature set, if the number of samples in the historical valid feature set has reached the preset maximum capacity, the historical sample with the earliest addition time will be eliminated according to the least recently used strategy.

[0013] Furthermore, if the current biometric authentication fails or the quality index of the current biometric sample does not meet the preset entry conditions, but the target user opens the access control through authentication of other biometric modalities, the current biometric sample is still prohibited from being added to the historical valid feature set.

[0014] Compared with the prior art, the present invention can achieve at least the following beneficial effects: 1. This invention replaces the traditional single centralized template by constructing a historical valid feature set, transforming identity authentication into a consensus judgment between the current sample and multiple historical samples, and establishing strict dual thresholds for entry into the pool. This fundamentally solves the template contamination problem driven by cross-modal confirmation in existing technologies. Specifically, a sample is only allowed to be added to the historical valid feature set when it has passed authentication for a certain biometric modality and meets the sample quality standards. Even if the high confidence of other modalities leads to the final opening of the access control, low-quality or unauthenticated samples are strictly prohibited from entering the pool. This mechanism cuts off the link for low-quality samples to enter the template library through the backdoor of multimodal fusion decision-making, eliminating the cumulative risk of chronic template drift from the source, ensuring that the historical valid feature set remains pure in the long term. This effectively avoids the problem of a sudden increase in false rejection rate and potential increase in security risks caused by template contamination during long-term system operation, significantly improving the long-term operational stability and security of the access control system.

[0015] 2. This invention replaces the traditional single centralized template by establishing a historical valid feature set and sets a strict dual-threshold entry mechanism. Current biometric samples must simultaneously meet two conditions: passing the authentication of this modality and meeting their own quality standards, in order to be eligible to enter the historical valid feature set. When a scenario occurs where the fingerprint quality is extremely low due to dry and cracked skin, but the face recognition score is high and the access control is opened, the low-quality fingerprint sample will be strictly prohibited from being added to the historical valid feature set because it cannot pass the consensus authentication of this modality or does not meet the quality threshold. This mechanism completely cuts off the pollution path of low-quality samples being mixed into the template library due to cross-modal strong confirmation, so that the historical valid feature set always remains pure. Because low-quality samples are systematically blocked and cannot enter the historical valid feature set to participate in subsequent consensus comparisons, the set always retains the user's recent high-quality feature samples. Even if the system runs for several weeks or months, the feature distribution of the historical valid feature set in the background will not drift due to the continuous absorption of low-quality samples, and will always remain consistent with the user's real biometric status. This fundamentally eliminates the risk of legitimate users being mistakenly rejected during normal authentication due to template drift, and the risk of a sudden increase in the high false rejection rate. By acquiring users' historical authentication behavior sequences, the consensus threshold is dynamically determined based on the degree of regularity, enabling the system to detect abnormal fluctuations in authentication behavior. When consecutive authentication failures or drastic state fluctuations occur, the degree of regularity decreases, and the consensus threshold automatically increases, requiring the current sample to obtain the approval of more historical samples before it can pass. This mechanism only makes moderate adjustments when normal users experience occasional state fluctuations, but it can significantly improve the defense level when subjected to probing attacks using forged features. It achieves intelligent and dynamic adjustment of security strategies, which is an additional technical effect that existing fixed-threshold methods do not possess. Instead of relying on a single, constantly modified central template for authentication comparison, the authentication process is transformed into a consensus vote between the current sample and multiple independent historical samples in the set of valid historical features. Only when a sufficient number of historical samples jointly approve the current sample can the authenticator pass. This distributed consensus mechanism avoids the risk that the authentication system will fail directly due to the contamination of a single template. The collective judgment of multiple independent samples significantly improves the reliability and stability of the authentication results. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic diagram of the overall system principle of the present invention.

[0017] Figure 2 This is a flowchart of the collection construction module of the present invention.

[0018] Figure 3 This is a flowchart of the consensus statistics module of the present invention.

[0019] Figure 4 This is a flowchart of the collection update module of the present invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] Please refer to Figures 1 to 4 As shown, the present invention provides a banking business database access control system based on multimodal identity verification, including a set construction module, a sample acquisition module, a consensus statistics module, a threshold determination module, a decision passing module, and a set update module; Please see Figure 2 The set includes a construction module for at least one biometric modality of a target user, which pre-collects multiple initial biometric samples, constructs and stores the historical valid feature set of the biometric modality.

[0022] It should be noted that the historical valid feature set refers to a fixed-capacity set of feature samples maintained separately for each biometric modality of each user. This set stores high-quality biometric samples recently collected during the authentication process for that user and that modality. This set is initialized during user registration by collecting initial biometric samples multiple times and is dynamically updated according to preset rules after each successful authentication. Unlike traditional single centralized templates, the historical valid feature set retains multiple independent original feature samples, rather than a single template obtained by averaging or fusing these samples. This allows subsequent authentication decisions to be based on the consensus between the current sample and multiple historical samples, thus avoiding the risk of the overall template being contaminated by a single low-quality sample.

[0023] In an optional embodiment, the biometric modality in the collection building module includes at least one of the following: Fingerprint modality, face modality, finger vein modality, palm vein modality, iris modality, and voiceprint modality.

[0024] In an optional embodiment, the set construction module constructs and stores a historical valid feature set of the biometric modality, including: Based on the intra-class similarity distribution of the multiple initial biometric samples, a baseline consensus threshold for the biometric modality is determined, wherein the more concentrated the intra-class similarity distribution, the lower the baseline consensus threshold.

[0025] It should be noted that the baseline consensus threshold refers to the minimum number of consensuses required to determine whether the current biometric sample has received sufficient "recognition" from a sufficient number of historical samples. Its initial value is determined during the registration phase based on the stability of the user's modality feature. The intra-class similarity distribution is used to measure the degree of feature consistency among multiple initial samples of the same modality for that user. Specifically, the intra-class similarity distribution can be characterized by calculating the feature distances between all pairs of initial samples and statistically analyzing the mean and standard deviation of these distances. The more concentrated the intra-class similarity distribution, the more stable and less varied the user's modality feature is across different collections. In this case, a lower baseline consensus threshold can be set, making authentication easier under normal conditions. Conversely, if the differences between initial samples are large, it indicates that the modality feature itself has a certain degree of volatility, requiring a higher baseline consensus threshold to ensure the security of authentication.

[0026] For example, in a bank vault fingerprint access control registration scenario, 10 initial fingerprint samples were collected from employee Zhang San. The calculated mean of the feature distance between samples was 0.12, and the standard deviation was 0.03, indicating a relatively concentrated distribution. Based on this, the baseline consensus threshold can be set to 15. However, for another employee, Li Si, with a pre-existing hand injury, 10 samples were also collected. The mean of the feature distance between samples was 0.25, and the standard deviation was 0.08, indicating a more dispersed distribution. Therefore, the baseline consensus threshold can be set to 20. This method of setting thresholds based on the differences in the stability of individual characteristics balances the security and usability of authentication.

[0027] In an optional embodiment, after constructing and storing the historical valid feature set, the method further includes: Cluster the historical samples in the historical valid feature set to obtain multiple feature prototypes, and establish a hierarchical feature index that includes a feature prototype layer and a cluster member layer.

[0028] It should be noted that clustering historical samples in the historical valid feature set and establishing a hierarchical feature index is to solve the efficiency problem of directly comparing them one by one with all historical samples in the set. The feature prototype is the representative vector of each group obtained by grouping historical samples using a clustering algorithm. Usually, the mean of all sample vectors in each group is taken as the feature prototype of that group. The hierarchical feature index consists of two layers: a feature prototype layer and a cluster member layer. The feature prototype layer stores the center vector of each cluster, and the cluster member layer stores the specific historical samples belonging to each prototype. When performing subsequent similarity comparisons, the currently collected samples can be quickly compared with a small number of prototypes in the feature prototype layer to identify the most similar one or a few prototypes. Then, a precise comparison is performed only with the historical samples in the clusters corresponding to these prototypes. Since the number of feature prototypes is much smaller than the total number of historical samples, this hierarchical retrieval method can reduce the number of comparisons from the full set to the scale of a single or a few clusters, thereby significantly reducing computational overhead and meeting the millisecond-level response requirements of access control systems for authentication.

[0029] The sample acquisition module is used to acquire the current biometric sample of the target user that is currently awaiting authentication.

[0030] It should be noted that the current biometric sample refers to the biometric vector collected and extracted in real time by the biometric acquisition device installed at the bank's business vault access control point during this access authentication. Unlike the historical valid feature sets of previous authentication samples, the current biometric sample reflects the user's biometric status at this moment. It may differ from historically stored samples due to factors such as changes in the user's physiological state, differences in the collection environment, or fluctuations in device operating conditions. This difference is the core object that needs to be judged through the consensus mechanism.

[0031] Please see Figure 3 The consensus statistics module is used to compare the current biometric sample with each historical sample in the historical valid feature set, and count the number of historical samples that meet the preset similarity conditions, which is used as the consensus quantity.

[0032] In an optional embodiment, the consensus statistics module compares the current biometric sample with each historical sample in the historical valid feature set to count the number of historical samples that meet a preset similarity condition, including: Calculate the feature distance between the current biometric sample and each historical sample in the historical valid feature set; Historical samples whose feature distance is less than a preset distance threshold are determined to meet the preset similarity condition, and the number of historical samples that meet the preset similarity condition is counted as the consensus quantity.

[0033] It should be noted that the consensus count is a quantitative statistical result of the similarity between the current biometric sample and historical samples in the historical valid feature set. Specifically, it means the number of historical samples in the historical valid feature set whose similarity to the current biometric sample reaches a preset standard. The higher the consensus count, the more consistent the current sample is with the samples retained from the user's multiple successful authentications in the past, and the greater the likelihood that the sample belongs to the user. Conversely, the lower the consensus count, the greater the difference between the current sample and the historically retained samples, and the higher the likelihood of an abnormal state or operation not performed by the user.

[0034] It should be noted that the feature distance refers to the metric difference between two biometric vectors in the feature space, used to quantify the similarity between two samples. The smaller the feature distance, the closer the two samples are in the feature space, and the higher the similarity. The method for calculating the feature distance can be selected according to the type of biometric modality. For example, for fingerprint modality, Euclidean distance can be used to calculate the square root of the sum of the squares of the differences in each dimension of the two feature vectors; for face modality, cosine distance can be used to calculate the cosine value of the angle between the two feature vectors as a similarity measure.

[0035] It should be noted that the preset distance threshold is the boundary value for determining whether a single comparison constitutes "valid consensus". If the feature distance between the current biometric sample and a certain historical sample is less than the threshold, it is considered that the historical sample has cast a "vote of approval" for the current sample, and the historical sample is counted as meeting the preset similarity condition; otherwise, it is not counted in the consensus count. The preset distance threshold can be set according to the intra-class feature distance distribution of the user's modality in the initial samples during the registration phase. For example, the threshold can be set as the mean of the feature distance of the initial samples plus two standard deviations, so that samples within the normal fluctuation range can be recognized by historical samples, while samples that deviate abnormally cannot obtain sufficient consensus.

[0036] For example, assuming the mean feature distance between each pair of the ten initial samples of a user's fingerprint modality is 0.12 and the standard deviation is 0.03, the preset distance threshold can be set to 0.18. In one authentication, the current sample is compared one by one with 50 historical samples in the historical valid feature set. If the feature distance between the current sample and 22 historical samples is less than 0.18, then the consensus count is 22.

[0037] It should be noted that the "historical samples" involved in the above comparison process can, in actual execution, be either all historical samples in the set of valid historical features, or, after using hierarchical feature indexing, only a portion of historical samples selected from the clusters corresponding to the feature prototypes most similar to the current sample. Regardless of the retrieval method used, the statistical rule for the number of consensus samples remains consistent, that is, the final result is the number of historical samples that actually participate in the precise comparison and meet the preset similarity conditions.

[0038] The threshold determination module is used to obtain the historical authentication behavior sequence of the target user and dynamically determine the consensus threshold based on the regularity of the historical authentication behavior sequence, wherein the lower the regularity, the higher the consensus threshold.

[0039] It should be noted that the historical authentication behavior sequence refers to the time-series record of authentication results generated by the target user for the current biometric modality within a recent period or in several recent authentication processes. Each authentication result in this sequence can be classified and labeled according to preset rules; for example, high-quality passes are labeled as category 1 results, low-quality passes as category 2 results, and authentication failures as category 3 results. The order of the sequence reflects the evolution trend of authentication results over time and serves as the foundational data for subsequent analysis of the regularity of user authentication behavior.

[0040] It should be noted that the degree of regularity is used to measure whether the user's recent authentication behavior is stable and predictable. A lower degree of regularity indicates greater volatility or unpredictability in the user's recent authentication result sequence, which may manifest as consecutive authentication failures, alternating high and low quality results, or an irregular random distribution of authentication results. These abnormal patterns may be caused by various reasons, such as temporary changes in the user's physiological characteristics, fluctuations in the status of the data collection device, or may even suggest that an imposter is attempting a probe attack using forged features. When the system detects such abnormal fluctuations, it automatically raises the consensus threshold, requiring the current sample to obtain more approval from historical samples before authentication can be passed, thereby intelligently enhancing the security level without affecting normal user experience.

[0041] In an optional embodiment, the threshold determination module dynamically determines the consensus threshold based on the regularity of the historical authentication behavior sequence, including: Symbolize each authentication result in the historical authentication behavior sequence to form a symbol sequence; It should be noted that symbolizing the authentication results is the process of converting the original authentication records into a symbol sequence that facilitates computational complexity analysis. Specifically, based on the decision result and sample quality of each authentication, the results can be mapped to a preset symbol category. For example, "authentication passed and sample quality is higher than the entry threshold" can be mapped to symbol A, "authentication passed but sample quality is lower than the entry threshold" can be mapped to symbol B, and "authentication failed" can be mapped to symbol C. Thus, an original sequence containing several authentication results can be converted into a symbol sequence composed of a finite number of symbol categories.

[0042] Calculate the complexity of the symbol sequence, and use the complexity to characterize the degree of regularity; It should be noted that the complexity mentioned is a quantitative indicator measuring the degree of regularity in a sequence of symbols. The more regular and predictable the arrangement of symbols in the sequence, the lower its complexity; conversely, the more frequent the changes in symbols and the more difficult the pattern to predict, the higher its complexity. The complexity can be calculated using known sequence complexity measurement methods in this field. For example, for the symbol sequence "A, A, A, A, A", its change pattern is simple, resulting in low complexity; for the symbol sequence "A, A, C, C, C", new symbols appear and the pattern changes, increasing complexity; and for the symbol sequence "A, C, A, C, A", its symbols alternate frequently without a clear pattern, resulting in even higher complexity.

[0043] Based on the difference between the complexity and the preset baseline complexity, the consensus threshold is adjusted to obtain the consensus threshold. The higher the complexity, the greater the upward adjustment of the consensus threshold relative to the baseline consensus threshold.

[0044] It should be noted that the preset baseline complexity is a reference baseline used to measure the complexity of the authentication sequence under normal and stable user usage conditions. It can be calculated and set based on the authentication sequences collected during a stable operating period in the early stages of user usage. The baseline consensus threshold is the basic consensus threshold value determined during the registration phase based on the intra-class similarity distribution of the initial samples. When the complexity calculated this time is higher than the preset baseline complexity, it indicates that the unpredictability of the current authentication behavior exceeds the normal level. Based on this, the system adds a certain safety margin above the baseline consensus threshold to obtain the dynamic consensus threshold, and the greater the complexity exceeds the baseline value, the larger the increase. When the complexity is equal to or lower than the preset baseline complexity, it indicates that the current authentication behavior is in a normal and stable state, and the baseline consensus threshold can be directly used as the dynamic consensus threshold.

[0045] It should be noted that the mechanism for dynamically adjusting the consensus threshold based on complexity differences can be illustrated with a specific scenario. Suppose an employee's baseline consensus threshold for this modality is 15, and the preset baseline complexity is 0.8. During normal use, the symbol sequence of their most recent ten authentication results is primarily "A, A, A, A, B, A, A, A, A, A," resulting in a complexity of 0.6, lower than the preset baseline complexity. This indicates that the current authentication behavior is highly regular, and the dynamic consensus threshold can be maintained at 15 or even appropriately lowered. One day, the employee experiences a temporary skin abnormality on their hand, causing the authentication sequence to change to "A, B, A, B, B, A, B, A, B," increasing the complexity to 1.2. After detecting this abnormal fluctuation, the system, according to the adjustment rules, increases the adjustment amount based on the baseline threshold of 15, resulting in a dynamic consensus threshold of 17, appropriately raising the authentication threshold. In another scenario, if someone attempts a probing attack using forged features, resulting in multiple authentication failures in the sequence (e.g., "A, A, C, C, C, F, C, C"), the complexity increases significantly to 1.6. The system then raises the dynamic consensus threshold to 19, making it more difficult for attackers to obtain a sufficient number of historical sample consensuses, thus effectively resisting attacks. Through this mechanism, the system achieves automatic and dynamic adjustment of security policies, remaining tolerant of minor fluctuations in the state of normal users while automatically raising the defense level for persistent abnormal or suspected attack behavior.

[0046] The decision-making module is used to determine whether the current authentication of the biometric modality is passed when the number of consensuses is not less than the consensus threshold.

[0047] It's important to note that the authentication decision is a single-modal authentication result for the current biometric modality, not a final access control decision based on multimodal fusion. This scheme determines whether the authentication of the current modality itself passes based solely on the degree of consensus between the current sample and the historical valid feature set. In a multimodal access control system, whether the access control is ultimately opened also depends on the authentication results of other modalities and the fusion decision strategy. Therefore, even if the current authentication of a particular modality fails, the access control may still be opened due to high-confidence authentication from other modalities. However, in this case, the current sample of that modality will be prohibited from being added to the historical valid feature set in subsequent steps to prevent low-quality samples from contaminating the set.

[0048] Please see Figure 4 The set update module is used to add the current biometric sample to the historical valid feature set only when the current authentication of the biometric modality is successful and the quality index of the current biometric sample meets the preset entry conditions; otherwise, it is prohibited from being added.

[0049] It's important to note that a dual-threshold mechanism is established for updating the historical valid feature set. The first threshold is the authentication threshold, meaning the biometric modality must achieve a sufficient number of consensuses to pass authentication. Only after authentication is the sample qualified to enter the set. The second threshold is the quality compliance threshold, meaning the current sample must meet preset inclusion criteria. Even if authentication is successful, if the sample quality is substandard, it is prohibited from entering the set. The core purpose of this dual-threshold mechanism is to ensure that only samples that are widely recognized by historical samples and possess excellent quality are eligible to become new historical samples for subsequent consensus comparisons. This effectively blocks low-quality samples from entering the set at the source, completely eliminating the risk of chronic template contamination caused by cross-modal confirmation.

[0050] In an optional embodiment, in the set update module, the quality metric is a composite quality score, which is obtained by weighted fusion of at least two of the following: image quality metric, feature extraction confidence, and distance anomaly between the current biometric sample and the center vector of the historical valid feature set.

[0051] It should be noted that the composite quality score is a quantitative indicator that comprehensively measures the quality of current biometric sample collection, obtained by weighted fusion of multiple quality assessment factors. Single-dimensional quality assessments often fail to fully reflect the true usability of a sample; the composite scoring method effectively compensates for the limitations of a single indicator. Specifically, image quality indicators are used to assess the clarity, contrast, and completeness of the original acquired image itself; for example, fingerprint images can be scored using the NFIQ standard. Feature extraction confidence assesses the reliability of the process of extracting feature vectors from the original image; a higher confidence indicates more reliable extracted features. Distance anomaly measures the degree of deviation between the current sample and the center vector of the historical effective feature set; a larger deviation indicates a more significant difference between the current sample and the user's historical normal state. These three factors are evaluated separately from three different dimensions: the quality of the collection source, the reliability of the extraction process, and consistency with historical states. They are then weighted and fused into a single comprehensive score according to preset weights, which can more accurately determine whether a sample is suitable for inclusion in the historical effective feature set.

[0052] It should be noted that the calculation of the composite quality score described above can be explained in conjunction with specific scenarios. Suppose that in a fingerprint authentication process, the acquired fingerprint image is clear and complete, with an NFIQ score of 1 (the highest level), a feature extraction confidence level of 0.98, and the normalized distance between this sample and the center vector of the historical valid feature set is 0.1, indicating extremely low anomaly. If the weights of the three factors are all set to 1 / 3, the composite quality score is 1×1 / 3 + 0.98×1 / 3 + 1×1 / 3, which is approximately equal to 0.99. This is far higher than the preset entry threshold (e.g., 0.7), and this sample can be allowed to be added to the historical valid feature set. Conversely, in another authentication, the user's hands were dry due to the use of disinfectant, resulting in a blurry fingerprint image. The NFIQ score was 4 (poor), the feature extraction confidence dropped to 0.6, the anomaly of the distance to the center vector increased to 0.8, and the composite quality score dropped to approximately 0.6, which is below the entry threshold of 0.7. Even if the access control is opened due to successful authentication in other modalities, this sample will be prohibited from being added to the historical valid feature set.

[0053] It should be noted that the weights of each factor are real numbers greater than or equal to 0 and less than or equal to 1, and the sum of the three is equal to 1. Specific values ​​can be configured differently based on the actual application scenario. Generally speaking, the distance anomaly directly measures the degree of deviation between the current sample and the user's historical normal characteristic state, and is a key indicator for determining whether a sample is suitable for inclusion in the historical effective feature set. It is usually recommended to assign a relatively high weight, with a value range of 0.3 to 0.6. The image quality index is a basic guarantee for the usability of the sample, with a value range of 0.2 to 0.5. The feature extraction confidence reflects the reliability of the transformation process from the original image to the feature vector, with a value range of 0.2 to 0.4. The basis for weight adjustment mainly includes: the higher the security level requirement, the higher the weight of the distance anomaly; the worse the stability of the biometric modality acquisition, the higher the weight of the image quality index; the worse the stability of individual user characteristics, the higher the weight of the distance anomaly; in the early stages of system operation, the weight of the distance anomaly can be appropriately reduced, and gradually increased after the set tends to stabilize.

[0054] In an optional embodiment, when adding the current biometric sample to the historical valid feature set, if the number of samples in the historical valid feature set has reached a preset maximum capacity, the historical sample that was added earliest is eliminated according to the least recently used strategy.

[0055] It should be noted that the preset maximum capacity is a storage limit set for the historical valid feature set. The purpose of setting the maximum capacity is twofold: firstly, to control storage overhead and comparison computation, preventing the set from growing indefinitely and impacting system performance; and secondly, to ensure that the set retains only recent, high-quality feature samples by eliminating outdated samples, allowing the set to naturally follow the slow changes in the user's biometrics. The Least Recently Used (LRU) strategy is triggered when the feature set is full. Specifically, it identifies the oldest historical sample from the current set, removes it from the set, and then adds the currently eligible sample to the set. The inclusion time, rather than comparison frequency, is used as the elimination criterion because the main value of historical samples lies in providing a reference for feature distribution over time; retaining more recent samples better reflects the user's recent true feature state.

[0056] In an optional embodiment, if the current authentication of the biometric modality fails or the quality index of the current biometric sample does not meet the preset entry conditions, but the target user opens the access control through authentication of other biometric modalities, the current biometric sample is still prohibited from being added to the historical valid feature set.

[0057] It's important to note that the above set update rules include a crucial boundary condition for the prohibition rule in cross-modal verification scenarios. In a multimodal access control system, the following situation may occur: a biometric modality (such as fingerprint) fails to obtain sufficient consensus due to poor sample quality or feature distortion, resulting in authentication failure. However, the system integrates high-confidence authentication results from other modalities such as face and iris recognition, ultimately determining the user's identity as legitimate and opening the access control. In this scenario, even though the access control is open, the current sample of this fingerprint modality is strictly prohibited from being added to its historical valid feature set. This rule is a key design element in blocking cross-modal template contamination. Its technical logic is as follows: if a sample fails to pass the consensus authentication of this modality, it indicates a significant deviation from the historical samples retained by this modality. Including it in the set would introduce low-quality or abnormal samples, which would gradually contaminate the entire set over time, ultimately leading to authentication performance degradation. By explicitly prohibiting such samples from entering the pool, the system ensures the availability of multimodal access control while safeguarding the purity of each modality's feature set through a mechanism that maintains each modality independently.

[0058] In summary, this invention application replaces the traditional single-center template with a distributed consensus mechanism based on historical valid feature sets, and establishes a dual entry threshold of passing this modality authentication and meeting its own quality standards. This fundamentally cuts off the link between cross-modal confirmation driving low-quality samples to pollute the template library, thereby solving the technical problem of a sudden increase in high false rejection rate and potential security risk caused by long-term cumulative drift of templates.

[0059] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0060] In the embodiments provided by this invention, it should be understood that the disclosed system or method can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0061] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0062] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0063] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0064] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A bank vault access control system based on multimodal identity verification, characterized in that, include: A set of construction modules is used to pre-collect multiple initial biometric samples for at least one biometric modality of a target user, and to construct and store the historical effective feature set of the biometric modality; The sample acquisition module is used to acquire the current biometric sample of the target user that is currently awaiting authentication. The consensus statistics module is used to compare the current biometric sample with each historical sample in the historical valid feature set, and count the number of historical samples that meet the preset similarity conditions, which is used as the consensus quantity. The threshold determination module is used to obtain the historical authentication behavior sequence of the target user and dynamically determine the consensus threshold based on the regularity of the historical authentication behavior sequence, wherein the lower the regularity, the higher the consensus threshold. The decision-making module is used to determine whether the current authentication of the biometric modality is passed when the consensus count is not lower than the consensus threshold. The set update module is used to add the current biometric sample to the historical valid feature set only when the current authentication of the biometric modality is successful and the quality index of the current biometric sample meets the preset entry conditions; otherwise, it is prohibited from being added.

2. The banking vault access control system based on multimodal identity verification as described in claim 1, characterized in that: In the collection construction module, the biometric modalities include at least one of the following: Fingerprint modality, face modality, finger vein modality, palm vein modality, iris modality, and voiceprint modality.

3. The banking vault access control system based on multimodal identity verification as described in claim 1, characterized in that: The set construction module constructs and stores the historical valid feature set of the biometric modality, including: Based on the intra-class similarity distribution of the multiple initial biometric samples, a baseline consensus threshold for the biometric modality is determined, wherein the more concentrated the intra-class similarity distribution, the lower the baseline consensus threshold.

4. The banking vault access control system based on multimodal identity verification as described in claim 1, characterized in that: After constructing and storing the historical valid feature set, the following is also included: Cluster the historical samples in the historical valid feature set to obtain multiple feature prototypes, and establish a hierarchical feature index that includes a feature prototype layer and a cluster member layer.

5. The banking vault access control system based on multimodal identity verification as described in claim 1, characterized in that: In the consensus statistics module, the current biometric sample is compared with each historical sample in the historical valid feature set to count the number of historical samples that meet the preset similarity conditions, including: Calculate the feature distance between the current biometric sample and each historical sample in the historical valid feature set; Historical samples whose feature distance is less than a preset distance threshold are determined to meet the preset similarity condition, and the number of historical samples that meet the preset similarity condition is counted as the consensus quantity.

6. The banking vault access control system based on multimodal identity verification as described in claim 3, characterized in that: In the threshold determination module, the consensus threshold is dynamically determined based on the regularity of the historical authentication behavior sequence, including: Symbolize each authentication result in the historical authentication behavior sequence to form a symbol sequence; Calculate the complexity of the symbol sequence, and use the complexity to characterize the degree of regularity; Based on the difference between the complexity and the preset baseline complexity, the consensus threshold is adjusted to obtain the consensus threshold. The higher the complexity, the greater the upward adjustment of the consensus threshold relative to the baseline consensus threshold.

7. The banking vault access control system based on multimodal identity verification as described in claim 1, characterized in that: In the set update module, the quality index is a composite quality score, which is obtained by weighted fusion of at least two of the following: image quality index, feature extraction confidence, and distance anomaly between the current biometric sample and the center vector of the historical effective feature set.

8. The banking vault access control system based on multimodal identity verification as described in claim 1 or 4, characterized in that: When adding the current biometric sample to the historical valid feature set, if the number of samples in the historical valid feature set has reached the preset maximum capacity, the historical sample with the earliest addition time will be eliminated according to the least recently used strategy.

9. The banking vault access control system based on multimodal identity verification as described in claim 1, characterized in that: If the current biometric authentication fails or the quality index of the current biometric sample does not meet the preset entry conditions, but the access control is opened by the target user through authentication of other biometric modalities, the current biometric sample is still prohibited from being added to the historical valid feature set.

Images