A security authentication method and system based on multi-modal biometric feature fusion

Abstract

The application discloses a kind of based on multimodal biological feature fusion security authentication method and system, it is related to biological feature identification technical field, by constructing environmental factor influence performance index dataset, environmental factor influence performance index dataset is grouped using clustering algorithm, the environmental condition classification set of group containing insufficient light and normal noise is determined, for the performance fluctuation under the condition such as insufficient light and noise interference, sliding window method is used to analyze change trend, generate core adaptive rule set, and the corresponding adaptive adjustment parameter is verified by simulation environment, finally form biological recognition authentication output rule set, realize constructing a comprehensive analysis framework, systematically monitor and comb each performance data in authentication process, and find the law of performance change through these data, ensure the authentication stability and accuracy under complex environment, significantly improve the adaptability and reliability of biological recognition technology.

Description

Technical Field

[0001] This invention relates to the field of biometric recognition technology, and in particular to a secure authentication method and system based on multimodal biometric fusion. Background Technology

[0002] The field of multimodal biometric fusion security authentication has received considerable attention in recent years. Its importance lies in its ability to significantly improve the security and reliability of identity verification by combining multiple human characteristics, such as facial features, fingerprints, and voice. This technology has become an indispensable safeguard in scenarios such as financial payments, public safety, and smart device unlocking. Especially in environments with high security requirements, multimodal fusion is considered a key direction for addressing complex threats.

[0003] However, existing methods still face many challenges in practical applications, particularly in comprehensively evaluating and optimizing authentication effectiveness. Many solutions focus only on the performance of a single step, neglecting the interactions between different factors during the authentication process. This makes it difficult for the system to maintain stability and accuracy when facing complex environments or changes in user behavior. This limitation casts doubt on the overall reliability of the authentication system, especially its adaptability in dynamic scenarios, which urgently needs improvement.

[0004] In this field, the core technical challenges lie primarily in how to conduct comprehensive and detailed analysis and monitoring of the authentication process. Firstly, authentication involves multiple performance metrics, such as the recognition accuracy of each feature, the false rejection rate, and the overall effect after fusion. These metrics are intricately correlated, and failure to systematically collect and analyze them can lead to inaccurate assessments of system performance. More critically, this analysis needs to cover changes in user behavior over different times and in different environments. For instance, in low-light or noisy environments, the collection quality of certain features may degrade, thus affecting the reliability of fusion decisions. These two factors are closely related, and the lack of metric analysis directly exacerbates the difficulty of adapting the system to environmental changes. Summary of the Invention

[0005] This invention provides a secure authentication method and system based on multimodal biometric fusion, which aims to build a comprehensive analysis framework to systematically monitor and analyze various performance data during the authentication process, and discover patterns in performance changes through these data, especially the fluctuations in user authentication experience under different scenarios.

[0006] In a first aspect, the present invention provides a security authentication method based on multimodal biometric fusion, executed by a computer, comprising:

[0007] The accuracy of facial recognition and the pass rate of voice verification related to biometric authentication are obtained. At the same time, the classification value of ambient light intensity level and the label of ambient noise decibel range at the corresponding time are obtained to obtain a dataset of performance indicators affected by environmental factors.

[0008] Based on the dataset of performance indicators affected by environmental factors, a clustering algorithm was used to group the data, resulting in a set of environmental condition classifications including groups with insufficient light and normal noise, groups with normal light and noise interference, groups with dual environmental interference, and groups with sufficient light and low noise.

[0009] Based on the low-light and normal-noise group, the accuracy change trend over a continuous time period is calculated using the sliding window method, and the performance change pattern of biometrics in low-light environment is obtained.

[0010] Based on the performance change pattern sequence, a classification algorithm is used to mine rules and generate a set of core adaptive rules that include branch judgments based on environmental conditions.

[0011] Based on the core set of adaptation rules, the adaptive adjustment parameters corresponding to each environmental condition classification set are determined, and the adaptive adjustment parameters are applied in a simulated biometric authentication environment to obtain an optimized performance index dataset.

[0012] The optimized performance index dataset and the environmental factor-affected performance index dataset are merged. The merged dataset is then grouped using a clustering algorithm to obtain a set of biometric authentication output rules. This set of biometric authentication output rules is used to perform secure authentication operations based on biometrics.

[0013] Optionally, obtain the facial recognition accuracy and voice verification pass rate related to biometric authentication, and simultaneously obtain the classification values ​​of ambient light intensity levels and ambient noise decibel range labels at the corresponding times to obtain a dataset of performance indicators affected by environmental factors, including:

[0014] Based on the record entries in the log file of the biometric authentication system, the accuracy of facial recognition and the pass rate of voice verification related to biometric authentication are obtained, and the performance index sequence data is obtained.

[0015] Based on the recorded entries, the classification values ​​of ambient light intensity levels and the decibel range labels of ambient noise at the corresponding time are extracted to obtain the environmental factor label sequence.

[0016] If the record item includes facial recognition accuracy, then the ambient light intensity level classification value of the corresponding record item is associated with the facial occlusion object category label to generate a facial recognition environmental impact sample pair;

[0017] If the record entry contains the sound verification pass rate, then the environmental noise decibel range label of the corresponding record entry will be associated and bound with the sound-affected object category label to generate a sound verification environmental impact sample pair;

[0018] Based on performance index sequence data, environmental factor label sequence, facial recognition environmental impact sample pairs, and sound verification environmental impact sample pairs, the environmental factor impact performance index dataset is determined.

[0019] Optionally, based on the dataset of performance indicators affected by environmental factors, a clustering algorithm is used to group the data, resulting in a set of environmental condition classifications including groups with insufficient light and normal noise levels, including:

[0020] Based on the dataset of performance indicators affected by environmental factors, a classification algorithm was used to perform correlation analysis on the classification values ​​of facial recognition accuracy and ambient light intensity level, and the first correlation matrix of the influence of facial recognition accuracy on the classification values ​​of ambient light intensity level was obtained.

[0021] Based on the dataset of performance indicators affected by environmental factors, a classification algorithm was used to perform correlation analysis on the sound verification pass rate and the environmental noise decibel range label to obtain the second correlation matrix of the sound verification pass rate affected by the environmental noise decibel range label.

[0022] Based on the first association matrix, the first association strength between facial recognition accuracy and the classification value of each ambient light intensity level is determined. If the first association strength is greater than the first preset threshold, the corresponding facial recognition accuracy and ambient light intensity level classification value are determined to be the main influencing factor pair for facial recognition.

[0023] Based on the second correlation matrix, the second correlation strength between the sound verification pass rate and each environmental noise decibel range label is determined. If the second correlation strength is greater than the second preset threshold, the corresponding sound verification pass rate and environmental noise decibel range label are determined as the main influencing factor pair for sound verification.

[0024] The main influencing factors of facial recognition and voice verification are combined to determine the target association matrix;

[0025] The influence paths of environmental factor changes are extracted from the target correlation matrix, and based on the influence paths of environmental factor changes, a clustering algorithm is used to group them to obtain a set of environmental condition classifications, including groups with insufficient light and normal noise.

[0026] Optionally, based on the low-light and normal-noise group, the accuracy change trend over a continuous time period is calculated using the sliding window method to obtain the performance change pattern sequence of biometrics in low-light environments, including:

[0027] The dataset of environmental factors affecting performance indicators corresponding to the group with insufficient light and normal noise is used as a subsequence dataset, and the fluctuation value of face recognition accuracy is determined based on the subsequence dataset.

[0028] If the fluctuation value exceeds the preset fluctuation threshold, the face recognition accuracy value is extracted window by window based on the subsequence dataset to obtain the face recognition accuracy subset.

[0029] Based on a subset of facial recognition accuracy, the sequence changes of accuracy values ​​within each sliding window are used to determine the subsequence of accuracy differences between adjacent windows;

[0030] Based on the direction and magnitude of the accuracy difference subsequence, the trend of accuracy change over a continuous time period is determined.

[0031] Based on the accuracy change trend, the performance change pattern sequence of biometrics in low-light environments was determined.

[0032] Optionally, based on the performance change pattern sequence, a classification algorithm is used for rule mining to generate a core set of adaptation rules that includes branch judgments based on environmental conditions, including:

[0033] Based on the performance change pattern sequence, the adjustment parameter range sequence for facial image texture enhancement in low-light scenes and the dynamic correction coefficient sequence for sound signal feature weights are determined.

[0034] Based on the adjusted parameter range sequence and the dynamic correction coefficient sequence, a classification algorithm is used to perform rule mining on the two types of sequence data to obtain a set of candidate adaptation rules covering multiple environmental conditions and the corresponding rule coverage degree.

[0035] If the rule coverage reaches the preset ratio, then based on the candidate adaptive rule set, the branch conditions for the two dimensions of ambient light intensity and noise level are sorted out to generate the branch judgment structure of the environmental conditions.

[0036] Based on the branch-based decision structure, a set of core adaptation rules containing environmental condition branch decisions is obtained by filling in the conditions using preset environmental condition thresholds.

[0037] Optionally, adaptive adjustment parameters corresponding to each environmental condition classification set are determined based on the core adaptation rule set, and these adaptive adjustment parameters are applied in a simulated biometric authentication environment to obtain an optimized performance index dataset, including:

[0038] The adaptive adjustment parameters corresponding to each environmental condition classification set are determined based on the core adaptation rule set;

[0039] Based on adaptive parameter adjustment, the corresponding parameter application operation is performed in a simulated biometric authentication environment to obtain the first performance index dataset;

[0040] Based on the first performance index dataset, support vector machine is used to classify the indexes under different environmental conditions to obtain a set of classification labels;

[0041] Based on the set of classification labels, extract the subset of adjustment parameters corresponding to each classification label from the adaptive adjustment parameters;

[0042] Based on the adjusted parameter subset, the parameters are applied again in a simulated biometric authentication environment to obtain a second performance index dataset.

[0043] The dataset with the lower false recognition rate between the first and second performance metric datasets is used as the optimized performance metric dataset.

[0044] Optionally, the optimized performance index dataset and the environmental factor impact performance index dataset are merged, and the merged dataset is grouped using a clustering algorithm to obtain a set of biometric authentication output rules, including:

[0045] The optimized performance index dataset and the environmental factor impact performance index dataset are merged to obtain the merged dataset;

[0046] Based on the merged dataset, a clustering algorithm is used to group the data, resulting in preliminary grouping data.

[0047] Based on the preliminary grouping results data, the first central location data of each group was determined, and based on the environmental condition classification set, the second central location data of the environmental condition classification set was determined.

[0048] By comparing the first center location data with the second center location data, the center distance data between each group is obtained;

[0049] Based on the center distance data, the average center distance of all groups is calculated to obtain the average calculation result;

[0050] If the average value is lower than the preset average threshold, the preliminary grouping result data is determined to meet expectations, and the preliminary grouping result data is determined to be the set of biometric authentication output rules.

[0051] In a first aspect, the present invention provides a security authentication system based on multimodal biometric fusion, comprising:

[0052] The dataset generation module is used to obtain the accuracy of facial recognition and the pass rate of voice verification related to biometric authentication, and at the same time, obtain the classification value of ambient light intensity level and the label of ambient noise decibel range at the corresponding time, so as to obtain the dataset of performance indicators of environmental factors.

[0053] The grouping module is used to group the dataset of performance indicators affected by environmental factors using a clustering algorithm to obtain a set of environmental condition classifications, including groups with insufficient light and normal noise.

[0054] The regular sequence generation module is used to calculate the accuracy change trend over a continuous time period based on a group with low light and normal noise, and obtain the performance change regular sequence of biometrics in low light environment by using the sliding window method.

[0055] The rule mining module is used to perform rule mining based on the performance change pattern sequence and a classification algorithm to generate a core adaptive rule set that includes branch judgments based on environmental conditions.

[0056] The application module is used to determine the adaptive adjustment parameters corresponding to each environmental condition classification set based on the core adaptive rule set, and apply the adaptive adjustment parameters in a simulated biometric authentication environment to obtain an optimized performance index dataset.

[0057] The merging module is used to merge the optimized performance index dataset with the environmental factor-affected performance index dataset. The merged dataset is then grouped using a clustering algorithm to obtain a set of biometric authentication output rules. This set of biometric authentication output rules is used to perform secure authentication operations based on biometrics.

[0058] This invention offers the following advantages: It provides a secure authentication method and system based on multimodal biometric fusion. Addressing the performance fluctuations of biometric systems under different environmental conditions, it integrates correlation analysis between facial recognition accuracy and voice verification pass rate with ambient light intensity and noise levels to construct a dataset of environmental factors affecting performance. Subsequently, a clustering algorithm is used to group this dataset, identifying a set of environmental condition classifications that includes groups with insufficient light and normal noise levels. For performance fluctuations under conditions of insufficient light and noise interference, a sliding window method is employed to analyze trends and generate a core set of adaptation rules. By simulating environmental verification and adjusting corresponding adaptive parameters, a set of biometric authentication output rules is ultimately formed. This establishes a comprehensive analytical framework that systematically monitors and analyzes various performance data during the authentication process. By leveraging this data, patterns in performance changes are discovered, ensuring authentication stability and accuracy in complex environments and significantly improving the adaptability and reliability of biometric technology. Attached Figure Description

[0059] Figure 1 This is one of the flowcharts of a security authentication method based on multimodal biometric fusion provided in an embodiment of the present invention;

[0060] Figure 2This is the second flowchart of a security authentication method based on multimodal biometric fusion provided in an embodiment of the present invention;

[0061] Figure 3 This is the third flowchart of a security authentication method based on multimodal biometric fusion provided in this embodiment of the invention;

[0062] Figure 4 This is the fourth flowchart of a security authentication method based on multimodal biometric fusion provided in this embodiment of the invention;

[0063] Figure 5 This is the fifth flowchart of a security authentication method based on multimodal biometric fusion provided in this embodiment of the invention;

[0064] Figure 6 This is the sixth flowchart of a security authentication method based on multimodal biometric fusion provided in this embodiment of the invention;

[0065] Figure 7 This is the seventh flowchart of a security authentication method based on multimodal biometric fusion provided in this embodiment of the invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Reference Figure 1 This invention provides a security authentication method based on multimodal biometric fusion, comprising the following steps:

[0068] Step 100: Obtain the facial recognition accuracy and voice verification pass rate related to biometric authentication, and at the same time obtain the classification value of the ambient light intensity level and the label of the ambient noise decibel range at the corresponding time to obtain the dataset of environmental factors affecting performance indicators.

[0069] In some embodiments, this step is used to correlate and synchronously collect the performance of the biometric authentication system with its specific operating environment, thereby constructing a structured dataset that can quantitatively and qualitatively reflect the intrinsic correlation between environment and performance. Traditional biometric authentication systems typically treat facial recognition and voice verification as independent identity determination modules. Their performance evaluation often focuses on improvements to the algorithm itself or the quality of user feature registration, while ignoring the systematic and regular impact of complex and changing environmental factors on authentication results in real-world application scenarios. First, the accuracy of facial recognition and the pass rate of voice verification related to biometric authentication are obtained, and key environmental parameters in two dimensions are captured simultaneously when authentication occurs, including a classification value for the ambient light intensity level for quantifying visual acquisition conditions, and a label for the ambient noise decibel range for quantifying auditory acquisition conditions.

[0070] By combining facial recognition accuracy and voice verification pass rate, along with classification values ​​for ambient light intensity and decibel range labels for ambient noise, a dataset of performance indicators influenced by environmental factors can be obtained. This dataset is essentially a mapping database that records the actual performance fluctuations of the multimodal biometric fusion authentication system under different combinations of environmental conditions. For example, this dataset may contain potential patterns such as a specific downward trend in facial recognition accuracy under low-light but normal-noise conditions, while the voice verification pass rate remains relatively stable; or optimal and stable performance in both modalities under sufficient light and low-noise conditions.

[0071] Step 200: Based on the dataset of performance indicators affected by environmental factors, a clustering algorithm is used to group the data to obtain a set of environmental condition classifications, including groups with insufficient light and normal noise.

[0072] In some embodiments, after constructing a dataset of environmental factors affecting performance indicators related to environment and performance, an unsupervised clustering algorithm is used to intelligently sort and summarize the complex and multidimensional dataset of environmental factors affecting performance indicators, so as to divide the seemingly disordered and continuous environmental state space into several environmental scene categories with different characteristics and practical analytical significance, thereby obtaining different environmental condition classification sets. Among them, the environmental condition classification set includes a group with insufficient light and normal noise, and may also include a group with normal light and noise interference, a group with dual environmental interference, and a group with sufficient light and low noise.

[0073] The advantage of clustering algorithms lies in their ability to discover and aggregate frequently occurring or similar combinations of environmental conditions based on the inherent similarity and distribution density of the data itself (i.e., the numerical combination of the two environmental dimensions of light and noise) without predefined rules. Each group in the resulting environmental condition classification set represents a typical environmental pattern that repeatedly occurs during the authentication process and may have a similar impact on system performance. Applying clustering algorithms can surpass the fixed threshold judgment methods that may be used in traditional techniques, adapt to the continuity of interwoven factors in real-world environments, and capture naturally formed boundaries in the data.

[0074] Furthermore, this step breaks down the general issue of environmental impact on performance into several specific sub-problems. For example, how does system performance evolve over time in a specific environment with insufficient light and normal noise, or which mode's performance degradation dominates in a scenario with dual environmental interference? By categorizing the dataset into different specific environmental scenarios, subsequent analysis can be conducted within each homogeneous scenario, thereby eliminating the interference of other variables and more reliably revealing the intrinsic correlation between single or combined environmental factors and biometric authentication performance.

[0075] Step 300: Based on the group with insufficient light and normal noise, the accuracy change trend in a continuous time period is calculated using the sliding window method to obtain the performance change pattern sequence of biometrics under insufficient light conditions.

[0076] In some embodiments, after clustering environmental conditions and separating scenes with insufficient light and normal noise, this step delves into the specific scene to uncover potential dynamic patterns in biometric performance over time. By introducing a time-dimensional analytical perspective, a sliding window method is used to scan and extract trends from the continuous time series authentication data within the group with insufficient light and normal noise, thus obtaining a sequence of biometric performance change patterns under insufficient light conditions.

[0077] The sliding window method simulates an observation window that slides along the time axis. For the low-light, normal-noise group, it calculates the local variation characteristics of facial recognition accuracy and voice verification pass rate within consecutive, overlapping time segments. For example, it uses five authentication data points as a window, sliding one data point at a time. These local variation characteristics can be the direction of mean shift, fluctuation amplitude, or stability indicators. The sliding window method captures the gradual changes, periodic fluctuations, or critical turning points that performance indicators for the low-light, normal-noise group may exhibit under specific environmental constraints. For example, in a continuously low-light environment, facial recognition accuracy does not simply remain at a constant low level, but may show a curve of initial decline followed by a slow recovery as the user continues to be exposed to the environment or the device is fine-tuned, or reveal a correlation between the decrease in accuracy and the number of consecutive authentication attempts.

[0078] This yields a performance variation sequence, a series of high-dimensional features characterizing how system performance evolves over time or a sequence of authentication events under a fixed environmental condition of insufficient light and normal noise. This process transforms the general phenomenon of performance degradation into specific, describable dynamic behavioral patterns. For example, in a continuously dim environment, the accuracy of facial authentication decreases slowly in the first three authentication attempts, then tends to stabilize.

[0079] Step 400: Based on the performance change pattern sequence, a classification algorithm is used to perform rule mining to generate a set of core adaptive rules that include branch judgments based on environmental conditions;

[0080] In some embodiments, after obtaining the set of dynamic performance evolution patterns (i.e., the performance change pattern sequence) under a specific environment, the performance change patterns, represented as sequences or patterns, are transformed into structured decision logic that the system can directly understand and execute, i.e., the core adaptation rule set. Specifically, a classification algorithm is used to perform pattern recognition and rule induction on this high-dimensional feature object of the performance change pattern sequence. By analyzing the key features contained in the performance change pattern sequence, such as the slope of the trend, the inflection point of fluctuation, and the performance decay rate within a specific time window, different performance evolution scenarios are learned and defined. For each identified scenario, there is a specific adaptive strategy. Therefore, the core adaptation rule set obtained by rule mining is essentially a data-driven mapping relationship between the dynamic trajectory of performance changes and the optimal system adjustment actions.

[0081] The generated core adaptation rule set is a context-aware decision tree or rule network containing environmental condition branch judgments. One type of environmental condition branch judgment logic structure of this core adaptation rule set can be expressed as follows: when in a low-light and normal-noise group, if the detected performance change sequence shows a pattern of continuous and rapid decline in facial accuracy in the first N attempts, a first-type parameter adjustment strategy is triggered; if it shows a pattern of low-level oscillation in accuracy but stable sound throughput, a second-type fusion weight adjustment strategy is triggered. Through this step, the biometric authentication system can intelligently select the most suitable authentication strength threshold, modality fusion weight, or backup verification mechanism based on real-time analysis of performance evolution patterns, rather than just the absolute value at the current moment, thus providing intelligent decision support for achieving stable and reliable environmental adaptive authentication.

[0082] Step 500: Determine the adaptive adjustment parameters corresponding to each environmental condition classification set based on the core adaptive rule set, and apply the adaptive adjustment parameters in the simulated biometric authentication environment to obtain the optimized performance index dataset;

[0083] In some embodiments, after generating a core set of adaptive rules containing decision-making logic, the core task of this step is to achieve the mapping from qualitative rules to quantitative parameters, moving from the logical decision-making stage to the engineering implementation stage of converting logical decision-making rules into specific numerical values ​​and verifying their effectiveness. First, adaptive adjustment parameters are determined based on the core set of adaptive rules; this is a process of parameterization and concretization. Each rule in the core set of adaptive rules, containing conditional branches, points to an adaptive strategy that is quantified into an operational instruction that the system can directly execute. For example, if a rule determines that the weight of a voice modality needs to be enhanced, the specific percentage or absolute value of the weight increase from the base value is determined; if a rule suggests relaxing the face matching threshold to reduce the false rejection rate, the specific numerical range of threshold relaxation needs to be calculated. Adaptive adjustment parameters can be assigned, and preliminarily calculated, sets of adjustment parameters for each environmental condition classification and its different performance evolution modes, based on the historical performance data distribution associated with the corresponding rule, through statistical optimization, cost function minimization, or simulation-based parameter tuning.

[0084] Subsequently, on the platform of this simulated biometric authentication environment, the determined adaptive adjustment parameters are substituted into the corresponding historical environmental sequences and biometric data to run simulated authentication calculations. This evaluates the changes in the accuracy or pass rate of simulated authentication after adopting the adaptive adjustment parameters, thereby obtaining an optimized performance index dataset. The process of simulating the biometric authentication environment is essentially a large-scale, data-driven simulation and verification of the effects of the mined rules and adaptive adjustment parameters. The resulting optimized performance index dataset includes the expected performance improvement results achieved under each environmental condition after applying the adaptive adjustment parameters.

[0085] Step 600: Merge the optimized performance index dataset and the environmental factor impact performance index dataset, and group the merged dataset using a clustering algorithm to obtain a set of biometric authentication output rules. The set of biometric authentication output rules is used to perform biometric security authentication operations.

[0086] In some embodiments, merging the optimized performance metric dataset with the original environmental factor-affected performance metric dataset is not a simple data overlay, but rather a knowledge fusion. The environmental factor-affected performance metric dataset records the system's actual performance under no-intervention or default parameter conditions, reflecting the natural impact of the environment on performance; while the optimized performance metric dataset records the expected performance after applying the core set of adaptation rules and adaptive adjustment parameters under the same environmental input. By merging the optimized performance metric dataset and the environmental factor-affected performance metric dataset, a richer knowledge base is formed, namely the merged dataset. Each data point in the merged dataset is simultaneously associated with three pieces of information: environmental conditions, baseline performance, and optimized performance. Therefore, by merging the dataset, we can understand how performance typically behaves in a certain environment, and more importantly, what adjustments can improve it to what level.

[0087] For this merged dataset, grouping using a clustering algorithm has a fundamentally different purpose than the first clustering. The first clustering simply divides the input space based on environmental features (light, noise). The second clustering, however, because its clustering objects include the merged dataset of environment, performance, and optimization actions, can group data based on stable correlation patterns between environmental states and optimal response strategies. Through the clustering algorithm, data samples with specific environmental characteristics and original performance patterns are identified, corresponding to the same type of optimization parameters and similar performance improvement results. These are then grouped into a decision cluster, resulting in a set of biometric authentication output rules. For example, a cluster might represent a scenario belonging to environment type A and exhibiting a performance decline trend B; adjusting parameter group C can stably improve the overall authentication success rate to within the target range D.

[0088] The final set of biometric authentication output rules is a set of directly executable instructions that correspond to conditions and actions. In actual authentication, real-time sensing of ambient light or noise maps to a specific environmental condition category; further, real-time analysis of performance time-series trends allows for matching specific output rules, i.e., the biometric authentication output rule set. This set of biometric authentication output rules indicates the specific fusion weights, confidence thresholds, or authentication processes that the system should employ. Thus, the entire method completes a process from learning from historical data to verification in a simulation environment, ultimately solidifying into intelligent decision-making in a real-time system for executing secure biometric authentication operations. This evolves the biometric authentication system into a dynamic, environmentally-aware, and adaptive intelligent entity.

[0089] This invention provides a secure authentication method and system based on multimodal biometric fusion, which realizes the construction of a comprehensive analysis framework to monitor and sort out various performance data in the authentication process, and discover the pattern of performance changes through these data, ensuring the stability and accuracy of authentication in complex environments, and significantly improving the adaptability and reliability of biometric technology.

[0090] In one embodiment, please refer to Figure 2 This study obtains the accuracy of facial recognition and the pass rate of voice verification related to biometric authentication. It also obtains the classification values ​​of ambient light intensity levels and the decibel range labels of ambient noise at corresponding times, resulting in a dataset of performance indicators related to environmental factors, including:

[0091] Step 101: Based on the record entries in the log file of the biometric authentication system, obtain the facial recognition accuracy and voice verification pass rate related to biometric authentication to obtain performance index sequence data;

[0092] Step 102: Based on the recorded entries, extract the classification value of the ambient light intensity level and the label of the ambient noise decibel range at the corresponding time to obtain the environmental factor label sequence;

[0093] Step 103: If the recorded item includes facial recognition accuracy, then associate and bind the ambient light intensity level classification value of the corresponding recorded item with the facial occlusion object category label to generate a facial recognition environmental impact sample pair.

[0094] Step 104: If the record item contains the sound verification pass rate, then associate and bind the environmental noise decibel range label of the corresponding record item with the sound-affected object category label to generate a sound verification environmental impact sample pair.

[0095] Step 105: Based on the performance index sequence data, environmental factor label sequence, facial recognition environmental impact sample pairs, and sound verification environmental impact sample pairs, determine the environmental factor impact performance index dataset.

[0096] In some embodiments, firstly, the performance index sequence and environmental factor label sequence of time synchronization are extracted from the log file of the biometric authentication system. Specifically, this includes obtaining the facial recognition accuracy and voice verification pass rate related to biometric authentication to obtain performance index sequence data; and extracting the ambient light intensity level classification value and ambient noise decibel range label at the corresponding time to obtain the environmental factor label sequence. Then, the environmental state is associated with the specific physical object or interference source category that caused the state. That is, the ambient light intensity level classification value of the corresponding record entry is associated with the facial occlusion object category label to generate facial recognition environmental impact sample pairs, and the ambient noise decibel range label of the corresponding record entry is associated with the sound impact object category label to generate sound verification environmental impact sample pairs. For example, the facial recognition environmental impact sample not only records the insufficient light intensity level classification value, but also further binds the current facial occlusion object category label (such as hat brim occlusion, mask, strong backlight silhouette, etc.) to it to generate facial recognition environmental impact sample pairs. Similarly, for sound verification, the ambient noise decibel range label is associated with the specific sound impact object category label (such as continuous traffic background noise, intermittent keyboard typing, reverberation from multi-person conversation, etc.). This approach essentially involves adding semantic annotations based on scene understanding to the original environmental sensor data.

[0097] Finally, based on performance index sequence data, environmental factor label sequence, facial recognition environmental impact sample pairs, and sound verification environmental impact sample pairs, the data is integrated to complete the transformation from multi-dimensional raw data to advanced analysis dataset. The final result is the environmental factor impact performance index dataset, which is a multi-dimensional structured data collection containing performance characteristics, macro-environmental parameters, and micro-interference source semantic information.

[0098] For example, the process of obtaining performance index data from the log records of a biometric authentication system and constructing a dataset on the impact of environmental factors can be achieved through the following specific methods. First, assuming the system logs are stored in a distributed database and contain data on facial recognition accuracy and voice verification pass rate, the daily averages over the past 30 days are extracted. For example, the facial recognition accuracy is 92.5%, and the voice verification pass rate is 88.3%. Simultaneously, environmental data is collected. The ambient light intensity at the corresponding moment is obtained using a sensor interface API and classified into three levels: low (0-50 lux), medium (51-200 lux), and high (201+ lux). Assuming the light intensity on a certain day is 120 lux, it is classified as medium. The ambient noise decibel range is collected using a sound pressure sensor and labeled as quiet (0-40 dB), normal (41-70 dB), and noisy (71+ dB). At a certain moment, the noise level is 55 dB and it is labeled as normal. The category of facial occlusion objects is identified using image processing algorithms to determine whether the occlusion is a mask, hat, or no occlusion. Assuming a certain detection result is a mask, it is labeled as "mask". Subsequently, performance metrics and environmental data were aligned with timestamps to construct an initial dataset, generating a table containing time, performance metric values, and environmental factor labels. For example, a record might be labeled "2023-10-15 14:00, 92.5%, 1.8%, 88.3%, Medium lighting, General noise, Mask". Finally, a preliminary analysis of the impact of environmental factors on performance was conducted. Correlation analysis algorithms such as the Pearson correlation coefficient were used. The correlation coefficient between light intensity classification and facial recognition accuracy was -0.65, indicating that accuracy decreases with stronger light. The correlation coefficient between noise and sound verification pass rate was -0.72, showing that increased noise reduces the sound verification pass rate. These findings form the preliminary conclusion regarding the impact of environmental factors on performance metrics.

[0099] In this embodiment, by identifying the type of interference source, it is possible to distinguish environmental conditions that are similar in appearance but different in cause. For example, even in low light conditions, the light spectrum and characteristic influence patterns at night without lighting and under the shade of trees during the day may be different, and differentiated optimal adaptation parameters can be derived to achieve more refined scene adaptation.

[0100] In one embodiment, please refer to Figure 3 Based on a dataset of environmental factors affecting performance indicators, a clustering algorithm was used to group the data, resulting in a set of environmental condition classifications including groups with insufficient light and normal noise levels.

[0101] Step 201: Based on the dataset of performance indicators affected by environmental factors, a classification algorithm is used to perform correlation analysis on the classification values ​​of facial recognition accuracy and ambient light intensity level to obtain the first correlation matrix of the influence of facial recognition accuracy on the classification values ​​of ambient light intensity level.

[0102] Step 202: Based on the dataset of performance indicators affected by environmental factors, a classification algorithm is used to perform correlation analysis on the sound verification pass rate and the environmental noise decibel range label to obtain the second correlation matrix of the influence of the sound verification pass rate on the environmental noise decibel range label.

[0103] Step 203: Based on the first association matrix, determine the first association strength between the face recognition accuracy and the classification values ​​of each ambient light intensity level. If the first association strength is greater than the first preset threshold, determine the corresponding face recognition accuracy and ambient light intensity level classification values ​​as the main influencing factors of face recognition.

[0104] Step 204: Determine the second correlation strength between the sound verification pass rate and each environmental noise decibel range label based on the second correlation matrix. If the second correlation strength is greater than the second preset threshold, then determine the corresponding sound verification pass rate and environmental noise decibel range label as the main influencing factor pair for sound verification.

[0105] Step 205: Combine the main influencing factors of facial recognition and the main influencing factors of voice verification to determine the target association matrix;

[0106] Step 206: Extract the influence paths of environmental factor changes from the target association matrix, and based on the influence paths of environmental factor changes, use a clustering algorithm to group them to obtain an environmental condition classification set including groups with insufficient light and normal noise.

[0107] In some embodiments, quantitative association analysis and causal path mining are introduced to replace or enhance simple statistical clustering, thereby enabling environmental grouping to be based on more solid evidence of causality rather than relying solely on the superficial similarity of data distribution.

[0108] This step does not treat facial recognition accuracy and voice verification pass rate as simply lumped together with environmental factors. Instead, it first performs a single-modal, isolated correlation analysis, constructing a first correlation matrix for the influence of ambient light intensity level classification values ​​on facial recognition accuracy, and a second correlation matrix for the influence of ambient noise decibel range labels on voice verification pass rate. These two correlation matrices are constructed using classification algorithms. Essentially, these matrices quantify the influence pattern and degree of each specific light level classification value on facial accuracy, and the influence pattern and degree of each noise range label on voice pass rate. This step eliminates cross-interference between the two modalities, clarifying the relationship between each biometric modality and its most significant environmental interference source. The classification algorithm can be a decision tree, rule induction, or correlation analysis model.

[0109] Then, by calculating the correlation strength and comparing it with a preset threshold, the system can automatically identify those environmental and performance relationship pairs that truly have a decisive impact, i.e., the main influencing factor pairs, including the main influencing factor pairs for facial recognition and voice verification. This process filters out a large number of correlations that exist but have a weak or accidental impact, ensuring that subsequent analysis focuses on the core contradictions. For example, it may find that moderate noise in a specific frequency band has a strong correlation with voice verification, while the correlation of uniform high-decibel white noise is relatively weak, thus making an accurate distinction.

[0110] The target association matrix is ​​a fused view of the two independent analysis results of face and voice. Therefore, the influence paths of environmental factor changes can be extracted from the target association matrix. For example, the environmental factors and performance indicators contained in the target association matrix can be identified, and the influence paths of environmental factor changes can be identified. This means that the system no longer views the environmental state statically, but dynamically analyzes the chain changes in bimodal performance when light changes from type A to type B and noise changes from label X to label Y. Among them, the influence paths of environmental factor changes describe the typical trajectory of joint performance evolution caused by changes in environmental combinations. Based on this dynamic change path of the influence paths of environmental factor changes, the final clustering and grouping are performed to obtain a set of environmental condition classifications. Each environmental condition classification combination not only represents a combination of environmental states, but also embeds the known and quantified influence mechanism and expected evolution direction of environmental factors on bimodal performance under that state.

[0111] For example, firstly, based on the predefined causal relationships of environmental factors in the association matrix, such as light intensity → work efficiency, noise decibels → distraction, and the interaction between light and noise → overall interference level, influence paths are extracted and a set of paths is selected with environmental light intensity level classification values ​​and environmental noise decibel range labels as the core. Next, each record in the dataset is divided into four levels according to light intensity: less than 200 lux is marked as insufficient light, 200 to 500 lux as low light, 500 to 1000 lux as normal light, and greater than 1000 lux as sufficient light. Simultaneously, noise decibels are divided into ranges: less than 50 dB as low noise, 50 to 65 dB as normal noise, 65 to 80 dB as noise interference, and greater than 80 dB as severe. Interference was eliminated; then, the K-means clustering algorithm was used, with the median value of light intensity (e.g., 100, 350, 750, 1500) and the median value of noise (e.g., 25, 57.5, 72.5, 90) as two-dimensional feature vectors to group all samples. The number of clusters was set to 4, and Euclidean distance was used as the similarity measure. Iterative optimization was performed until the cluster center shift was less than 0.01. After clustering, four groups were obtained. The first group with a center point close to (100, 57.5) was defined as a group with insufficient light and normal noise. The second group with a center point close to (750, 72.5) was defined as a group with normal light and noise interference. The third group with a center point close to (350, 90) was defined as a group with dual environmental interference. The fourth group with a center point close to (1500, 25) was defined as a group with sufficient light and low noise. The first preset threshold is used to determine whether the correlation between facial recognition accuracy and the classification value of ambient light intensity level is strong enough. If it exceeds the threshold, the relationship pair is marked as a "major influencing factor pair for facial recognition". It is usually preset based on domain knowledge, historical data distribution, or cross-validation methods to filter strong correlations. The second preset threshold is used to determine whether the correlation between sound verification pass rate and ambient noise decibel range label is strong enough. If it exceeds the threshold, it is marked as a "major influencing factor pair for sound verification".

[0112] In this embodiment, by analyzing the impact path, it is possible to understand and predict to some extent the potential impact of new environmental combinations that have never appeared in historical data, thereby enhancing the system's adaptability to complex and unknown environmental conditions and thus strengthening the system's inference and generalization capabilities.

[0113] In one embodiment, please refer to Figure 4 Based on the low-light and normal-noise group, the accuracy change trend over a continuous time period was calculated using the sliding window method, resulting in a sequence of biometric performance change patterns in low-light environments, including:

[0114] Step 301: Take the environmental factor impact performance index dataset corresponding to the group with insufficient light and normal noise as a subsequence dataset, and determine the fluctuation value of face recognition accuracy based on the subsequence dataset.

[0115] Step 302: If the fluctuation value exceeds the preset fluctuation threshold, then based on the subsequence dataset, the face recognition accuracy value is extracted window by window using the sliding window method to obtain the face recognition accuracy subset.

[0116] Step 303: Based on the facial recognition accuracy subset, determine the accuracy difference subsequence between adjacent windows by analyzing the sequence changes of accuracy values ​​within each sliding window.

[0117] Step 304: Based on the direction and magnitude of the accuracy difference subsequence, determine the accuracy change trend over a continuous time period;

[0118] Step 305: Based on the accuracy change trend, determine the performance change pattern sequence of biometrics in low light conditions.

[0119] In some embodiments, a necessity assessment is first performed. The dataset representing the environmental factors affecting performance indicators in the low-light and normal-noise group is used as a subsequence dataset to determine the overall fluctuation value of facial recognition accuracy within this subsequence dataset. Only when the fluctuation value exceeds a preset fluctuation threshold, indicating significant performance instability or potential analyzable patterns in that environment, will subsequent, more refined time-series analysis be initiated. This reflects the idea of ​​efficient resource allocation, avoiding unnecessary complex calculations on stable data. The preset fluctuation threshold is used to determine whether the fluctuation level of facial recognition accuracy in the low-light and normal-noise group is significant. Subsequent sliding window time-series analysis is only initiated when the fluctuation value exceeds this threshold, avoiding invalid calculations on stable data. This threshold can be set by analyzing the fluctuation levels of historical data or in conjunction with business requirements.

[0120] When the fluctuation value exceeds a preset fluctuation threshold, the analysis process is triggered. Using a sliding window method, the accuracy values ​​within each window are extracted to form a subset, resulting in a subset of facial recognition accuracy, which captures the local performance level. Furthermore, by calculating the accuracy difference sequence between adjacent windows, the system identifies how performance changes, shifting from first-order statistics to second-order features characterizing the speed and direction of change. Finally, based on the accuracy difference sequence, the direction (positive / negative) and magnitude (size) of these differences are used to comprehensively determine the accuracy change trend over a continuous time period, such as a continuous slow decline, a V-shaped recovery after an initial drop followed by a rise, or sharp oscillations at a low level.

[0121] Finally, the accuracy change trends were integrated, completing the transformation from raw data to high-level knowledge, and ultimately determining the performance change pattern sequence. This determined performance change pattern sequence is an organic integration of the aforementioned multi-level features. This performance change pattern sequence is a structured sequence description containing time segments, change gradients, and evolution patterns. For example, it can be described as exhibiting a moderate downward trend in the initial stage, continuing for three windows before entering a period of slight fluctuation and plateau.

[0122] For example, firstly, facial recognition accuracy sequence data is extracted for each environmental condition classification set: sufficient light and normal noise, sufficient light and abnormal noise, insufficient light and normal noise, and insufficient light and abnormal noise. Taking the group with insufficient light and normal noise as an example, its accuracy data contains 1000 samples. The fluctuation level is calculated using the standard deviation, resulting in σ=7.82% and the coefficient of variation CV=9.14%. The preset fluctuation threshold is set to 6.5%. Since 7.82%>6.5%, the subsequent analysis process is triggered. Next, the accuracy data of this group is arranged in ascending order according to the collection timestamp to form a time series with a time interval of 5 seconds. Then, a sliding window of length 120, corresponding to a continuous 10-minute period, is used to calculate the linear regression slope within each window as the trend change rate by sliding point by point on the time series. The specific algorithm is to use least squares on the data points within the window. A straight line was fitted using a method to obtain the slope, which represents the accuracy trend value for that 10-minute period. The sliding step size was set to 12 data points, or 1 minute, and the trend sequence was calculated sequentially. The trend sequence was then smoothed using the Exponentially Weighted Moving Average (EWMA) method with a smoothing coefficient α=0.15 to obtain a smoothed trend sequence, which characterizes the change pattern of biometric performance in low-light environments. Finally, analysis of the smoothed trend sequence revealed that when the ambient light level remained below 20 lux for more than 8 minutes, the accuracy decreased at an average rate of -0.38% / minute. The rate of decrease accelerated to -0.71% / minute in the light level range of 15 lux to 5 lux. When the light level recovered to above 30 lux, the accuracy rebounded rapidly at a rate of +0.52% / minute, thus forming a complete sequence of the change pattern of biometric performance in low-light environments.

[0123] In this embodiment, difference sequence analysis can capture the instantaneous rate and acceleration of performance changes, rather than just the final state, thereby predicting the trend of performance degradation earlier. The output performance change pattern sequence is a feature expression that includes temporal logic, enabling subsequent rule mining algorithms to generate intelligent rules containing temporal conditions.

[0124] In one embodiment, please refer to Figure 5 Based on the performance change pattern sequence, a classification algorithm is used for rule mining to generate a core set of adaptation rules that includes branch judgments based on environmental conditions, including:

[0125] Step 401: Based on the performance change pattern sequence, determine the adjustment parameter range sequence for facial image texture enhancement in low-light scenarios, and the dynamic correction coefficient sequence for sound signal feature weights.

[0126] Step 402: Based on the adjusted parameter range sequence and the dynamic correction coefficient sequence, a classification algorithm is used to perform rule mining on the two types of sequence data to obtain a set of candidate adaptation rules covering multiple environmental conditions and the corresponding rule coverage degree.

[0127] Step 403: If the rule coverage reaches the preset ratio, then based on the candidate adaptation rule set, the branch conditions for the two dimensions of ambient light intensity and noise level are sorted out to generate the branch judgment structure of the environmental conditions.

[0128] Step 404: Based on the branch judgment structure, the conditions are filled using preset environmental condition thresholds to obtain a core adaptation rule set containing environmental condition branch judgments.

[0129] In some embodiments, firstly, based on the performance change pattern sequence obtained from the prior analysis, two specific and operable adjustment dimensions and their dynamic ranges are derived, including: a sequence of adjustment parameters for facial image texture enhancement in the visual modality, and a sequence of dynamic correction coefficients for the sound signal feature weights in the auditory modality. This means that the system not only knows that performance is declining, but also knows more precisely that it should enhance image texture to compensate for detail loss caused by insufficient lighting, and adjust sound feature weights to optimize judgment in a fixed noise background, thereby responding to environmental changes, and providing a quantified adjustment range for each pattern state.

[0130] Next, the rule knowledge is automatically extracted and initially screened for quality. Classification algorithms are used to mine the two types of parameter sequence data to discover stable correspondences between different performance change patterns and optimal parameter adjustment combinations, thereby generating a candidate set of adaptation rules and their corresponding rule coverage. Rule coverage is a quality indicator of the candidate set of adaptation rules, used to evaluate whether this set of preliminary rules is sufficient to explain and handle most of the observed scenarios.

[0131] When the rule coverage reaches a preset ratio, indicating that the coverage of the candidate adaptive rule set meets the standard, the rule organization and solidification stage begins. Based on the candidate adaptive rule set, a branch judgment structure for environmental conditions is organized, which essentially constructs a decision logic framework for rule application. This structure is then populated with preset environmental condition thresholds, transforming this decision logic framework into a specific set of core adaptive rules with clearly defined numerical judgment conditions. This step ensures that the rules have clear, unambiguous, and quickly matching trigger conditions, completing the transformation from knowledge patterns mined from data to executable conditions and action statements in the software system. The preset ratio is used to evaluate whether the "rule coverage" of the candidate adaptive rule set meets the standard. Only when this ratio is reached does the rule organization and solidification stage begin, preventing overfitting. This ratio is usually set based on experience to ensure that the rules cover most observed scenarios. The preset environmental condition thresholds are used to populate the specific numerical conditions in the branch judgment structure, giving the rules clear trigger boundaries. These thresholds can be predefined based on the historical distribution of environmental parameters or industry standards.

[0132] In this embodiment, a rule generation coverage verification mechanism is introduced. By evaluating the rule coverage, it is ensured that the final rule set is not an overfitted product for a few special cases, but a robust solution that can broadly cover various identified scenarios, thereby improving overall reliability.

[0133] In one embodiment, please refer to Figure 6 Based on the core adaptation rule set, adaptive adjustment parameters corresponding to each environmental condition classification set are determined, and these parameters are applied in a simulated biometric authentication environment to obtain an optimized performance index dataset, including:

[0134] Step 501: Determine the adaptive adjustment parameters corresponding to each environmental condition classification set based on the core adaptation rule set;

[0135] Step 502: Based on adaptive adjustment parameters, perform the corresponding parameter application operation in a simulated biometric authentication environment to obtain the first performance index dataset;

[0136] Step 503: Based on the first performance index dataset, support vector machine is used to classify the indexes under different environmental conditions to obtain a set of classification labels;

[0137] Step 504: Based on the set of classification labels, extract the subset of adjustment parameters corresponding to each classification label from the adaptive adjustment parameters;

[0138] Step 505: Based on the adjusted parameter subset, the parameters are applied again in a simulated biometric authentication environment to obtain the second performance index dataset;

[0139] Step 506: Select the dataset with the lower false recognition rate from the first performance metric dataset and the second performance metric dataset as the optimized performance metric dataset.

[0140] In some embodiments, firstly, adaptive adjustment parameters are determined based on the core rule set. The generated adaptive adjustment parameters are then fully applied in a simulation environment to obtain a first performance index dataset. This first performance index dataset records the preliminary effects of these parameters in various environments and can be used to verify the effectiveness of the rules.

[0141] Subsequently, the Support Vector Machine (SVM) classification algorithm was introduced to perform in-depth analysis on the first performance indicator dataset. Based on the patterns of the performance indicators (such as combinations of accuracy, pass rate, and false recognition rate), the SVM algorithm automatically categorizes the application results under different environmental conditions into different label sets. These label sets essentially represent the type of effect of the parameter application, for example, significantly improved effect, stable effect, poor effect, or even worsening effect.

[0142] Next, performance-based parameter optimization is performed. Based on the performance classification labels generated by the support vector machine algorithm (i.e., the set of classification labels), the subset of adjusted parameters most closely associated with significant performance improvements is extracted. Subsequently, this subset of adjusted parameters is applied again in a simulated biometric authentication environment to conduct a second round of validation, generating a second performance metric dataset. This realizes the reinforcement learning process, learning which parameter combinations are more effective from the first attempt and concentrating resources to validate and confirm these advantageous combinations a second time. Finally, a final selection decision is made. By comparing the false recognition rates of the two datasets, the dataset with the lower false recognition rate is established as the final optimized performance metric dataset. This ensures that the knowledge base provided for generating the final output rules in subsequent stages is a validated dataset selected with the most stringent security standards.

[0143] In this embodiment, the application effect is automatically classified by the support vector machine algorithm. It can intelligently identify the truly efficient parts in the rule set and focus on optimizing the parameters of these parts, thereby realizing the self-refinement and evolution of the rules.

[0144] In one embodiment, please refer to Figure 7 The optimized performance index dataset and the environmental factor-affected performance index dataset are merged. A clustering algorithm is then used to group the merged datasets to obtain a set of biometric authentication output rules, including:

[0145] Step 601: Merge the optimized performance index dataset with the environmental factors affecting performance index dataset to obtain a merged dataset;

[0146] Step 602: Based on the merged dataset, a clustering algorithm is used to group the datasets to obtain preliminary grouping results.

[0147] Step 603: Based on the preliminary grouping results data, determine the first central location data of each group, and based on the environmental condition classification set, determine the second central location data of the environmental condition classification set.

[0148] Step 604: Compare the first center location data with the second center location data to obtain the center distance data between each group;

[0149] Step 605: Based on the center distance data, calculate the average center distance of all groups to obtain the average calculation result;

[0150] Step 606: If the average value is lower than the preset average threshold, the preliminary grouping result data is determined to meet expectations, and the preliminary grouping result data is determined to be the biometric authentication output rule set.

[0151] In some embodiments, firstly, the optimized performance index dataset is merged with the original environmental factor impact performance index dataset to form an enhanced knowledge base that simultaneously includes both the natural state and the optimized state, i.e., a merged dataset. Then, the merged dataset is clustered to obtain preliminary grouping results. This preliminary grouping results represent the understanding of the relationship pattern between environment and performance based on all experience (including intervention effects) and forms the basis for generating new rules.

[0152] Next, the center positions are compared and verified. First, the first center position data of each group in the preliminary grouping results and the second center position data of the previously determined environmental condition classification set are calculated. The difference between the first and second center position data is calculated to obtain the center distance data, and its average value is calculated. This average value essentially measures the degree of geometric deviation between the old and new cognitive frameworks for defining different environmental scenarios in the feature space. If the environmental scenario center defined by the new rule deviates very little from the scene center in the original cognition, that is, the average value is lower than the preset mean threshold, it proves that the optimization process is robust and has not fundamentally changed the system's basic classification of the environment, but only enhanced its performance. Only when the average value is lower than the preset mean threshold will the preliminary grouping result data be established as the final set of biometric authentication output rules, ensuring that the final output rules achieve performance optimization without deviating from the physical constraints of the real environment. The preset mean threshold is used to verify the consistency between the grouping results obtained from the second clustering (preliminary grouping result data) and the original environmental classification (environmental condition classification set). If the average distance between the centers of all groups is below the threshold, the optimization process is considered robust, and the new group can be used as the final output rule. It needs to be set through experiments or experience to ensure that the center deviation between the old and new groups is within an acceptable range.

[0153] In this embodiment, a preset mean threshold and consistency check are set for the authentication system, which means that the intelligent improvement of the system is achieved within a stable and verifiable framework, thereby outputting a set of authentication decision rules that are both intelligent and advanced as well as robust and reliable, thus improving the confidence and trustworthiness of the final rule set.

[0154] The following describes a security authentication system based on multimodal biometric fusion provided by the present invention. The security authentication system based on multimodal biometric fusion described below and the security authentication method based on multimodal biometric fusion described above can be referred to and correspond to each other.

[0155] This invention also provides a security authentication system based on multimodal biometric fusion, comprising:

[0156] The dataset generation module is used to obtain the accuracy of facial recognition and the pass rate of voice verification related to biometric authentication, and at the same time, obtain the classification value of ambient light intensity level and the label of ambient noise decibel range at the corresponding time, so as to obtain the dataset of performance indicators of environmental factors.

[0157] The grouping module is used to group the dataset of performance indicators affected by environmental factors using a clustering algorithm to obtain a set of environmental condition classifications, including groups with insufficient light and normal noise.

[0158] The regular sequence generation module is used to calculate the accuracy change trend over a continuous time period based on a group with low light and normal noise, and obtain the performance change regular sequence of biometrics in low light environment by using the sliding window method.

[0159] The rule mining module is used to perform rule mining based on the performance change pattern sequence and a classification algorithm to generate a core adaptive rule set that includes branch judgments based on environmental conditions.

[0160] The application module is used to determine the adaptive adjustment parameters corresponding to each environmental condition classification set based on the core adaptive rule set, and apply the adaptive adjustment parameters in a simulated biometric authentication environment to obtain an optimized performance index dataset.

[0161] The merging module is used to merge the optimized performance index dataset with the environmental factor-affected performance index dataset. The merged dataset is then grouped using a clustering algorithm to obtain a set of biometric authentication output rules. This set of biometric authentication output rules is used to perform secure authentication operations based on biometrics.

[0162] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A security authentication method based on multimodal biometric fusion, characterized in that, Executed by a computer, including: To obtain a performance index dataset for environmental factors, the following steps are taken: First, based on records in the log files of the biometric authentication system, obtain facial recognition accuracy and voice verification pass rate to generate a performance index sequence data. Second, based on records in the log files, extract the corresponding ambient light intensity level classification value and ambient noise decibel range label to generate an environmental factor label sequence. Third, if the records contain facial recognition accuracy, associate the ambient light intensity level classification value with the category label of the object obstructing the face to generate a facial recognition environmental impact sample pair. Fourth, if the records contain voice verification pass rate, associate the corresponding ambient noise decibel range label with the category label of the object affecting the sound to generate a voice verification environmental impact sample pair. Finally, based on the performance index sequence data, the environmental factor label sequence, the facial recognition environmental impact sample pair, and the voice verification environmental impact sample pair, determine the performance index dataset for environmental factors. Based on the aforementioned dataset of environmental factors affecting performance indicators, a clustering algorithm was used to group the data, resulting in a set of environmental condition classifications including groups with insufficient light and normal noise. Based on the group with insufficient light and normal noise, the accuracy change trend over a continuous time period is calculated using the sliding window method to obtain the performance change pattern sequence of biometrics under insufficient light conditions. Specifically, this includes: using the environmental factor impact performance index dataset corresponding to the group with insufficient light and normal noise as a subsequence dataset, and determining the fluctuation value of facial recognition accuracy based on the subsequence dataset; if the fluctuation value exceeds a preset fluctuation threshold, then extracting facial recognition accuracy values ​​window by window using the sliding window method based on the subsequence dataset to obtain a facial recognition accuracy subset; based on the facial recognition accuracy subset, determining the accuracy difference subsequence between adjacent windows for the sequence change of accuracy values ​​within each sliding window; determining the accuracy change trend over a continuous time period based on the direction and amplitude of the accuracy difference subsequence; and determining the performance change pattern sequence of biometrics under insufficient light conditions based on the accuracy change trend. Based on the performance change pattern sequence, a classification algorithm is used for rule mining to generate a core adaptation rule set containing environmental condition branch judgments. Specifically, this includes: determining the adjustment parameter range sequence for facial image texture enhancement in low-light scenarios and the dynamic correction coefficient sequence for sound signal feature weights based on the performance change pattern sequence; performing rule mining on the two types of sequence data using a classification algorithm based on the adjustment parameter range sequence and the dynamic correction coefficient sequence to obtain a candidate adaptation rule set covering multiple environmental conditions and the corresponding rule coverage degree; if the rule coverage degree reaches a preset proportion, then based on the candidate adaptation rule set, branch conditions are organized for the two dimensions of ambient light intensity and noise level to generate a branch judgment structure for environmental conditions; based on the branch judgment structure, a preset environmental condition threshold is used for condition filling to obtain a core adaptation rule set containing environmental condition branch judgments. Based on the core set of adaptation rules, the adaptive adjustment parameters corresponding to each environmental condition classification set are determined, and the adaptive adjustment parameters are applied in a simulated biometric authentication environment to obtain an optimized performance index dataset. The optimized performance index dataset and the environmental factor impact performance index dataset are merged, and the merged dataset is grouped using a clustering algorithm to obtain a set of biometric authentication output rules. This set of biometric authentication output rules is used to perform secure biometric authentication operations. Specifically, the process includes: merging the optimized performance index dataset and the environmental factor impact performance index dataset to obtain a merged dataset; grouping the merged dataset using a clustering algorithm to obtain preliminary grouping results; determining the first center position data for each group based on the preliminary grouping results, and determining the second center position data for the environmental condition classification set based on the environmental condition classification set; comparing the first center position data with the second center position data to obtain the center distance data between each group; calculating the average center distance of all groups based on the center distance data to obtain the average calculation result; if the average calculation result is lower than a preset average threshold, the preliminary grouping results are deemed to meet expectations, and the preliminary grouping results are determined to be the set of biometric authentication output rules.

2. The security authentication method based on multimodal biometric fusion according to claim 1, characterized in that, The dataset of performance indicators affected by environmental factors is grouped using a clustering algorithm to obtain a set of environmental condition classifications, including groups with insufficient light and normal noise levels. Based on the aforementioned dataset of performance indicators affected by environmental factors, a classification algorithm was used to perform a correlation analysis on the classification values ​​of facial recognition accuracy and ambient light intensity level, resulting in the first correlation matrix of the influence of facial recognition accuracy on the classification values ​​of ambient light intensity level. Based on the aforementioned dataset of performance indicators affected by environmental factors, a classification algorithm was used to perform correlation analysis on the sound verification pass rate and the environmental noise decibel range label to obtain a second correlation matrix of the influence of the sound verification pass rate on the environmental noise decibel range label. Based on the first association matrix, the first association strength between facial recognition accuracy and the classification value of each ambient light intensity level is determined. If the first association strength is greater than the first preset threshold, the corresponding facial recognition accuracy and ambient light intensity level classification value are determined to be the main influencing factor pair for facial recognition. Based on the second correlation matrix, the second correlation strength between the sound verification pass rate and each environmental noise decibel range label is determined. If the second correlation strength is greater than the second preset threshold, the corresponding sound verification pass rate and environmental noise decibel range label are determined to be the main influencing factor pair for sound verification. The main influencing factors of facial recognition and the main influencing factors of voice verification are combined to determine the target association matrix; The influence paths of environmental factor changes are extracted from the target correlation matrix, and based on the influence paths of environmental factor changes, a clustering algorithm is used to group them to obtain an environmental condition classification set including groups with insufficient light and normal noise.

3. The security authentication method based on multimodal biometric fusion according to claim 1, characterized in that, The adaptive adjustment parameters corresponding to each environmental condition classification set are determined based on the core adaptive rule set, and the adaptive adjustment parameters are applied in a simulated biometric authentication environment to obtain an optimized performance index dataset, including: Based on the core set of adaptation rules, determine the adaptive adjustment parameters corresponding to each environmental condition classification set; Based on the adaptive adjustment parameters, the corresponding parameter application operation is performed in a simulated biometric authentication environment to obtain the first performance index dataset; Based on the first performance index dataset, a support vector machine is used to classify the indexes under different environmental conditions to obtain a set of classification labels. Based on the set of classification labels, extract the subset of adjustment parameters corresponding to each classification label from the adaptive adjustment parameters; Based on the adjusted parameter subset, the parameters are applied again in the simulated biometric authentication environment to obtain a second performance index dataset. The dataset with the lower false recognition rate between the first performance metric dataset and the second performance metric dataset is used as the optimized performance metric dataset.

4. A security authentication system based on multimodal biometric fusion, used to implement the security authentication method based on multimodal biometric fusion as described in any one of claims 1 to 3, characterized in that, include: The dataset generation module is used to obtain the accuracy of facial recognition and the pass rate of voice verification related to biometric authentication, and at the same time, obtain the classification value of ambient light intensity level and the label of ambient noise decibel range at the corresponding time, so as to obtain the dataset of performance indicators affected by environmental factors. The grouping module is used to perform grouping processing based on the dataset of performance indexes affected by environmental factors using a clustering algorithm to obtain a set of environmental condition classifications including groups with insufficient light and normal noise. The regular sequence generation module is used to calculate the accuracy change trend over a continuous time period based on the group with insufficient light and normal noise, and obtain the performance change regular sequence of biometrics under insufficient light conditions by using the sliding window method. The rule mining module is used to perform rule mining based on the performance change pattern sequence, using a classification algorithm to generate a core adaptive rule set that includes environmental condition branch judgments; The application module is used to determine the adaptive adjustment parameters corresponding to each environmental condition classification set based on the core adaptive rule set, and apply the adaptive adjustment parameters in a simulated biometric authentication environment to obtain an optimized performance index dataset. The merging module is used to merge the optimized performance index dataset with the environmental factor impact performance index dataset, and to group the merged dataset using a clustering algorithm to obtain a set of biometric authentication output rules, wherein the set of biometric authentication output rules is used to perform biometric security authentication operations.

Images