A multimodal biometric feature fusion intelligent door lock safety control method

Abstract

The present application belongs to the technical field of Internet of Things security, and specifically relates to a multi-modal biological feature fusion intelligent door lock security control method, which collects at least two types of original biological features in parallel through a multi-modal biological feature collection unit and performs a live anti-fake check, performs a space-time two-dimensional adaptive weight calculation based on real-time environmental parameters and user historical identification behavior data, applies controllable micro-Gauss desensitization disturbance processing to a feature vector, adjusts an identity determination threshold in real time according to risk control factors such as an unlocking period, position, historical abnormal frequency, and performs hierarchical check in combination with a global fusion confidence and a single-modal live body credibility, and performs security control actions such as unlocking, review or rejection alarm according to a check result. By adopting the technical scheme, the present application can effectively improve the anti-interference performance of an identification system, realize biological feature privacy protection, balance safety and convenience requirements, and has a fine risk disposal capability.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things security technology, and in particular relates to a smart door lock security control method that integrates multimodal biometric features. Background Technology

[0002] With the rapid development of IoT and AI technologies, the smart home industry has entered a stage of large-scale application. As the physical entry point for a smart home system, the security capabilities of smart locks directly impact user experience. In recent years, smart lock products equipped with biometric recognition technologies such as fingerprint recognition, facial recognition, and finger vein recognition have rapidly gained popularity, gradually replacing traditional mechanical keys and password input authentication methods, bringing users a seamless and convenient entry experience.

[0003] However, single biometric modalities are revealing increasingly significant technical bottlenecks in practical applications. From the perspective of the acquisition environment, fingerprint recognition is easily affected by factors such as finger humidity, wear and tear, and the pressure applied during acquisition, resulting in a significant decrease in recognition stability in dry or humid environments. Facial recognition experiences large fluctuations in acquisition quality under complex lighting conditions such as backlighting, strong light, and low light. Finger vein recognition is highly sensitive to the acquisition location and blood flow rate. From a security perspective, single biometric modalities face a severe threat from forgery attacks. Attack methods such as photo hijacking, 3D-printed masks, and silicone simulated finger membranes continue to evolve, and the limitations of traditional single-modal anti-counterfeiting capabilities constitute a security risk that cannot be ignored.

[0004] As multimodal fusion solutions are widely used in actual products and their deployment scale continues to expand, deep-seated defects in their technical implementation are gradually becoming apparent.

[0005] First, existing multimodal fusion schemes generally use a fixed weight mechanism for feature fusion. This inherent characteristic means that when the acquisition quality of a certain biometric modality is significantly reduced due to environmental interference, the system still includes the feature data of the low-quality modality in the fusion operation according to the preset fixed weight, inevitably introducing noise feature components, which in turn causes false rejection or false approval events, resulting in poor robustness and scene adaptability of the recognition system.

[0006] Secondly, existing technologies suffer from a significant lack of privacy protection mechanisms throughout the entire biometric data processing chain. Specifically, traditional solutions directly input the raw biometric vectors into the fusion algorithm for plaintext computation. Sensitive biometric data such as user fingerprints, facial features, and finger vein patterns are transferred and stored in plaintext form during the algorithm processing chain, lacking effective protection from edge-side desensitization mechanisms.

[0007] Furthermore, existing identity verification mechanisms suffer from significant static and fixed limitations in their judgment strategies. Most solutions employ fixed judgment thresholds for binary identity authentication. These threshold parameters are calibrated at the factory and fixed in the system configuration, and do not adaptively adjust to changes in the risk level of the unlocking scenario during use.

[0008] Finally, existing technologies suffer from limitations in scenario risk assessment capabilities due to their singular focus and coarse quantification. Some solutions only consider the success or failure of the matching result during the identity authentication decision-making process, failing to incorporate multi-dimensional spatiotemporal information such as unlocking time, unlocking location, and historical abnormal unlocking frequency into a comprehensive assessment. This makes it impossible to conduct a refined and quantitative assessment of the security risk level of the current unlocking scenario, significantly limiting the targeting and effectiveness of risk control strategies.

[0009] Therefore, this invention aims to address the aforementioned technical contradictions and challenges by proposing a smart door lock security control method based on multimodal biometric fusion. Summary of the Invention

[0010] The purpose of this invention is to provide a smart door lock security control method based on multimodal biometric fusion, in order to solve the technical problems existing in the prior art.

[0011] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A smart door lock security control method based on multimodal biometric fusion, applied to a smart door lock system equipped with a multimodal biometric acquisition unit, a liveness detection unit, an environmental sensing unit, a main control unit, and a door lock actuator, includes the following steps: S1. At least two different types of raw biometric data of the current user are collected in parallel by the multimodal biometric acquisition unit, and the real-time environmental parameters and device operating status parameters of the smart door lock are collected simultaneously by the environmental perception unit. S2. Perform liveness detection verification on each type of original biometric data using the liveness detection unit, and calculate the liveness confidence score for each type of biometric data. ,filter Biometric data exceeding the preset liveness confidence threshold proceeds to step S3; S3. Based on the preset feature quality evaluation rules, the selected biometric data are scored. Biometric data with a quality score lower than the preset quality score threshold are identified as invalid modal data and removed. Valid modal biometric data with a quality score that meets the standard and has completed the liveness verification are retained. S4. Using the real-time environmental parameters as spatial dimension factors and the user's historical recognition behavior data and the long-term recognition stability of each modality as time dimension factors, perform spatiotemporal dual-dimensional adaptive weight calculation on the biometric data of each effective modality to determine the dynamic fusion weight corresponding to each effective modality. S5. Apply a controllable micro Gaussian desensitization perturbation to the feature vectors corresponding to each effective modality of biometric data to obtain desensitized biometric vectors. Perform fusion operation on all desensitized biometric vectors based on the dynamic fusion weights corresponding to each effective modality and output the global identity fusion confidence. S6. Based on the unlocking time period, unlocking location, and historical unlocking anomaly frequency parameters of the smart door lock, determine the security risk level of the current unlocking scenario in real time, and adaptively adjust the preset identity determination threshold according to the security risk level. S7. Combining the liveness confidence scores of each modality and the adaptively adjusted identity determination threshold, the global identity fusion confidence is subjected to hierarchical verification. Based on the hierarchical verification results, the door lock actuator is driven to perform corresponding security control actions. If the global identity fusion confidence score and the liveness confidence score of each modality both meet the high-level threshold conditions, the door lock actuator will be driven to perform the unlocking operation. If the global identity fusion confidence score or liveness confidence score is in the medium-level threshold range, a secondary biometric verification and collection process will be triggered. If the global identity fusion confidence level is not up to standard or there is a modal liveness verification failure, the unlocking operation will be refused, the abnormal unlocking data will be recorded, and the hardware alarm mechanism will be triggered.

[0012] The specific process of step S2 is as follows: S21: For the loaded first Based on the original biometric data, liveness feature parameters are extracted simultaneously from three anti-counterfeiting dimensions: texture, depth, and dynamics. The specific extraction rules are as follows: Extracting live texture feature matching degree Extract detailed textures, edge contours, and feature point distribution information of the current biometric features, compare them with the legal liveness texture benchmark library pre-stored locally on the device, and output a texture matching score. Characterizing the realism of static textures: Calculate the first Biological-like live texture feature matching degree : will the first The keypoint texture feature vectors of the original biological features are compared with the locally pre-stored standard live texture reference vectors of this modality using cosine similarity matching, as shown in the following formula: ; in: : No. The currently acquired texture feature vector for bio-like characteristics; : No. Pre-stored standard live texture reference vectors for bio-like features; Vector 2 norm operation, used to normalize characteristic magnitudes; Texture matching degree range is: The higher the value, the better the texture details match the real living features.

[0013] Extracting effective values ​​of live depth features Collect 3D spatial depth information of current biometric features, calculate the completeness, continuity, and spatial hierarchy of the depth information, and output an effective depth score. Distinguish between two-dimensional forged media and the three-dimensional features of a real living object; Calculate the first Effective value of bio-like in vivo depth features Based on 3D depth point cloud data of biological features, and combining depth integrity and depth value variance, the effective values ​​of living organism depth features are quantified. ; in: : No. Number of effective depth point clouds for biological-like features; : No. The total number of depth point clouds collected for bio-like features; : No. Variance of bio-like feature depth values; : Depth completeness weighting coefficient; Valid values ​​and ranges of live depth features It effectively distinguishes between two-dimensional planar forgery and three-dimensional live features.

[0014] Extracting dynamic change feature scores from living organisms Continuously acquire multiple frames of homologous biometric data to capture unique micro-dynamic features of living organisms, including skin micro-deformation, organ micro-movement, and changes in texture tension, and output dynamic effective scores. To resist static fake body attacks.

[0015] Calculate the first Biological characteristics of living dynamic changes feature score Based on homologous biological feature data acquired in multiple consecutive frames, the mean value of inter-frame feature differences characterizes the micro-dynamic characteristics of living organisms and suppresses static prosthetic features. The calculation formula is as follows: ; in: : No. The mean inter-frame feature difference of consecutive multi-frame images of bio-like characteristics characterizes the degree of micro-dynamic changes in living organisms. : Dynamic characteristic gain coefficient; : Score for dynamic changes in living organisms.

[0016] S22: Single-modal liveness confidence score fusion calculation: Based on the three types of liveness feature parameters extracted in S21, the score is calculated. Biomimetic characteristics liveness confidence score The calculation formula is as follows: ; in, : No. The liveness confidence score of bio-like characteristics; : No. Biomimetic live texture feature matching degree; : No. Effective values ​​of bio-like depth features in living organisms; : No. The score of the dynamic change characteristic of living organisms; Fixed weighting coefficients for texture, depth, and dynamic features.

[0017] S23: Full Modal Traversal Calculation and Data Integration: Repeat steps S21-S22 to complete all operations sequentially. The liveness scores of bio-feature modalities are calculated and integrated to obtain a complete set of modal liveness confidence scores. .

[0018] Preferably, the specific process of step S3 is as follows: S31: Load the biometric data of each modality that passed the S2 liveness verification one by one, and extract the first... Three basic quality assessment parameters for bio-like characteristics include feature completeness. Feature clarity Working condition adaptability ; Calculate feature completeness Calculated based on the ratio of the number of valid key feature points detected by current biometrics to the total number of standard feature points for this modality: ; in: : No. The number of valid key feature points detected in the currently acquired image of the modality; : No. The total number of standard complete key feature points preset for the class modality.

[0019] Calculate feature sharpness : ; in: : No. The gradient variance of the current feature image of the class modality; : No. The standard clear image gradient variance benchmark value for the modality preset.

[0020] Calculate the adaptability of the working conditions Normalized calculation based on multi-dimensional limit constraints of device signal-to-noise ratio, ambient light interference, and temperature and humidity interference: ; in: : Current number Real-time signal-to-noise ratio of the modal acquisition unit; The device has a preset standard signal-to-noise ratio threshold for data acquisition. Real-time ambient light intensity; : Preset optimal standard light intensity for data acquisition.

[0021] S32: A scoring mechanism combining minimum value bottleneck constraint and mean balance correction is adopted. This mechanism utilizes the bottleneck effect in the quality dimension to suppress non-conforming data and calculates the... Raw quality score of the modality : ; in: : No. Raw quality scores of bio-like modalities; : No. Modal feature completeness; : No. Modal feature clarity; S33: To avoid misjudgments due to qualified liveness detection but low feature reliability or inflated original quality, a liveness reliability score output from S2 is introduced. The original quality scores are corrected to obtain the final comprehensive quality score used for modality screening. This enables a dual-coupling evaluation of safety and quality attributes. ； in: : No. The final comprehensive quality score of bio-like modalities; : No. Modal raw quality score; : The output of S2 Modal liveness confidence score.

[0022] S34: Preset uniform modal quality threshold The final comprehensive quality score for each modality. Perform threshold determination and tiered filtering: like The modality is determined to be valid modality biometric data, retained, and included in the subsequent weight calculation process; like If the mode is determined to be a failed mode, it will be discarded directly.

[0023] The specific process of step S4 is as follows: S41: For each type of valid modality biometric data, extract the following: Spatial Dimension Environmental Adaptation Factor Based on the quantification of S1 real-time environmental parameters and equipment operating status parameters, it characterizes the impact of current illumination, signal-to-noise ratio, and environmental interference on the first... The better the adaptation of modal acquisition results, the less environmental interference and the better the acquisition conditions. The higher the score; Time-dimensional steady-state adaptation factor The stability of a modality is quantified by combining its long-term recognition stability with historical user recognition behavior data. It characterizes the long-term adaptability of the modality to the current user; the higher the historical recognition success rate, the stronger the feature stability. The higher the score.

[0024] S42: Modal composite quality score combined with the output of the preceding S3 By integrating spatial and temporal factors, the first... Comprehensive fit of effective modes : ; in: : No. Comprehensive fit of effective modes; : No. Effective modal composite quality score; : No. Modal space dimension environment adaptation factor; : No. Modal time dimension steady-state adaptation factor.

[0025] The use of multi-factor product coupling logic means that defects in any dimension will significantly reduce the modal synthesis fit.

[0026] S43: Perform global normalization on the overall fit of all effective modes to obtain the final dynamic fusion weights for each mode. ,: ; in: : No. Dynamic fusion weights for effective modalities; : No. Comprehensive fit of effective modes; : No. Comprehensive fit of effective modes, For valid modal traversal indexes; : Total number of valid modes after S3 screening.

[0027] Weight constraint property: The dynamic fusion weights of all effective modes satisfy... .

[0028] Spatial Dimension Environmental Adaptation Factor and time-dimensional steady-state adaptation factor The calculation process is as follows: Spatial Dimension Environmental Adaptation Factor Calculation: Based on the real-time environmental parameters collected by S1, the equipment operating status parameters, and the operating condition adaptability calculated by S3. An exponential penalty correction is applied based on the device's signal-to-noise ratio offset, calculated using the following formula: ; in: : No. Modal space dimension environment adaptation factor; : No. Modal condition adaptability characterizes the overall degree of adaptability to illumination and environmental interference. : No. Real-time signal-to-noise ratio of the modal acquisition unit; : The corresponding modal preset standard optimal acquisition signal-to-noise ratio benchmark value; Boundary constraints: the calculated value is forced to be 1 if it is greater than 1, and forced to be 0 if it is less than 0.

[0029] The exponential term represents the signal-to-noise ratio (SNR) deviation penalty. The greater the deviation of the SNR acquired by the equipment from the standard operating conditions, the stronger the penalty attenuation. The lower the value, the better. Only when the operating condition adaptability is high and the equipment signal-to-noise ratio meets the standard, the spatial adaptability factor approaches 1, which represents the current environment's true adaptability to the acquisition of this mode.

[0030] Time-dimensional steady-state adaptation factor calculation: Based on local historical recognition behavior data, it is calculated by coupling the dual indicators of modality historical recognition success rate and feature stability. ; in: : No. Modal time dimension steady-state adaptation factor; : The first in the preset statistical period Historical success rate of modality recognition; : The first in the preset statistical period Historical variance of modality feature scores; Taking the minimum of success rate and stability is an unweighted bottleneck constraint, which avoids the incorrect overestimation of modes that are "high success rate but large fluctuations" or "good stability but low success rate", and is consistent with the real steady-state adaptation logic of long-term use.

[0031] The specific process of step S5 is as follows: S51: Extraction of Original Feature Vectors for Effective Modalities: For each class of effective modal biometric data identified in S4, extract standardized original feature vectors with uniform dimensions, denoted as the first... The original feature vector of the effective mode is The original feature vector To create a composite feature vector that integrates the dimensions of live anti-counterfeiting and biological ontology features, the original feature vector... Specifically, it includes: texture feature components: corresponding to S2 texture feature matching degree derived quantization features; depth feature components: corresponding to S2 3D depth space effective features; dynamic micro-variation feature components: corresponding to S2 liveness dynamic change features, the specific expression of which is: ; S52: To avoid high accuracy loss or desensitization failure caused by fixed noise, an adaptive micro-Gaussian desensitization perturbation coefficient is designed based on the preceding modal quality parameters and environmental operating condition parameters. ; in: : No. Modal adaptive differential Gaussian desensitization perturbation coefficient; The system presets the basic disturbance reference coefficients; : No. The overall quality score for each modality; : No. Modal space dimension environment adaptation factor.

[0032] High-quality, high-performance reliable modes employ minute perturbations to retain identity feature information to the maximum extent; low-quality, poor-performance modes strengthen desensitization to avoid the risk of reverse leakage of dirty feature data, achieving a dynamic balance between security and accuracy.

[0033] S53: Based on an adaptive infinitesimal Gaussian desensitization perturbation coefficient, a controllable infinitesimal Gaussian noise perturbation is applied to the original feature vector to generate a desensitized biometric feature vector, avoiding the inverse reconstruction of the original biometric features. The formula is as follows: ; in: : No. Modal desensitization biometric vectors; : No. Modal-based original biological feature vectors; The mean is 0 and the standard deviation is 0. Gaussian infinitesimal noise.

[0034] S54: Call the dynamic fusion weights of each modality output by S4. We then perform a weighted fusion of all desensitized biometric feature vectors to obtain a global fused feature vector: ; in: : Global desensitization feature vector after multimodal fusion; : No. Effective modal dynamic fusion weights; : No. Modal desensitization of biological feature vectors.

[0035] S55: Perform cosine similarity matching between the global de-identified fusion feature vector and the user's locally stored standard fusion benchmark feature vector to quantify and generate the global identity fusion confidence score, thereby quantifying the degree of identity trustworthiness. ; in: Global identity fusion confidence level; : Standard fusion baseline feature vector pre-stored locally by the user; Vector 2-norm operation. Boundary constraints: The result is forced to be set to 1 if it is greater than 1, and forced to be set to 0 if it is less than 0.

[0036] The specific process of step S6 is as follows: S61: Calculate the three core risk control factors for the current unlocking scenario, including the risk factor during the unlocking period. Unlocking location risk factors and historical anomaly frequency risk factors The higher the risk, the greater the corresponding factor score; Unlocking Period Risk Factors Calculation: Based on the quantification of the deviation between the current unlock time and the user's regular authorization period, a time deviation normalization penalty is calculated using the following formula: ; in: : Risk factor for the unlocking period, value range [0,1]; : Current real-time unlock time (24-hour quantized time value); : Average of standard times for users to frequently authorize unlocking on a daily basis; : Preset the maximum legal time deviation threshold, used to normalize the time deviation amount.

[0037] The closer the unlock time is to the regular time period, The closer it gets to 0; abnormal unlocking times such as late at night and early morning, which deviate from the norm. Approaching 1.

[0038] Unlock location risk factors Calculation: Based on the spatial distance deviation quantification between the current unlocked location and commonly used trusted locations, the distance deviation index penalty is used for calculation, as follows: ; in: Unlock location risk factor, value range [0,1]; : The real-time spatial distance deviation between the current unlock location and the user's commonly used trusted unlock locations on the local network; : The maximum permissible deviation distance of the preset reliable position is used to normalize the spatial deviation.

[0039] The deviation is minimal when unlocking at a commonly used local location. The value approaches 0; however, the deviation is extremely large when unlocking in unfamiliar locations. Approaching 1.

[0040] Historical anomaly frequency risk factor Calculation: Based on the proportion of abnormal unlocking times to the total number of recognition times within the statistical period, frequency saturation normalization is used for calculation, as shown in the following formula: ; in: : Historical unlocking anomaly frequency risk factor, value range [0,1]; : The cumulative number of abnormal unlock attempts, including liveness detection, feature matching, and illegal attempts, within a preset statistical period; Total number of unlocks within the same statistical period.

[0041] Combining the proportion of anomalies with the saturation characteristics of cumulative anomalies, minor and sporadic anomalies slightly increase the risk, while frequent and repeated anomalies will quickly saturate and raise the risk score.

[0042] S62: Calculation of Scenario-Based Comprehensive Risk Coefficient: A logic-constraint cumulative unweighted risk quantification algorithm is adopted. The core principle is: low-risk factors are cumulatively superimposed, while high-risk factors are subject to mandatory capping constraints. This preserves the cumulative effect of multi-dimensional anomalies while ensuring that high risk in any dimension can significantly increase the overall risk. The formula is as follows: ; in: : The overall risk coefficient of the current unlocking scenario, with a value range of [0,1]. The larger the value, the higher the security risk of the unlocking scenario. : Risk factor for the unlocking period, value range [0,1]; Unlock location risk factor, value range [0,1]; : Historical unlocking frequency risk factor, value range [0,1].

[0043] Logical constraint description: It consists of a geometric mean cumulative term and a single-point high-risk correction term; the geometric mean term retains the cumulative superposition effect of three-dimensional risk, and the overall risk steadily increases when multiple dimensions are abnormal at the same time; max Single-point correction provides a safety net for high-risk scenarios. If any one dimension of the scenario is extremely risky, the overall risk coefficient can be significantly increased. This effectively solves the risk control loopholes where minor anomalies in multiple dimensions are averaged out and high-risk single points are masked, thus improving the reliability of door lock security.

[0044] Boundary constraints: Set the result to 0 when it is less than 0 and to 1 when it is greater than 1, so as to maintain consistency with the parameter value range system of the whole text.

[0045] S63: Safety Risk Level Classification Determination: Based on Comprehensive Risk Coefficient The system categorizes security risks into three levels, with thresholds fixed within the system to adapt to different risk control intensities in various scenarios. Specific grading rules are as follows: Low risk level: This corresponds to conventional trusted unlocking scenarios; Medium risk level: This corresponds to suspicious unlocking scenarios with minor anomalies; High risk level: This is for high-risk, unfamiliar, and frequently abnormal unlocking scenarios.

[0046] S64: Adaptive Adjustment Calculation of Identity Determination Threshold: Setting the System Baseline Identity Determination Threshold By combining comprehensive risk coefficients, continuous dynamic threshold adjustment is achieved. In high-risk scenarios, the judgment threshold is automatically raised to increase the verification rigor, while in low-risk scenarios, the threshold is appropriately lowered to improve unlocking smoothness. The calculation formula is as follows: ; in: : Adaptively adjusted real-time identity determination threshold, value range [ ,1]; The system's preset basic identity determination threshold is a fixed constant with a value range of (0,1), representing the default verification standard in normal scenarios.

[0047] The higher the risk coefficient, the greater the increase in the threshold and the stricter the identity verification; when the risk coefficient is 0, the threshold reverts to the benchmark value.

[0048] The beneficial effects of this invention include: First, through a spatiotemporal dual-dimensional adaptive weighting mechanism, real-time environmental parameters are used as spatial dimension factors, and user historical recognition behavior data and the long-term recognition stability of each modality are used as temporal dimension factors. The fusion weights of each biometric modality are dynamically adjusted according to the current environmental conditions and user habits. This allows the weight of a modality to be automatically reduced when its quality is degraded due to environmental interference, thus preventing low-quality noise features from entering the fusion calculation and improving the recognition system's adaptability and anti-interference performance in complex application scenarios.

[0049] Second, through controllable micro-Gaussian desensitization perturbation processing technology, an adaptive-level Gaussian micro-noise perturbation is applied to the original biometric vector. The perturbation coefficient is dynamically adjusted according to the modal quality and environmental conditions. The higher the quality and the better the conditions, the smaller the perturbation is to maximize the preservation of identity feature information. The worse the quality and the worse the conditions, the stronger the desensitization intensity is to avoid the risk of reverse leakage of dirty feature data, thus achieving a dynamic balance between security and recognition accuracy.

[0050] Third, through the scene risk adaptive threshold adjustment mechanism, the scene risk level is comprehensively quantified based on multi-dimensional information such as unlocking time period, unlocking location, and historical abnormal frequency. The identity determination threshold is continuously and dynamically adjusted according to the risk level. The threshold is automatically raised in high-risk scenes to enhance the security level, while the threshold is appropriately lowered in low-risk scenes to improve the user experience.

[0051] Fourth, by using a graded verification and dual-dimensional security control mechanism, a three-level judgment threshold system of high, medium and low is constructed. The system combines global fusion confidence and single-modal liveness credibility for dual verification. A secondary review process is initiated for suspicious states in the critical range. This avoids the problem of user experience degradation caused by direct rejection, prevents the security risks caused by direct approval, and eliminates the security vulnerability of overall score meeting the standard but some modal liveness failures.

[0052] Fifth, through a multi-dimensional scenario risk comprehensive assessment mechanism, multiple risk factors in time, space and behavior dimensions are comprehensively considered. An algorithm combining geometric mean accumulation and single-point high-risk correction is used to achieve a refined quantitative assessment of risk, making the risk control strategy more targeted and effective. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the smart door lock security control method based on multimodal biometric fusion of the present invention. Figure 2 This is a schematic diagram of the process for live anti-counterfeiting verification and feature quality evaluation of the present invention; Figure 3 This is a schematic diagram illustrating the process of adaptive threshold adjustment and hierarchical verification security control for scenario risks according to the present invention. Detailed Implementation

[0054] The following is in conjunction with the appendix Figures 1-3 The present invention will be further described in detail below: A multimodal biometric fusion-based smart lock security control method is applied to a smart lock system equipped with a multimodal biometric acquisition unit, a liveness detection unit, an environmental sensing unit, a main control unit, and a lock actuator. A bidirectional data channel is established between the multimodal biometric acquisition unit and the liveness detection unit. The output of the liveness detection unit is connected to the input of the main control unit, and the output of the environmental sensing unit is also connected to the input of the main control unit. The output of the main control unit is then connected to both the liveness detection unit and the lock actuator to transmit control commands and drive execution actions.

[0055] In the specific implementation process, the user first collects biometric data through the multimodal biometric acquisition unit. This acquisition unit supports the parallel acquisition of at least two different types of biometric features. A typical implementation may include any two or more combinations of fingerprint features, facial features, iris features, and vein features. During the acquisition process of each type of biometric feature, real-time environmental parameters and device operating status parameters are acquired simultaneously. These parameters are synchronously acquired by the environmental sensing unit and transmitted to the main control unit for subsequent processing.

[0056] The multimodal biometric acquisition unit comprises a high-resolution image sensor array, a depth camera module, a near-infrared illumination component, and a signal preprocessing circuit. The image sensor array acquires two-dimensional texture information of the biometric features; the depth camera module, with millimeter-level measurement accuracy, obtains three-dimensional spatial depth information of the biometric features; the near-infrared illumination component provides auxiliary illumination for biometric acquisition in low-light environments; and the signal preprocessing circuit amplifies, filters, and performs analog-to-digital conversion on the raw sensor signals, outputting standardized raw biometric data.

[0057] The hardware components of the environmental sensing unit include a light intensity sensor, a temperature and humidity sensor, a vibration sensor, an electromagnetic interference detection circuit, and an equipment operation status monitoring circuit. The electromagnetic interference detection circuit monitors the degree of interference from the surrounding electromagnetic environment on the signals acquired by the equipment; the equipment operation status monitoring circuit collects operating parameters such as battery voltage, charging status, storage capacity, and processor load rate in real time.

[0058] See appendix Figure 1 As shown, the method includes the following steps: S1. At least two different types of raw biometric data of the current user are collected in parallel by the multimodal biometric acquisition unit, and the real-time environmental parameters and device operating status parameters of the smart door lock are collected simultaneously by the environmental perception unit. S2. Perform liveness detection verification on each type of original biometric data using the liveness detection unit, and calculate the liveness confidence score for each type of biometric data. ,filter Biometric data exceeding the preset liveness confidence threshold proceeds to step S3; S3. Based on the preset feature quality evaluation rules, the selected biometric data are scored. Biometric data with a quality score lower than the preset quality score threshold are identified as invalid modal data and removed. Valid modal biometric data with a quality score that meets the standard and has completed the liveness verification are retained. S4. Using the real-time environmental parameters as spatial dimension factors and the user's historical recognition behavior data and the long-term recognition stability of each modality as time dimension factors, perform spatiotemporal dual-dimensional adaptive weight calculation on the biometric data of each effective modality to determine the dynamic fusion weight corresponding to each effective modality. S5. Apply a controllable micro Gaussian desensitization perturbation to the feature vectors corresponding to each effective modality of biometric data to obtain desensitized biometric vectors. Perform fusion operation on all desensitized biometric vectors based on the dynamic fusion weights corresponding to each effective modality and output the global identity fusion confidence. S6. Based on the unlocking time period, unlocking location, and historical unlocking anomaly frequency parameters of the smart door lock, determine the security risk level of the current unlocking scenario in real time, and adaptively adjust the preset identity determination threshold according to the security risk level. S7. Combining the liveness confidence scores of each modality and the adaptively adjusted identity determination threshold, the global identity fusion confidence is subjected to hierarchical verification. Based on the hierarchical verification results, the door lock actuator is driven to perform corresponding security control actions. If the global identity fusion confidence score and the liveness confidence score of each modality both meet the high-level threshold conditions, the door lock actuator will be driven to perform the unlocking operation. If the global identity fusion confidence score or liveness confidence score is in the medium-level threshold range, a secondary biometric verification and collection process will be triggered. If the global identity fusion confidence level is not up to standard or there is a modal liveness verification failure, the unlocking operation will be refused, the abnormal unlocking data will be recorded, and the hardware alarm mechanism will be triggered.

[0059] See Figure 2 As shown, the specific process of step S2 is as follows: S21: Multi-dimensional live cell core feature parameter extraction: targeting the loaded first... Similar to original biometric data, liveness feature parameters are extracted simultaneously from three anti-counterfeiting dimensions: texture, depth, and dynamics. All parameter values ​​are uniformly set within a certain range. The specific extraction rules are as follows: (1) Extracting live texture feature matching degree Extract detailed textures, edge contours, and feature point distribution information of the current biometric features, compare them with the legal liveness texture benchmark library pre-stored locally on the device, and output a texture matching score. This characterizes the realism of static textures; (2) Extracting effective values ​​of live depth features Collect 3D spatial depth information of current biometric features, calculate the completeness, continuity, and spatial hierarchy of the depth information, and output an effective depth score. Distinguish between two-dimensional forged media and the three-dimensional features of a real living object; (3) Extracting the feature scores of dynamic changes in living organisms Continuously acquire multiple frames of homologous biometric data to capture unique micro-dynamic features of living organisms (micro-deformation of skin, micro-movement of organs, changes in texture tension, etc.) and output dynamic effective scores. To resist static fake body attacks.

[0060] S22: Single-modal liveness confidence score fusion calculation: Based on the three types of liveness feature parameters extracted in S21, the score is calculated. Biomimetic characteristics liveness confidence score The calculation formula is as follows: ; in, : No. The liveness confidence score of bio-like characteristics, with a range of values. The higher the score, the higher the credibility of the modal liveness detection. : No. Biological-like live texture feature matching degree, value range ; : No. Effective values ​​and ranges of bio-like depth features in living organisms ; : No. Biological characteristics of in vivo dynamic change feature score, value range ; Fixed weighting coefficients for texture, depth, and dynamic features, satisfying the constraints. ,and The coefficients are fixed in the liveness detection unit to ensure a stable and consistent calibration benchmark.

[0061] S23: Full Modal Traversal Calculation and Data Integration: Repeat steps S21-S22 to complete all operations sequentially. The liveness scores of bio-feature modalities are calculated and integrated to obtain a complete set of modal liveness confidence scores. .

[0062] In step S22 Biological-like live texture feature matching degree Effective value of live depth features and scores for dynamic changes in living organisms The calculation processes are as follows: Calculate the first Biological-like live texture feature matching degree : will the first The keypoint texture feature vectors of the original biological features are compared with the locally pre-stored standard live texture reference vectors of this modality using cosine similarity matching, as shown in the following formula: ; in: : No. The currently acquired texture feature vector for bio-like characteristics; : No. Pre-stored standard live texture reference vectors for bio-like features; Vector 2 norm operation, used to normalize characteristic magnitudes; Texture matching degree, with a value range of: The higher the value, the better the texture details match the real living features; Calculate the first Effective value of bio-like in vivo depth features Based on 3D depth point cloud data of biological features, and combining depth integrity and depth value variance, the effective value of the depth feature of a living organism is quantified, as shown in the following formula: ; in: : No. Number of effective depth point clouds for biological-like features; : No. The total number of depth point clouds collected for bio-like features; : No. Biological feature depth value variance is used to characterize the sense of three-dimensional depth hierarchy; Depth completeness weighting coefficient, with a fixed value. ; Valid values ​​and ranges of live depth features It effectively distinguishes between two-dimensional planar forgery and three-dimensional live features.

[0063] Calculate the first Biological characteristics of living dynamic changes feature score Based on homologous biological feature data acquired in multiple consecutive frames, the mean value of inter-frame feature differences characterizes the micro-dynamic characteristics of living organisms and suppresses static prosthetic features. The calculation formula is as follows: ; in: : No. The mean inter-frame feature difference of consecutive multi-frame images of bio-like characteristics characterizes the degree of micro-dynamic changes in living organisms. : Dynamic characteristic gain coefficient, which is a preset fixed constant used to normalize the dynamic difference amplitude; : Live organism dynamic change characteristic score, value range Static forgeries have no inter-frame differences, and their corresponding scores approach 0.

[0064] The specific process of step S3 is as follows: S31: Load the biometric data of each modality that passed the S2 liveness verification one by one, and extract the first... Three basic quality assessment parameters for bio-like characteristics were normalized to the [0,1] interval: Feature completeness : Characterizes the completeness of key feature points of a single-modality biological feature collection; the fewer missing feature points, the higher the score. Feature clarity It characterizes the clarity of biological features, textures, and contours, avoiding data distortion caused by blurring, ghosting, and noise interference. Operating condition adaptability The environment and equipment adaptation score, obtained based on real-time illumination, temperature and humidity, and the signal-to-noise ratio of equipment acquisition, characterizes the degree of interference of the current working conditions on feature acquisition.

[0065] S32: Modal raw quality score calculation: A scoring mechanism of minimum value short-board constraint + mean balance correction is adopted. The short-board effect of the quality dimension is used to suppress non-compliant data, and the score is calculated for the [missing value]. Raw quality score of the modality : ; in: : No. The original quality score of the bio-characteristic modality, with a value range of [0,1]; : No. Modal feature completeness, with a value range of [0,1]. : No. Class modality feature clarity, with a value range of [0,1]; S33: Calculation of the final comprehensive quality score integrating liveness confidence: To avoid misjudgments due to qualified liveness but low feature confidence or inflated original quality, the liveness confidence score output from S2 is introduced. The original quality scores are corrected to obtain the final comprehensive quality score used for modality screening. This achieves a dual-coupling evaluation of safety and quality attributes, as shown in the following formula: ； in: : No. The final comprehensive quality score of the biometric modality ranges from [0,1]. A higher score indicates higher security and usability of the data in that modality. : No. Modal raw quality score; : The output of S2 Modal liveness confidence score.

[0066] Logical constraint explanation: The lower the liveness confidence, the stronger the correction attenuation of the original quality score, which can effectively reduce the flawed modality score of "doubtful liveness, seemingly qualified collection quality", and prevent flawed data from entering the fusion process.

[0067] S34: Failure Mode Identification and Removal Screening: Preset Uniform Modal Quality Threshold The final comprehensive quality score for each modality. Perform threshold determination and tiered filtering: like The modality is determined to be valid modality biometric data, retained, and included in the subsequent weight calculation process; like If the mode is determined to be a failed mode, it will be discarded directly.

[0068] Feature completeness Feature clarity and working condition adaptability The specific calculation process is as follows: Calculate feature completeness : The formula is calculated based on the ratio of the number of valid key feature points detected by current biometrics to the total number of standard feature points for this modality: ; in: : No. The number of valid key feature points detected in the currently acquired image of the modality; : No. The total number of standard complete key feature points preset for each modality; value constraint: if the calculated result is greater than 1, it will be forcibly truncated to 1 to ensure... The more feature points are missing, the lower the completeness.

[0069] Calculate feature sharpness : Image gradient variance is used to characterize texture sharpness; higher scores are obtained when there is no blur or noise interference. The formula is: ; in: : No. The gradient variance of the current feature image of the class modality represents the sharpness of texture edges; : No. The predefined standard clear image gradient variance baseline value for the modality; value constraints: The blurrier the image, the more noise it has, the smaller the gradient variance, and the lower the sharpness score.

[0070] Calculate the adaptability of the working conditions : Normalization calculation based on multi-dimensional limit constraints of device signal-to-noise ratio, ambient light interference, and temperature and humidity interference: ; in: : Current number Real-time signal-to-noise ratio of the modal acquisition unit; The device has a preset standard signal-to-noise ratio threshold for data acquisition. Real-time ambient light intensity; : Preset optimal standard illumination intensity; Value constraints: set to 0 when the calculation result is less than 0, and set to 1 when the result is greater than 1. The greater the illumination offset and the lower the signal-to-noise ratio, the worse the adaptability to operating conditions.

[0071] The specific process of step S4 is as follows: S41: For each type of valid modality biometric data, extract the spatial dimension factor and the temporal dimension factor respectively, and normalize all parameters to the [0,1] interval: Spatial Dimension Environmental Adaptation Factor Based on the quantification of S1 real-time environmental parameters and equipment operating status parameters, it characterizes the impact of current illumination, signal-to-noise ratio, and environmental interference on the first... The better the adaptation of modal acquisition results, the less environmental interference and the better the acquisition conditions. The higher the score; Time-dimensional steady-state adaptation factor The stability of a modality is quantified by combining its long-term recognition stability with historical user recognition behavior data. It characterizes the long-term adaptability of the modality to the current user; the higher the historical recognition success rate, the stronger the feature stability. The higher the score.

[0072] S42: Single-modal synthesis fit calculation: Combined with the modal synthesis quality score output from the preceding S3. By integrating spatial and temporal factors, the first... Comprehensive fit of effective modes : ; in: : No. The comprehensive fit of the effective modalities, with a value range of [0,1]. The higher the score, the stronger the recognition reliability and fit of the modalities in the current spatiotemporal scene. : No. Effective modal composite quality score; : No. The modal space dimension environment adaptation factor, with a value range of [0,1]; : No. The steady-state adaptation factor for the modal time dimension has a value range of [0,1].

[0073] Employing multi-factor product coupling logic, defects in any dimension (poor quality, large environmental interference, weak long-term stability) will significantly reduce the modal synthesis fit.

[0074] S43: Calculation of Dynamic Weights for Effective Modalities: To meet the allocation constraints of multimodal fusion weights, the overall fitness of all effective modalities is globally normalized to obtain the final dynamic fusion weights for each modality. To ensure that the weight allocation aligns with the current global modal quality and spatiotemporal conditions, the formula is as follows: ; in: : No. The dynamic fusion weights of the effective modalities take values ​​in the range (0,1); : No. Comprehensive fit of effective modes; : No. Comprehensive fit of effective modes, For valid modal traversal indexes; : Total number of valid modes after S3 screening.

[0075] Weight constraint property: The dynamic fusion weights of all effective modes satisfy... .

[0076] Spatial Dimension Environmental Adaptation Factor and time-dimensional steady-state adaptation factor The calculation process is as follows: Spatial Dimension Environment Adaptability Factor Calculation: Based on real-time environmental parameters collected by S1, equipment operating status parameters, and operating condition adaptability calculated by S3. An exponential penalty correction is applied based on the device's signal-to-noise ratio offset, calculated using the following formula: ; in: : No. The modal space dimension environment adaptation factor has a final constraint range of [0,1]. : No. Modal condition adaptability characterizes the overall degree of adaptability to illumination and environmental interference. : No. The real-time signal-to-noise ratio of the modal acquisition unit is derived from the operating status parameters of the S1 device. : The optimal signal-to-noise ratio benchmark value for the corresponding modal preset standard; Boundary constraints: Calculated values ​​greater than 1 are forcibly set to 1, and values ​​less than 0 are forcibly set to 0.

[0077] The calculated exponential term is the signal-to-noise ratio (SNR) deviation penalty term. The greater the deviation of the equipment's SNR from the standard operating condition, the stronger the penalty attenuation. The lower the value, the better. Only when the operating condition adaptability is high and the equipment signal-to-noise ratio meets the standard, the spatial adaptability factor approaches 1, which represents the current environment's true adaptability to the acquisition of this mode.

[0078] Time-dimensional steady-state adaptation factor calculation: Based on local historical recognition behavior data, the calculation is coupled with the dual indicators of modality historical recognition success rate and feature stability. Instead of weighted fusion, a bottleneck constraint mechanism is used to ensure that long-term unstable modalities are naturally downweighted. The calculation formula is as follows: ; in: : No. The steady-state adaptation factor for the modal time dimension, with a value range of [0,1]. : The first in the preset statistical period The historical recognition success rate of the modality is obtained by statistically analyzing the effective unlocking and recognition samples, with a value range of [0,1]. : The first in the preset statistical period The historical fluctuation variance of modality feature scores characterizes the long-term stability of user biometrics, with a value range of [0,1]. Taking the minimum of success rate and stability is an unweighted bottleneck constraint, which avoids the incorrect overestimation of modes that are "high success rate but large fluctuations" or "good stability but low success rate", and is consistent with the real steady-state adaptation logic of long-term use.

[0079] Boundary constraints: The calculation results are automatically clamped in the interval [0,1].

[0080] The specific process of step S5 is as follows: S51: Extraction of Original Feature Vectors for Effective Modalities: For each class of effective modal biometric data identified in S4, extract standardized original feature vectors with uniform dimensions, denoted as the first... The original feature vector of the effective mode is The original feature vector To create a composite feature vector that integrates the dimensions of live anti-counterfeiting and biological ontology features, the original feature vector... Specifically, it includes: Texture feature components: corresponding to the S2 texture feature matching degree derived quantization features; Depth feature components: corresponding to the effective features in the S2 3D depth space; Dynamic micro-variation feature components: corresponding to the dynamic change features of S2 in vivo.

[0081] The specific expression is: ; in: : No. The matching degree of live texture features of bio-like modalities, with values ​​[0,1]; : No. The effective value of the live depth feature of the bio-feature modality, taking values ​​[0,1]; : No. The feature score of the in vivo dynamic change of the bio-like feature mode is taken as [0,1].

[0082] S52: Adaptive Micro-Gaussian Desensitization Perturbation Coefficient Calculation: To avoid high accuracy loss or desensitization failure caused by fixed noise, this invention designs an adaptive micro-Gaussian desensitization perturbation coefficient based on prior modal quality parameters and environmental operating condition parameters. Higher quality and better operating conditions result in lower perturbation noise. When operating interference is high and feature quality is low, the perturbation intensity is appropriately increased to prevent privacy leakage while ensuring recognition accuracy. The formula for calculating the adaptive micro-Gaussian desensitization perturbation coefficient is as follows: ; in: : No. Modal adaptive infinitesimal Gaussian desensitization perturbation coefficient, with a value range of (0,1). The smaller the coefficient, the more minute the applied perturbation noise. The system presets a basic disturbance reference coefficient, which is a fixed minimum constant, to ensure basic desensitization protection in all scenarios; : No. The overall quality score for each modality; : No. Modal space dimension environment adaptation factor.

[0083] Logical constraint description: The high-quality, high-performance mode uses micro-perturbations to retain identity feature information to the maximum extent; the low-quality, poor-performance mode strengthens the desensitization intensity to avoid the risk of reverse leakage of dirty feature data and achieve a dynamic balance between security and accuracy.

[0084] S53: Modal Desensitization Biometric Vector Generation: Based on an adaptive infinitesimal Gaussian desensitization perturbation coefficient, a controllable infinitesimal Gaussian noise perturbation is applied to the original feature vector to generate a desensitized biometric vector, preventing the original biometric features from being reverse-engineered. The formula is as follows: ; in: : No. Modal desensitization biometric vectors; : No. Modal-based original biological feature vectors; The mean is 0 and the standard deviation is 0. Gaussian micro-noise ensures that the disturbance is unbiased and does not disrupt the distribution of core identity features.

[0085] S54: Multimodal desensitization feature weighted fusion operation: Calls the dynamic fusion weights of each modality output by S4. We then perform a weighted fusion of all desensitized biometric feature vectors to obtain a global fused feature vector, as shown in the following formula: ; in: : Global desensitization feature vector after multimodal fusion; : No. Effective modal dynamic fusion weights; : No. Modal desensitization of biological feature vectors.

[0086] S55: Global Identity Fusion Confidence Measurement Output: The global de-identified fusion feature vector is matched with the user's locally stored standard fusion benchmark feature vector using cosine similarity to quantify and generate the global identity fusion confidence score, thereby quantifying the degree of identity trustworthiness. The formula is as follows: ; in: Global identity fusion confidence score, with a value range of [0,1]. A higher score indicates a higher level of confidence in matching the current user's identity. The user's locally stored standard fusion benchmark feature vector is a de-identified compliant template with no original privacy data. Vector 2-norm operation.

[0087] Boundary constraints: When the calculation result is greater than 1, it is forced to be set to 1; when it is less than 0, it is forced to be set to 0.

[0088] See Figure 3 The specific process of step S6 is as follows: S61: Calculate the three core risk control factors for the current unlocking scenario, including the risk factor during the unlocking period. Unlocking location risk factors and historical anomaly frequency risk factors The higher the risk, the greater the corresponding factor score; Unlocking Period Risk Factors Calculation: Based on the quantification of the deviation between the current unlock time and the user's regular authorization period, a time deviation normalization penalty is calculated using the following formula: ; in: : Risk factor for the unlocking period, value range [0,1]; : Current real-time unlock time (24-hour quantized time value); : Average of standard times for users to frequently authorize unlocking on a daily basis; : Preset the maximum legal time deviation threshold, used to normalize the time deviation amount.

[0089] The closer the unlock time is to the regular time period, The closer it gets to 0; abnormal unlocking times such as late at night and early morning, which deviate from the norm. Approaching 1.

[0090] Unlock location risk factors Calculation: Based on the spatial distance deviation quantification between the current unlocked location and commonly used trusted locations, the distance deviation index penalty is used for calculation, as follows: ; in: Unlock location risk factor, value range [0,1]; : The real-time spatial distance deviation between the current unlock location and the user's commonly used trusted unlock locations on the local network; : The maximum permissible deviation distance of the preset reliable position is used to normalize the spatial deviation.

[0091] The deviation is minimal when unlocking at a commonly used local location. The value approaches 0; however, the deviation is extremely large when unlocking in unfamiliar locations. Approaching 1.

[0092] Historical anomaly frequency risk factor Calculation: Based on the proportion of abnormal unlocking times to the total number of recognition times within the statistical period, frequency saturation normalization is used for calculation, as shown in the following formula: ; in: : Historical unlocking anomaly frequency risk factor, value range [0,1]; : The cumulative number of abnormal unlock attempts, including liveness detection, feature matching, and illegal attempts, within a preset statistical period; Total number of unlocks within the same statistical period.

[0093] Combining the proportion of anomalies with the saturation characteristics of cumulative anomalies, minor and sporadic anomalies slightly increase the risk, while frequent and repeated anomalies will quickly saturate and raise the risk score.

[0094] S62: Calculation of Scenario-Based Comprehensive Risk Coefficient: A logic-constraint cumulative unweighted risk quantification algorithm is adopted. The core principle is: low-risk factors are cumulatively superimposed, while high-risk factors are subject to mandatory capping constraints. This preserves the cumulative effect of multi-dimensional anomalies while ensuring that high risk in any dimension can significantly increase the overall risk. The formula is as follows: ; in: : The overall risk coefficient of the current unlocking scenario, with a value range of [0,1]. The larger the value, the higher the security risk of the unlocking scenario. : Risk factor for the unlocking period, value range [0,1]; Unlock location risk factor, value range [0,1]; : Historical unlocking frequency risk factor, value range [0,1].

[0095] Logical constraint description: It consists of a geometric mean cumulative term and a single-point high-risk correction term; the geometric mean term retains the cumulative superposition effect of three-dimensional risk, and the overall risk steadily increases when multiple dimensions are abnormal at the same time; max Single-point correction provides a safety net for high-risk scenarios. If any one dimension of the scenario is extremely risky, the overall risk coefficient can be significantly increased. This effectively solves the risk control loopholes where minor anomalies in multiple dimensions are averaged out and high-risk single points are masked, thus improving the reliability of door lock security.

[0096] Boundary constraints: Set the result to 0 when it is less than 0 and to 1 when it is greater than 1, so as to maintain consistency with the parameter value range system of the whole text.

[0097] S63: Safety Risk Level Classification Determination: Based on Comprehensive Risk Coefficient The system categorizes security risks into three levels, with thresholds fixed within the system to adapt to different risk control intensities in various scenarios. Specific grading rules are as follows: Low risk level: This corresponds to conventional trusted unlocking scenarios; Medium risk level: This corresponds to suspicious unlocking scenarios with minor anomalies; High risk level: This is for high-risk, unfamiliar, and frequently abnormal unlocking scenarios.

[0098] S64: Adaptive Adjustment Calculation of Identity Determination Threshold: Setting the System Baseline Identity Determination Threshold By combining comprehensive risk coefficients, continuous dynamic threshold adjustment is achieved. In high-risk scenarios, the judgment threshold is automatically raised to increase the verification rigor, while in low-risk scenarios, the threshold is appropriately lowered to improve unlocking smoothness. The calculation formula is as follows: ; in: : Adaptively adjusted real-time identity determination threshold, value range [ ,1]; The system's preset basic identity determination threshold is a fixed constant with a value range of (0,1), representing the default verification standard in normal scenarios.

[0099] The higher the risk coefficient, the greater the increase in the threshold and the stricter the identity verification; when the risk coefficient is 0, the threshold reverts to the benchmark value.

[0100] The specific process of step S7 is as follows: S71: Construction of a three-level hierarchical verification threshold system: Real-time identity determination threshold obtained through adaptive adjustment by S6 Linked to S2 liveness confidence benchmark threshold A three-level (high, medium, and low) threshold system for adaptive hierarchical verification is constructed, reusing preceding parameters throughout the process without adding any new custom subjective parameters. The threshold grading formula is as follows: ; in: High-level security verification threshold, the value of which is equal to the real-time dynamic identity determination threshold, is the highest security standard for unlocking; The intermediate-level security verification threshold is a critical threshold for graded review, used to define the suspicious transition range. Low-level security verification threshold: This is the minimum legal threshold for unlocking. If the threshold is lower than this value, it is directly judged as identification failure or high-risk anomaly. The real-time identity determination threshold, which is adaptively adjusted by S6, inherits the results of previous steps and changes dynamically with the risk of the scenario.

[0101] S72: Dual-dimensional verification condition judgment: A dual judgment mechanism is adopted, which combines global confidence overall verification with single-modal liveness fallback verification. This ensures the overall accuracy of identity matching while eliminating single-modal liveness fraud vulnerabilities. The three-level judgment rules are as follows: High-level unlock criteria (completely secure): An identity is deemed highly trustworthy and legitimate if both of the following conditions are met: 1) Global identity fusion confidence level reaches a high-level threshold: ; 2) All valid modal liveness confidence levels met the standards: .

[0102] The global identity matching rate is high, and all biometrics involved in the fusion are real and alive, with no risk of forgery and no security vulnerabilities, allowing for direct and secure unlocking.

[0103] Intermediate-level review criteria (suspicious transition): If the criteria for a higher level are not met, but any of the following interval criteria are met, the status is determined to be suspicious and pending review: 1) The confidence level of global identity fusion is in the medium range: ; 2) Global confidence level meets the standard, but some modal liveness confidence levels only meet the critical standard: single modality Approaching No liveness check failed mode.

[0104] There is slight uncertainty in the overall identity matching or liveness status. There are no obvious signs of forgery or attack. The system does not directly unlock or reject the identity. Instead, it initiates a second verification to improve recognition accuracy, thus balancing security and user experience.

[0105] Low-level rejection criteria (abnormal risk): Meeting any of the following conditions will result in being deemed an illegal or high-risk identity: Global identity fusion confidence level did not reach the low threshold: ; There exists an instance where the liveness confidence level fails to meet the standard for any valid modality: .

[0106] If the overall identity matching is insufficient, or if there are issues such as modal liveness detection forgery or failure of liveness verification, there is a clear security risk. Therefore, the system will directly refuse to unlock the device and trigger a risk control alarm.

[0107] S73: Execution of graded safety control actions: Based on the aforementioned three-level judgment results, the corresponding drive door lock actuator executes differentiated safety control actions, forming a closed-loop risk control logic: High-level judgment passed → Unlock operation executed: The door lock actuator drives the bolt to open, completing this legal unlocking process. At the same time, the record of this normal unlocking is retained for subsequent user behavior steady-state data updates.

[0108] Medium-level suspicious judgment → Triggering the secondary biometric verification and collection process: Call the multimodal biometric collection unit to re-collect the current user's core biometric data, repeat the S2-S7 core verification process, and eliminate the judgment deviation caused by environmental interference and instantaneous collection error through secondary verification. If the secondary verification is successful, the alarm will be unlocked; if there is still suspicion, the alarm will be downgraded to rejection.

[0109] Low-level risk assessment → Refusal to unlock + anomaly recording + hardware alarm: Immediately prohibit the door lock unlocking action, and record the time, location, feature matching score, and liveness abnormal modal information of this unlocking to the local anomaly log. Simultaneously trigger the door lock hardware alarm mechanism (buzzer alarm, cloud anomaly push) to achieve risk tracing and proactive security warning.

[0110] The above embodiments are only used to illustrate specific implementations of the present invention. Equivalent substitutions or modifications made by those skilled in the art based on the technical solutions of the present invention still fall within the protection scope of the present invention.

Claims

1. A smart door lock security control method based on multimodal biometric fusion, applied to a smart door lock system equipped with a multimodal biometric acquisition unit, a liveness detection unit, an environmental sensing unit, a main control unit, and a door lock actuator, characterized in that, Includes the following steps: S1. Collect at least two different types of raw biometric data of the current user through the multimodal biometric acquisition unit, and collect the real-time environmental parameters and device operating status parameters of the door lock through the environmental perception unit. S2. The liveness detection unit calculates the liveness confidence score for each type of biological feature based on the original biological feature data for each type. ,filter Biometric data exceeding a preset liveness confidence threshold; S3. Perform quality scoring on the screened biometric data, and retain valid modal biometric data with quality scores greater than or equal to the preset quality score threshold and with liveness verification completed. S4. Using the real-time environmental parameters as spatial dimension factors and the user's historical recognition behavior data and the long-term recognition stability of each modality as time dimension factors, perform spatiotemporal dual-dimensional adaptive weight calculation on the biometric data of each effective modality to determine the dynamic fusion weight corresponding to each effective modality. S5. Apply a controllable micro Gaussian desensitization perturbation to the feature vectors corresponding to each effective modality of biometric data to obtain desensitized biometric vectors. Perform fusion operation on all desensitized biometric vectors based on the dynamic fusion weights corresponding to each effective modality and output the global identity fusion confidence. S6. Based on the unlocking time period, unlocking location, and historical unlocking anomaly frequency parameters of the smart door lock, determine the security risk level of the current unlocking scenario in real time, and adaptively adjust the preset identity determination threshold according to the security risk level. S7. Combining the liveness confidence scores of each modality and the adaptively adjusted identity determination threshold, the global identity fusion confidence is graded and verified, and the door lock actuator is driven to perform corresponding security control actions based on the graded verification results.

2. The smart door lock security control method based on multimodal biometric fusion according to claim 1, characterized in that, The specific process of step S2 is as follows: S21: For the loaded first Similar to original biometric data, liveness feature parameters are extracted simultaneously from three anti-counterfeiting dimensions: texture, depth, and dynamics, including liveness texture feature matching degree. Effective value of live depth features and scores for dynamic changes in living organisms ; S22: Based on live texture feature matching degree Effective value of live depth features and scores for dynamic changes in living organisms Calculate the first Biomimetic characteristics liveness confidence score ; S23: Repeat steps S21-S22 sequentially to complete all steps. The liveness scores of bio-feature modalities are calculated and integrated to obtain a complete set of modal liveness confidence scores. .

3. The smart door lock security control method based on multimodal biometric fusion according to claim 2, characterized in that, In step S22 Biological-like live texture feature matching degree Effective value of live depth features and scores for dynamic changes in living organisms The calculation processes are as follows: Calculate the first Biological-like live texture feature matching degree By passing the first The key point texture feature vectors of the original biological features are obtained by cosine similarity matching with the standard live texture reference vectors of the corresponding modalities stored locally. Calculate the first Effective value of bio-like in vivo depth features Based on 3D depth point cloud data of biological features, and combining depth integrity and depth value variance, the effective value of the depth feature of the living organism is quantified. Calculate the first Biological characteristics of living dynamic changes feature score Based on homologous biological feature data acquired in multiple consecutive frames, the mean value of inter-frame feature differences is used to characterize the micro-dynamic characteristics of living organisms, suppress static prosthetic features, and obtain the feature score of dynamic changes in living organisms.

4. The smart door lock security control method based on multimodal biometric fusion according to claim 1, characterized in that, The specific process of step S3 is as follows: S31: Load the biometric data of each modality that passed the S2 liveness verification one by one, and extract the first... Three basic quality assessment parameters for bio-like characteristics include feature completeness. Feature clarity and working condition adaptability ; S32: A scoring mechanism combining minimum value bottleneck constraint and mean balance correction is adopted. This mechanism utilizes the bottleneck effect in the quality dimension to suppress non-conforming data and calculates the... Raw quality score of the modality : S33: Introduce the liveness confidence score output by S2. The original quality scores are corrected to obtain the final comprehensive quality score used for modality screening. This enables a dual-coupling evaluation of safety and quality attributes. S34: Preset uniform modal quality threshold The final comprehensive quality score for each modality. Perform threshold determination and tiered filtering: like The modality is determined to be valid modality biometric data, retained, and included in the subsequent weight calculation process; like If the mode is determined to be a failed mode, it will be discarded directly.

5. The smart door lock security control method based on multimodal biometric fusion according to claim 4, characterized in that, Feature completeness Feature clarity and working condition adaptability The specific calculation process is as follows: Calculate feature completeness The value is calculated based on the ratio of the number of valid key feature points detected by current biometrics to the total number of standard feature points of the corresponding modality. Calculate feature sharpness The image gradient variance is used to characterize the sharpness of the texture, and then the feature sharpness is calculated. Calculate the adaptability of the working conditions The result is obtained by normalization calculation based on multi-dimensional limit constraints of device signal-to-noise ratio, ambient light interference, and temperature and humidity interference.

6. The smart door lock security control method based on multimodal biometric fusion according to claim 4, characterized in that, The specific process of step S4 is as follows: S41: For each type of effective modality of biometric data, extract the spatial dimension environmental adaptation factor. steady-state adaptation factor with time dimension ; S42: Overall quality score combined with the output of S3 Integrating spatial dimension environmental adaptation factors steady-state adaptation factor with time dimension Calculate the first Comprehensive fit of effective modes ; S43: Perform global normalization on the overall fit of all effective modes to obtain the final dynamic fusion weights for each mode. .

7. The smart door lock security control method based on multimodal biometric fusion according to claim 6, characterized in that, Spatial Dimension Environmental Adaptation Factor and time-dimensional steady-state adaptation factor The calculation process is as follows: Spatial Dimension Environmental Adaptation Factor Calculation: Based on the real-time environmental parameters collected by S1, the equipment operating status parameters, and the operating condition adaptability calculated by S3. By combining the device signal-to-noise ratio offset with an exponential penalty correction, the spatial dimension environment adaptation factor is obtained. Time-dimensional steady-state adaptation factor Calculation: Based on local historical recognition behavior data, the calculation is performed by coupling the success rate of modality historical recognition and feature stability as dual indicators, and a short-board constraint mechanism is adopted to calculate the steady-state adaptation factor in the time dimension.

8. The smart door lock security control method based on multimodal biometric fusion according to claim 1, characterized in that, The specific process of step S5 is as follows: S51: For each type of valid modal biometric data identified in S4, extract the original feature vector with uniform dimensions, and denote it as the first... The original feature vector of the effective mode is The original feature vector The composite feature vector integrates the dimensions of live anti-counterfeiting and biological ontology features, specifically including texture feature components, depth feature components, and dynamic micro-variation feature components. S52: Adaptive micro-Gaussian desensitization perturbation coefficient calculation based on comprehensive quality score and spatial dimension environmental adaptation factor; S53: Based on an adaptive infinitesimal Gaussian desensitization perturbation coefficient, a controllable infinitesimal Gaussian noise perturbation is applied to the original feature vector to generate a desensitized biofeature vector. ; S54: Call the dynamic fusion weights of each modality output by S4. All desensitized biometric feature vectors are weighted and fused to obtain a global fused feature vector. ; S55: Perform cosine similarity matching between the global de-identified fusion feature vector and the user's locally stored standard fusion benchmark feature vector to obtain the global identity fusion confidence score. .

9. The smart door lock security control method based on multimodal biometric fusion according to claim 1, characterized in that, The specific process of step S6 is as follows: S61: Calculate the three core risk control factors for the current unlocking scenario, including the risk factor during the unlocking period. Unlocking location risk factors and historical anomaly frequency risk factors ; S62: Calculate the comprehensive risk coefficient of the current unlocking scenario using a logic and constraint cumulative unweighted risk quantification algorithm. ; S63: Safety Risk Level Classification Determination: Based on Comprehensive Risk Coefficient The safety risk levels are divided into three categories: low risk, medium risk, and high risk. S64: Set the system baseline identity determination threshold By combining comprehensive risk coefficients, continuous dynamic threshold adjustment is achieved. In high-risk scenarios, the judgment threshold is automatically raised to increase the verification rigor, while in low-risk scenarios, the threshold is appropriately lowered to improve the unlocking smoothness.

10. The smart door lock security control method based on multimodal biometric fusion according to claim 9, characterized in that, Unlocking the risk factor during the time period in step S61 Unlocking location risk factors and historical anomaly frequency risk factors The calculation process is as follows: Unlocking Period Risk Factors Calculation: Based on the quantification of the deviation between the current unlock time and the user's regular authorization period, a time deviation normalization penalty calculation is used; Unlock location risk factors Calculation: Based on the spatial distance deviation quantification between the current unlocked location and commonly used trusted locations, the distance deviation index penalty is used for calculation; Historical anomaly frequency risk factor Calculation: Based on the proportion of abnormal unlocking times to the total number of recognition times within the statistical period, the calculation adopts frequency saturation normalization.

Images