An intelligent remote audit method

Abstract

The present application discloses an intelligent remote audit method, which includes: obtaining an audit request sent by a user terminal through an authentication interaction platform; retrieving verification data within a preset time window based on a user identifier, generating a target verification feature vector, inputting it into a preset risk assessment algorithm, and outputting a risk assessment value corresponding to the user terminal; if the risk assessment value is less than a preset risk threshold, obtaining a list of audit and certification tasks associated with the request type, and generating a matching result based on each audit and certification task using a matching diversion mechanism; establishing an information channel between the audit terminal and the user terminal that matches each audit and certification task, interacting with the audit and certification task, generating an audit result fed back by the audit terminal, and sending it to the user terminal through the authentication interaction platform. In this way, targeted matching is performed based on the risk differentiation requirements of the audit and certification task and the qualifications of the audit terminal, thereby improving the accuracy and efficiency of the audit.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a method. Background Art

[0002] Certification is a form of credit assurance, whereby a certification body verifies that an organization's products, services, and management systems comply with relevant standards, technical specifications, or other mandatory requirements, and issues a corresponding certificate. Currently, the certification industry faces challenges such as insufficient innovation, low efficiency, difficulty in traceability, and declining credibility, raising questions about the validity of certification results. The rapid development of new-generation information technology has further highlighted the shortcomings of traditional certification models. Consequently, due to factors such as distance and time constraints, certification bodies are unable to provide comprehensive certification services to geographically distant organizations, and remote audits remain inefficient.

[0003] The Chinese invention patent with patent application number 202110793174.8 discloses an intelligent remote audit method. In response to the audit request sent by the user terminal through the authentication interaction platform, if the verification parameters of the audit request are passed, at least one authentication task to be executed corresponding to the audit request is determined. The audit terminal is used to interact with the user terminal through the information channel in the authentication interaction platform to execute the authentication task to be executed, and return the task execution result of the authentication task to be executed according to the execution status of the authentication task to be executed, and send the audit result to the user terminal through the authentication interaction platform, so that the certification service provided by the certification agency is no longer limited by factors such as distance and time, thereby improving the efficiency of remote audits.

[0004] However, it is necessary to match the audit terminal based on the identification of the certification task to be performed. However, in actual scenarios, the audit terminal for remote audit is not unique, and the matching of relatively fixed or overly random audit terminals has limitations. It is impossible to perform differentiated matching and diversion based on the risk situation of audit requests from different user terminals and the qualification characteristics of the audit terminal. Summary of the Invention

[0005] This application improves the accuracy and efficiency of the audit by providing an intelligent remote audit method that performs targeted matching based on the risk differentiation requirements of the audit and certification tasks and the qualifications of the audit terminals.

[0006] This application provides an intelligent remote audit method, including:

[0007] S101, obtaining an audit request sent by a user terminal through an authentication interaction platform, the audit request including but not limited to user identification, request type, and timestamp information;

[0008] S102, based on the user identifier, retrieves verification data within a preset time window, generates a target verification feature vector, inputs it into a preset risk assessment algorithm, and outputs a risk assessment value corresponding to the user terminal;

[0009] S103, if the risk assessment value is less than the preset risk threshold, the user terminal is determined to have passed the verification, and step S104 is executed; otherwise, a warning message is returned to the user terminal for identity verification;

[0010] S104: Obtain an audit and authentication task list associated with the request type, the audit and authentication task list including at least one audit and authentication task, and generate a matching result for each audit and authentication task in the audit and authentication task list using a preset matching and diversion mechanism;

[0011] S105: Based on each audit and certification task, an information channel is established between the audit terminal and the user terminal that matches it, the audit and certification task is interacted, and the audit result fed back by the audit terminal is generated and sent to the user terminal through the authentication interaction platform.

[0012] Preferably, the verification data includes user historical behavior data and historical audit results within a preset time window; user historical behavior data includes audit timestamp information and request type; the target verification feature vector is composed of three risk factors: audit frequency jump value, request type trajectory feature value, and audit pass rate.

[0013] Preferably, the audit frequency jump value is set as the change amplitude value of the daily audit times within the preset time window; the audit pass rate is set as the proportion of audit results of all audit requests within the preset time window that are passed; the request type trajectory feature value is used to indicate the degree of disturbance of the request type trajectory feature by the current audit request.

[0014] Preferably, the preset risk assessment algorithm is specifically:

[0015] R=γ1×F+γ2×D+γ3×(1-S)

[0016] Among them, R is the risk assessment value, F is the audit frequency jump value, D is the request type trajectory characteristic value, S is the audit pass rate, γ1, γ2 and γ3 are the weight values ​​of the impact of the audit frequency jump value, request type trajectory characteristic value and audit pass rate on the risk assessment value, respectively, which are pre-set according to actual conditions and expert experience.

[0017] Preferably, the preset matching and diversion mechanism specifically includes:

[0018] S201: Obtain at least one audit terminal that matches the task identifier of the audit and certification task, and generate the current qualification feature vector of each audit terminal. The qualification feature vector includes the historical accuracy rate, processing rate, and task queue idle rate, expressed as V Ej =[S1,S2,S3],V Ej is the qualification feature vector, S1 is the historical accuracy, S2 is the processing rate, and S3 is the task queue idle rate;

[0019] S202: Based on the risk assessment value and its corresponding risk factors, the demand vector V of each audit and certification task is obtained. Tj =[e1×R,e2×(1-R),e3×R], where e1, e2, and e3 are preset adjustment coefficients. e1×R, e2×(1-R), and e3×R are used to represent the degree of demand for the historical accuracy rate, processing rate, and task queue idle rate of the audit and certification task under the risk assessment value.

[0020] S203: Calculate the matching degree between the audit and certification task and each audit terminal according to the following formula:

[0021]

[0022] Among them, Match ij is the matching degree between the audit and certification task and the audit terminal Ej, V Ej is the qualification feature vector of the audit terminal Ej, V Tj is the demand vector of the audit and certification task, V Ej ·V Tj Indicates V Ej With V Tj The dot product of ||V Ej || represents the modulus of the qualification feature vector, ||V Tj || represents the modulus of the demand vector;

[0023] S204: The audit terminal with the highest matching degree is used as the matching result of the audit and authentication task.

[0024] Preferably, the request type trajectory feature value is obtained in the following manner:

[0025] A1. Obtain all timestamp information of the request type within the preset time window to obtain the first occurrence trajectory {t1, t2, ..., t n-1}, based on the request type and timestamp information of the current audit request, update the first occurrence trajectory to obtain the second occurrence trajectory {t1, t2, ..., tn}, where tn is the current time point;

[0026] A2, based on the difference between the adjacent time points in the first and second occurrence trajectories, the corresponding first audit cycle sequence {d1, d2, ..., d i ,...,d n-2}, the second review cycle sequence {d1,d2,...,d j ,...,d n-2 , d n-1}, where d i =t i+1 -t i , d j =t j+1 -t j , i and j are used to represent numbers, and their value ranges are [1, n-2] and [1, n-1] respectively. The absolute value of the difference between the first floating index of the first audit cycle sequence and the second floating index of the second audit cycle sequence is calculated as the first disturbance index;

[0027] A3. Obtain other request types associated with the request type within a preset time window as target types, obtain all timestamp information of the target types within the preset time window to obtain a third occurrence trajectory and its corresponding third review period sequence, and calculate a second disruption index;

[0028] A4. Perform weighted summation based on the first disturbance index and the second disturbance index to obtain a request type trajectory feature value.

[0029] Preferably, the second disturbance index is obtained in the following manner:

[0030] Calculate the absolute value of the difference between the second floating index of the second review cycle sequence and the third floating index of the third review cycle sequence, and calculate the second disruption index according to the following formula:

[0031]

[0032] Where D is the second disturbance index, M is the number of target types associated with the request type within the preset time window, and w k The association weight value of the k-th target type is preset according to the business logic of the k-th target type and the request type.

[0033] Preferably, in said S202, the method for generating the adjustment coefficient in the demand vector of each audit and certification task specifically includes:

[0034] S301, obtain the risk assessment value of the audit request and the risk contribution value of each risk factor, and form a risk feature vector [R,w F ,w D ,w S ], R is the risk assessment value, wF 、w D and w S They represent the risk contribution value of the audit frequency jump value, the request type trajectory characteristic value, and the audit pass rate respectively;

[0035] S302: Input the risk feature vector into the pre-trained adjustment coefficient generation model, and output the adjustment coefficient combination (e1, e2, e3).

[0036] Preferably, the risk contribution value of the audit frequency jump value, the request type trajectory characteristic value, and the audit pass rate is calculated according to the following formula:

[0037] Among them, w F 、w D and w S They respectively represent the risk contribution values ​​of the audit frequency jump value, the request type trajectory characteristic value and the audit pass rate, R is the risk assessment value, F is the audit frequency jump value, D is the request type trajectory characteristic value, S is the audit pass rate, γ1, γ2 and γ3 are the impact weights of the audit frequency jump value, the request type trajectory characteristic value and the audit pass rate on the risk assessment value.

[0038] Preferably, the method for obtaining the pre-trained adjustment coefficient generation model specifically includes:

[0039] B1. Collect a large number of audit records for historical audit and certification tasks. The audit records include the risk assessment value and risk factors of the request type corresponding to the audit and certification task, the qualification feature vector of the corresponding audit terminal, the feedback audit results and the assessed audit accuracy, and obtain the risk feature vector for each historical audit and certification task.

[0040] B2. Filter out risk feature vectors whose audit accuracy rate reaches the accuracy threshold from the audit results fed back by the audit terminal as training samples;

[0041] B3. Label each training sample, setting the label content to the optimal adjustment coefficient combination corresponding to the risk feature vector;

[0042] B4. Use the labeled training samples as a training set, use the training set to train the pre-selected machine learning structure, optimize the model parameters, and obtain the final adjustment coefficient generation model.

[0043] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0044] By introducing a risk assessment mechanism, preliminary security verification is performed on high-risk users, unnecessary audit steps are reduced, and timely feedback warnings are provided for security identity audits, thereby improving the security of remote audits. By utilizing a matching and diversion mechanism, differentiated matching is performed based on the risk differentiation requirements of audit and certification tasks and the qualifications of audit terminals, thereby improving the accuracy and efficiency of audits. The matching and diversion mechanism takes into account the historical accuracy, processing rate and task queue idle rate of audit terminals, ensuring that audit tasks can be assigned to the most suitable audit terminals. The demand vector is dynamically adjusted according to the risk assessment value to adapt to the audit requirements under different risk situations. Through risk assessment and differentiated matching, the uncertainty caused by human intervention and random allocation is reduced, and the credibility of the audit is improved.

[0045] In traditional remote audits, terminal matching is rigid (fixed or randomly assigned) and cannot be adjusted dynamically according to risks. By introducing risk assessment values ​​to quantify user risks and achieve risk stratification, the design of demand vectors dynamically reflects the requirements for audit terminal qualifications in tasks under different risk scenarios, thereby improving the flexibility and accuracy of audit terminal allocation and reducing the misjudgment rate of high-risk tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the flow of the intelligent remote audit method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] To facilitate understanding of the present invention, the present application will be described more comprehensively below with reference to the relevant drawings; the drawings show preferred embodiments of the present invention, but the present invention can be implemented in many different forms and is not limited to the embodiments described herein; on the contrary, the purpose of providing these embodiments is to enable a more thorough and comprehensive understanding of the disclosed content of the present invention.

[0048] It should be noted that the terms “vertical”, “horizontal”, “up”, “down”, “left”, “right” and similar expressions used in this document are for illustrative purposes only and do not represent the only implementation method.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains; the terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention; the term "and / or" used herein includes any and all combinations of one or more of the associated listed items.

[0050] Example 1: Figure 1 It is a flowchart of the intelligent remote audit method according to an embodiment of the present invention.

[0051] like Figure 1As shown, an intelligent remote audit method includes the following steps:

[0052] S101, obtaining an audit request sent by a user terminal through an authentication interaction platform, where the audit request includes but is not limited to user identification, request type, and timestamp information.

[0053] Specifically, the audit request is received and parsed through the API interface. For example, a user submits an audit request through a mobile phone APP. The audit request contains the user's unique ID (user identification), the request type is "identity authentication", and the timestamp of the request being sent. Among them, the request type refers to the type information corresponding to the audit request. Each type of information corresponds to an audit and authentication task list containing at least one task. Therefore, the server can obtain the corresponding audit and authentication task list based on the request type carried by the audit request. For example, if the request information is "identity authentication", the corresponding audit and authentication task list is pulled from the pre-built task database: T1-real-name verification, T2-liveness detection.

[0054] S102: Retrieve verification data within a preset time window based on the user identifier, generate a target verification feature vector, input it into a preset risk assessment algorithm, and output a risk assessment value corresponding to the user terminal.

[0055] The verification data includes historical user behavior data and historical audit results within a preset time window. The preset time window can be set to the past month or the past week, depending on the actual situation and scenario. Historical user behavior data includes audit timestamp information and request type.

[0056] Specifically, the target verification feature vector is composed of three risk factors: the audit frequency jump value, the request type trajectory feature value, and the audit pass rate. The audit frequency jump value is set as the change in the number of daily audits within the preset time window (which can be set as variance or standard deviation) to reflect the volatility of user behavior. It is calculated according to the following formula: Where N is the total number of time units in the preset time window, x i is the number of audit requests in the i-th time unit within the preset time window, μ is the average number of audit requests in the time unit, and the time unit is set to one day; the audit pass rate is set to the proportion of audit results of all audit requests in the preset time window that are passed; the request type trajectory feature value is used to indicate the degree of disturbance of the request type trajectory feature by the current audit request.

[0057] In some embodiments, the preset risk assessment algorithm is specifically:

[0058] R=γ1×F+γ2×D+γ3×(1-S)

[0059] Among them, R is the risk assessment value, F is the audit frequency jump value, D is the request type trajectory characteristic value, S is the audit pass rate, γ1, γ2 and γ3 are the weight values ​​of the impact of the audit frequency jump value, the request type trajectory characteristic value and the audit pass rate on the risk assessment value, respectively. They are pre-set according to actual conditions and expert experience, and the sum of γ1, γ2, and γ3 is 1.

[0060] S103, if the risk assessment value is less than the preset risk threshold, it is determined that the user terminal has passed the verification and step S104 is executed. Otherwise, a warning message is returned to the user terminal for identity verification.

[0061] The preset risk threshold is set according to actual conditions and expert experience. For example, the preset risk threshold is set to 0.6.

[0062] S104, obtaining an audit and authentication task list associated with the request type, the audit and authentication task list including at least one audit and authentication task, and generating a matching result based on each audit and authentication task in the audit and authentication task list using a preset matching and diversion mechanism.

[0063] In some embodiments, the preset matching and diversion mechanism specifically includes:

[0064] S201: Obtain at least one audit terminal that matches the task identifier of the audit and certification task (it can be understood that the audit terminal containing the execution program corresponding to the task identifier is determined to be a match, each audit and certification task is assigned a task identifier, each task identifier corresponds to an execution program, and the audit and certification task can be implemented according to the execution program). Generate the current qualification feature vector of each audit terminal. The qualification feature vector includes the historical accuracy rate, processing rate and task queue idle rate, which is expressed as V Ej =[S1,S2,S3],V Ej is the qualification feature vector, S1 is the historical accuracy, which is expressed as the correct rate of the historical audit results of the audit terminal within the preset time window (the accuracy of the audit results of the audit terminal and its accuracy can be judged by experts). It is an important indicator for measuring the reliability of terminal audits. S2 is the processing rate, which is determined according to the current network bandwidth and performance of the audit terminal. The better the network condition, the greater the processing rate. S3 is the task queue idle rate, which is calculated based on the current task queue status of the audit terminal and reflects the terminal's immediate ability to process new tasks.

[0065] S202: Based on the risk assessment value and its corresponding risk factors, the demand vector V of each audit and certification task is obtained. Tj=[e1×R,e2×(1-R),e3×R], where e1, e2, and e3 are preset adjustment coefficients. e1×R, e2×(1-R), and e3×R are respectively used to represent the demand degree values ​​of the audit and certification task for the historical accuracy rate, processing rate, and task queue idle rate within the qualification feature vector under the risk assessment value. The combination of adjustment coefficients can achieve the demand vector tending towards high historical accuracy rate, high task queue idle rate, and low processing rate under higher risk assessment values.

[0066] S203: Calculate the matching degree between the audit and certification task and each audit terminal according to the following formula:

[0067]

[0068] Among them, Match ij is the matching degree between the audit and certification task and the audit terminal Ej, V Ej is the qualification feature vector of the audit terminal Ej, V Tj is the demand vector of the audit and certification task, V Ej ·V Tj Indicates V Ej With V Tj The dot product of ||V Ej || represents the modulus of the qualification feature vector, ||V Tj || represents the modulus of the demand vector.

[0069] S204: The audit terminal with the highest matching degree is used as the matching result of the audit and authentication task.

[0070] S105: Based on each audit and certification task, an information channel is established between the audit terminal and the user terminal that matches it, the audit and certification task is interacted, and the audit result fed back by the audit terminal is generated and sent to the user terminal through the authentication interaction platform.

[0071] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages:

[0072] By introducing a risk assessment mechanism, preliminary security verification is performed on high-risk users, unnecessary audit steps are reduced, and timely feedback warnings are provided for security identity audits, thereby improving the security of remote audits. By utilizing a matching and diversion mechanism, differentiated matching is performed based on the risk differentiation requirements of audit and certification tasks and the qualifications of audit terminals, thereby improving the accuracy and efficiency of audits. The matching and diversion mechanism takes into account the historical accuracy, processing rate and task queue idle rate of audit terminals, ensuring that audit tasks can be assigned to the most suitable audit terminals. The demand vector is dynamically adjusted according to the risk assessment value to adapt to the audit requirements under different risk situations. Through risk assessment and differentiated matching, the uncertainty caused by human intervention and random allocation is reduced, and the credibility of the audit is improved.

[0073] In traditional remote audits, terminal matching is rigid (fixed or randomly assigned) and cannot be adjusted dynamically according to risks. By introducing risk assessment values ​​to quantify user risks and achieve risk stratification, the design of demand vectors dynamically reflects the requirements for audit terminal qualifications in tasks under different risk scenarios, thereby improving the flexibility and accuracy of audit terminal allocation and reducing the misjudgment rate of high-risk tasks.

[0074] Example 2: Further limit the request type trajectory feature value in Example 1.

[0075] Therefore, the embodiments of the present application are optimized based on the above embodiments.

[0076] In some embodiments, the request type trajectory feature value is obtained in the following manner:

[0077] A1. Obtain all timestamp information of the request type within the preset time window to obtain the first occurrence trajectory {t1, t2, ..., t n-1}, based on the request type and timestamp information of the current audit request, the first occurrence trajectory is updated to obtain the second occurrence trajectory {t1, t2, ..., tn}, where tn is the current time point.

[0078] A2, based on the difference between the adjacent time points in the first and second occurrence trajectories, the corresponding first audit cycle sequence {d1, d2, ..., d i ,...,d n-2}, the second review cycle sequence {d1,d2,...,d j ,...,d n-2 , d n-1}, where d i =t i+1 -t i , d j =t j+1 -t j , i and j are both used to represent numbers, and their value ranges are [1,n-2] and [1,n-1] respectively. The absolute value of the difference between the first floating index of the first audit cycle sequence and the second floating index of the second audit cycle sequence is calculated as the first disturbance index, which can be understood as the degree of influence of the audit request at the current time point on the occurrence trajectory characteristics of the request type. The larger the first disturbance index, the greater the possibility of illegal audit at the current time point.

[0079] Among them, the first floating index is set to the variance value of the first review cycle sequence, which is used to reflect the regularity of the time dimension of the review request in the first appearance trajectory; the second floating index is set to the variance value of the second review cycle sequence, which is used to reflect the regularity of the time dimension of the review request in the second outgoing trajectory.

[0080] A3. Obtain other request types associated with the request type within a preset time window as target types (at least one), obtain all timestamp information of the target type within the preset time window to obtain a third occurrence trajectory and its corresponding third review period sequence, calculate the absolute value of the difference between the second floating index of the second review period sequence and the third floating index of the third review period sequence, and calculate the second disturbance index according to the following formula:

[0081]

[0082] Where D is the second disturbance index, M is the number of target types associated with the request type within the preset time window, and w k The association weight for the kth target type is preset based on the business logic between the kth target type and the request type. The greater the business logic relevance between the two, the larger the association weight. This value is set based on expert experience. It should be noted that target types are screened based on a preset business logic relevance. When the business logic relevance reaches a preset logical threshold, the target type is considered associated with the request type. The corresponding association weight is preset and ranges from 0 to 1.

[0083] A4. Perform a weighted summation based on the first and second disturbance indices to obtain a request type trajectory feature value. The sum of the two weights is 1, and the weight of the second disturbance index is slightly smaller than that of the first. For example, the weight of the second disturbance index is set to 0.4, and the weight of the first disturbance index is set to 0.6. The two indices are used in a coordinated manner to effectively mitigate the randomness of user terminal review requests.

[0084] Therefore, by analyzing the differences between the historical and current trajectory characteristics of the request type itself, and the current and related request type trajectory characteristics, we can obtain the first disturbance index and the second disturbance index from different perspectives, and comprehensively generate the request type trajectory feature value, thereby more comprehensively and reliably reflecting the rationality and security of the request type at the current time point, and effectively detecting some illegal review behaviors that have been attacked by hackers.

[0085] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages:

[0086] By introducing the first and second disturbance indices, the request type trajectory characteristics are analyzed from multiple perspectives, reflecting the rationality and security of the request type more comprehensively. The first disturbance index focuses on the historical and current trajectory characteristics of the request type itself, while the second disturbance index considers the current and related request type trajectory characteristics. The combination of the two can effectively detect illegal review behavior. The calculation of the request type trajectory characteristic value takes into account multiple factors, including timestamp information, related request type, business logic relevance, etc., making risk assessment more reliable. The weighted summation of the first and second disturbance indices can balance the impact of different factors, improve the robustness of risk assessment, and be applicable to different types of review requests and scenarios. Through refined analysis, it can better adapt to changes in user behavior and market demand, and improve the flexibility and adaptability of review.

[0087] By introducing the request type trajectory feature value into the target feature vector, a multi-dimensional trajectory analysis of the user terminal on this request type is performed. By combining its own history with the temporal characteristics of the associated request type, the anomaly detection capability is enhanced, and illegal requests (such as frequently initiating "identity authentication" and linked "fund transfer" requests in a short period of time) are effectively avoided. This allows for a more comprehensive and reliable risk assessment of the user terminal's current request type.

[0088] Embodiment 3: Further limit the combination of adjustment coefficients in the demand vector of each audit and certification task in embodiment 1.

[0089] Therefore, the embodiments of the present application are optimized based on the above embodiments.

[0090] In some embodiments, in order to more accurately adjust the adjustment coefficient according to the risk assessment value, an adjustment coefficient generation model is constructed. The generation method of the adjustment coefficient specifically includes:

[0091] S301, obtain the risk assessment value of the audit request and the risk contribution value of each risk factor (audit frequency jump value, request type trajectory characteristic value, audit pass rate), and form a risk feature vector [R, w F ,w D ,w S ], where R is the risk assessment value, w F 、w D and w S They represent the risk contribution value of the audit frequency jump value, the request type trajectory characteristic value and the audit pass rate respectively.

[0092] S302: Input the risk feature vector into a pre-trained adjustment coefficient generation model, and output the adjustment coefficient combination (e1, e2, e3). The adjustment coefficient combination is used to adjust the weights of historical accuracy, processing rate, and task queue idle rate in the demand vector to adapt to the audit requirements of different risk situations.

[0093] In some embodiments, a method for obtaining a pre-trained adjustment coefficient generation model specifically includes:

[0094] B1. Collect a large number of audit records of historical audit and certification tasks. The audit records include the risk assessment value and various risk factors of the request type corresponding to the audit and certification task, the qualification feature vector of the corresponding audit terminal, the feedback audit results and the assessed audit accuracy, and obtain the risk feature vector of each historical audit and certification task.

[0095] B2. Filter out risk feature vectors whose audit accuracy of the audit results fed back by the audit terminal reaches an accuracy threshold (which can be set to 95%) as training samples.

[0096] B3. Label each training sample, and set the label content to the optimal adjustment coefficient combination corresponding to the risk feature vector (labeling is performed based on expert experience. The labeling criteria of expert experience are: according to the risk assessment value and the contribution of risk factors, dynamically adjust the demand weight of the audit task for the audit terminal qualification (historical accuracy, processing rate, task queue idle rate). High-risk scenario: the demand vector determined by the adjustment coefficient tends to favor audit terminals with high historical accuracy and high task queue idle rate; low-risk scenario: the demand vector determined by the adjustment coefficient tends to favor audit terminals with fast processing rate; medium-risk scenario: the demand vector determined by the adjustment coefficient tends to favor audit terminals with balanced distribution of demand degree values ​​for the qualification feature vector. The adjustment coefficient can also be adaptively determined based on the risk contribution value of the risk factor to explore the relationship between the risk factor and each element in the qualification feature vector).

[0097] B4. Use the labeled training samples as a training set, use the training set to train the pre-selected machine learning structure, optimize the model parameters, and obtain the final adjustment coefficient generation model.

[0098] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages:

[0099] By building an adjustment coefficient generation model and training it based on a large amount of historical audit records, we can more accurately output adjustment coefficient combinations that are suitable for different risk scenarios based on risk assessment values ​​and the risk contribution values ​​of each risk factor. This avoids the subjectivity and arbitrariness of manually setting adjustment coefficients and improves their objectivity and accuracy. The dynamic adjustment of adjustment coefficient combinations enables the demand vector to adapt to audit needs in different risk situations. In high-risk scenarios, the demand vector tends to select audit terminals with high historical accuracy and high task queue idleness, reducing the misjudgment rate. In low-risk scenarios, the demand vector tends to select audit terminals with high processing speeds, improving efficiency. Training the adjustment coefficient generation model through machine learning technology enables intelligent decision-making in the audit process.

[0100] Historical data is used to learn the mapping relationship between risk scenarios and adjustment coefficients. High-accuracy samples are screened to prevent low-quality data from interfering with model training, and dynamic generation of adjustment coefficient combinations is achieved to adapt to the differentiated needs of terminals in different risk scenarios.

[0101] Embodiment 4: Further limiting the method for obtaining the adjustment coefficient generation model.

[0102] Therefore, the embodiments of the present application are optimized based on the above embodiments.

[0103] In some embodiments, in step B3, the method for determining the annotation content of each training sample further includes:

[0104] S401, define the optimization problem: the decision variables are the adjustment coefficient combination (e1, e2, e3), and the objective function is to maximize the qualification matching degree of the audit terminal: The constraint condition is e1×R+e2×(1-R)+e3×R=1, which realizes the normalization constraint. When realizing the constraint condition, the upper and lower limits of the demand degree value in the demand vector are set according to the risk assessment value and the contribution index of each risk factor:

[0105] High-risk scenario (R>R high ): e1×R≥0.6, e2×(1-R)≤0.2, e3×R≥0.2;

[0106] Medium risk scenario (R low ≤R≤R high ): 0.3≤e1×R≤0.6, e2×(1-R)≤0.5, e3×R≤0.5;

[0107] Low risk scenario (R<R low ): e1×R≤0.3, e2×(1-R)≥0.5, e3×R≤0.3; where R low Greater than or equal to the preset risk threshold, R high Greater than Rlow And it is less than 1. It should be set according to the actual situation.

[0108] S402, solve the optimal adjustment coefficient:

[0109] The above optimization problem is solved using the existing linear programming (LP) technology to obtain the optimal adjustment coefficient combination (e1, e2, e3) for each training sample.

[0110] S403: Based on a pre-set rule base, the optimal adjustment coefficient combination is integrated with expert experience and manually annotated and revised. Domain experts verify and adjust the optimal adjustment coefficient combination for some training samples. For example, for certain special request types (such as financial-grade identity verification), even if the risk assessment value is low, e1×R≥0.6 is still required to ensure accuracy.

[0111] Among them, expert experience is converted into a pre-set rule base, including special request types and their corresponding special constraint rules. For example, if a special request type is "fund transfer", then any risk assessment value is forced to be e1×R≥0.8, which is a special constraint rule.

[0112] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages:

[0113] When setting labels for model training samples, by defining an optimization problem and solving it using linear programming techniques, the optimal combination of adjustment coefficients can be determined more scientifically, avoiding the subjectivity and arbitrariness of manual settings; converting expert experience into a rule base and manually annotating and correcting the optimal combination of adjustment coefficients ensures the accuracy and reliability of the model in specific scenarios, especially for some special request types, such as financial-level identity authentication or fund transfers. By forcibly setting the lower limit of certain adjustment coefficients, the security and accuracy of the audit process are ensured; by considering the demand vector adjustment under different risk scenarios, the model can better adapt to the audit needs under different risk situations, further improving the accuracy of the adjustment coefficient and the adaptability of the model.

[0114] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An intelligent remote audit method, characterized in that: include: S101, obtaining an audit request sent by a user terminal through an authentication interaction platform, the audit request including but not limited to user identification, request type, and timestamp information; S102, based on the user identifier, retrieves verification data within a preset time window, generates a target verification feature vector, inputs it into a preset risk assessment algorithm, and outputs a risk assessment value corresponding to the user terminal; S103, if the risk assessment value is less than the preset risk threshold, the user terminal is determined to have passed the verification, and step S104 is executed; otherwise, a warning message is returned to the user terminal for identity verification; S104: Obtain an audit and authentication task list associated with the request type, the audit and authentication task list including at least one audit and authentication task, and generate a matching result for each audit and authentication task in the audit and authentication task list using a preset matching and diversion mechanism; The matching and diversion mechanism includes: S201, obtaining at least one audit terminal that matches the task identifier of the audit and certification task, generating the current qualification feature vector of each audit terminal, the qualification feature vector includes the historical accuracy rate, processing rate and task queue idle rate; S202, based on the risk assessment value and its corresponding risk factors, obtaining the demand vector of each audit and certification task [ ], 、 、 is the preset adjustment coefficient, 、 、 They are used to represent the degree of demand for the historical accuracy rate, processing rate and task queue idle rate in the qualification feature vector under the risk assessment value of the audit and certification task, and R is the risk assessment value; S203, calculate the matching degree between the audit and certification task and each audit terminal according to the following formula: , Certify the audit task and audit terminal The matching degree, For audit terminal The qualification feature vector of is the demand vector of the audit and certification task, express and The dot product of represents the modulus of the qualification feature vector, Indicates the modulus of the demand vector; S204, taking the audit terminal with the highest matching degree as the matching result of the audit and certification task; S105: Based on each audit and certification task, an information channel is established between the audit terminal and the user terminal that matches it, the audit and certification task is interacted, and the audit result fed back by the audit terminal is generated and sent to the user terminal through the authentication interaction platform.

2. The intelligent remote audit method according to claim 1, characterized in that: The verification data includes user historical behavior data and historical audit results within a preset time window; user historical behavior data includes audit timestamp information and request type; the target verification feature vector is composed of three risk factors: audit frequency jump value, request type trajectory feature value, and audit pass rate.

3. The intelligent remote audit method according to claim 2, characterized in that: The audit frequency jump value is set as the change amplitude of the number of daily audits within the preset time window; the audit pass rate is set as the proportion of audit results of all audit requests within the preset time window that are passed; the request type trajectory feature value is used to indicate the degree of disturbance of the request type trajectory feature by the current audit request.

4. The intelligent remote audit method according to claim 3, characterized in that: The preset risk assessment algorithm is specifically: ； Among them, R is the risk assessment value, is the audit frequency jump value, D is the request type trajectory feature value, S is the audit pass rate, 、 and They are the audit frequency jump value, request type trajectory characteristic value, and the weight value of the impact of the audit pass rate on the risk assessment value, which are pre-set according to actual conditions and expert experience.

5. The intelligent remote audit method according to claim 2, characterized in that: The method for obtaining the trajectory feature value of the request type is: A1. Obtain all timestamp information of the request type within the preset time window to obtain the first occurrence trajectory {t1, t2, ..., }, based on the request type and timestamp information of the current audit request, update the first occurrence trajectory to obtain the second occurrence trajectory {t1, t2, ..., tn}, where tn is the current time point; A2, based on the difference between the adjacent time points in the first and second occurrence trajectories, the corresponding first audit cycle sequence {d1, d2, ..., ,..., }、Second review cycle sequence {d1,d2,..., ,..., , },in, , , i and j are used to represent numbers, and their value ranges are [1, n-2] and [1, n-1] respectively. The absolute value of the difference between the first floating index of the first audit cycle sequence and the second floating index of the second audit cycle sequence is calculated as the first disturbance index; A3. Obtain other request types associated with the request type within a preset time window as target types, obtain all timestamp information of the target types within the preset time window to obtain a third occurrence trajectory and its corresponding third review period sequence, and calculate a second disruption index; A4. Perform weighted summation based on the first disturbance index and the second disturbance index to obtain a request type trajectory feature value.

6. The intelligent remote audit method according to claim 5, characterized in that: The second disturbance index is obtained as follows: Calculate the absolute value of the difference between the second floating index of the second review cycle sequence and the third floating index of the third review cycle sequence, and calculate the second disruption index according to the following formula: ； Where D is the second disturbance index, M is the number of target types associated with the request type within the preset time window, The association weight value of the k-th target type is pre-set according to the business logic of the k-th target type and the request type. The second floating index of the second review cycle sequence, The third floating index corresponding to the third review cycle sequence.

7. The intelligent remote audit method according to claim 4, characterized in that: In S202, the method for generating the adjustment coefficient in the demand vector of each audit and certification task specifically includes: S301, obtain the risk assessment value of the audit request and the risk contribution value of each risk factor, and form a risk feature vector [R, , , ], R is the risk assessment value, 、 and They represent the risk contribution value of the audit frequency jump value, the request type trajectory characteristic value, and the audit pass rate respectively; S302, input the risk feature vector into the pre-trained adjustment coefficient generation model, and output the adjustment coefficient combination ( , , ).

8. The intelligent remote audit method according to claim 7, characterized in that: The audit frequency jump value, request type trajectory characteristic value and audit pass rate risk contribution value are calculated according to the following formula: , , ,in, 、 and They represent the risk contribution value of the audit frequency jump value, the request type trajectory characteristic value and the audit pass rate, respectively. R is the risk assessment value. is the audit frequency jump value, D is the request type trajectory feature value, S is the audit pass rate, 、 and They are the impact weights of the audit frequency jump value, request type trajectory characteristic value, and audit pass rate on the risk assessment value.

9. The intelligent remote audit method according to claim 7, characterized in that: The method for obtaining the pre-trained adjustment coefficient generation model specifically includes: B1. Collect a large number of audit records for historical audit and certification tasks. The audit records include the risk assessment value and risk factors of the request type corresponding to the audit and certification task, the qualification feature vector of the corresponding audit terminal, the feedback audit results and the assessed audit accuracy, and obtain the risk feature vector for each historical audit and certification task. B2. Filter out risk feature vectors whose audit accuracy rate reaches the accuracy threshold from the audit results fed back by the audit terminal as training samples; B3. Label each training sample, setting the label content to the optimal adjustment coefficient combination corresponding to the risk feature vector; B4. Use the labeled training samples as a training set, use the training set to train the pre-selected machine learning structure, optimize the model parameters, and obtain the final adjustment coefficient generation model.

Images