Multi-party signature electronic payment method and system based on Internet

Abstract

The invention provides an internet-based multi-party signature electronic payment method and system. The method comprises the following steps: acquiring historical signature data, analyzing signature behavior statistical characteristics by adopting a machine learning algorithm, obtaining a signature abnormality judgment quantitative standard, and determining abnormal thresholds of signature duration, stroke number, pressure change and timestamp characteristics; according to the signature anomaly judgment quantification standard, obtaining multi-dimensional features of a signature track, speed and a timestamp, and performing comprehensive analysis on the multi-dimensional features by adopting a deep learning technology to obtain an abnormal signature recognition accuracy rate; establishing a signature behavior risk assessment model according to the abnormal signature recognition accuracy, and quantifying the signature behavior risk degree to obtain a signature intervention strategy decision basis; and if the difference between the signature timestamps exceeds a preset threshold value, triggering a signature delay abnormal warning, and starting a signature intervention measure.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an Internet-based multi-party signature electronic payment method and system. Background Art

[0002] In the scenario of multi-party signature electronic payment, signature intervention technology faces many technical difficulties. First, it is difficult to unify the judgment criteria for abnormal signature behavior. Different payment scenarios have different definitions of signature anomalies, which makes it difficult to determine the triggering conditions for intervention. For example, in some scenarios, signature delay may be regarded as abnormal behavior, while in other scenarios it may be regarded as normal. This difference makes it difficult to formulate a unified abnormal signature judgment standard. Secondly, it is difficult to balance the timeliness and accuracy of signature intervention. In order to prevent risks in a timely manner, the system needs to monitor and intervene in abnormal signature behavior in real time. However, too frequent intervention may affect the normal payment process and reduce user experience. At the same time, if the intervention strategy is too loose, it may not be able to effectively prevent risks, resulting in threats to fund security. In addition, signature intervention also faces the problem of data inconsistency. In the scenario of multi-party signature, the signature data of each participant may be inconsistent, such as the timestamp of the signature is not synchronized, the signature algorithm is inconsistent, etc. This data inconsistency increases the difficulty of judging abnormal signature behavior and further increases the complexity of signature intervention technology. In summary, in multi-party signature electronic payment scenarios, signature intervention technology faces technical contradictions such as difficulty in unifying judgment standards, difficulty in balancing timeliness and accuracy, and inconsistent data. How to minimize the impact on the normal payment process while ensuring the security of funds is a key issue that signature intervention technology needs to solve. This requires comprehensive consideration of factors such as the characteristics of payment scenarios, risk prevention needs, and user experience, and continuously improving the effectiveness and reliability of signature intervention technology by optimizing abnormal signature judgment algorithms, dynamically adjusting intervention strategies, and other means. Summary of the invention

[0003] The present invention provides an Internet-based multi-party signature electronic payment method, which mainly includes:

[0004] Obtain historical signature data, use machine learning algorithms to analyze the statistical characteristics of signature behavior, obtain quantitative standards for signature anomaly judgment, and determine the abnormal thresholds of signature duration, number of strokes, pressure changes, and timestamp characteristics; According to the signature abnormality judgment quantitative standard, the multi-dimensional features of signature trajectory, speed and timestamp are obtained, and the multi-dimensional features are comprehensively analyzed using deep learning technology to obtain the abnormal signature recognition accuracy; According to the abnormal signature recognition accuracy, a signature behavior risk assessment model is established to quantify the risk level of the signature behavior and obtain a decision-making basis for the signature intervention strategy; Acquire the signature timestamps of each participant in real time from the Internet, and determine whether the difference between the signature timestamps exceeds a preset threshold based on the abnormal thresholds of the signature duration, number of strokes, pressure change and timestamp characteristics and the signature behavior risk assessment model; If the difference between the signature timestamps exceeds a preset threshold, a signature delay abnormality warning is triggered and signature intervention measures are initiated; Based on the transaction amount, user credit score and transaction frequency factors, the payment risk level is comprehensively assessed, and an adaptive signature intervention strategy is adopted. Through a dynamic weight allocation mechanism, the intervention frequency and intensity are automatically adjusted according to the importance of different factors.

[0005] The present invention provides an Internet-based multi-party signature electronic payment system, which mainly includes: A signature anomaly judgment module is used to establish a quantitative standard for signature anomalies; Abnormal signature recognition module, used to analyze multi-dimensional features to improve recognition accuracy; Risk assessment module, used to build models to quantify the degree of risk; Real-time monitoring module, used to obtain timestamps to determine abnormalities; The intervention strategy module is used to take adaptive intervention measures according to the risk level.

[0006] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0007] The present invention discloses a multi-party signature electronic payment method based on the Internet. The method establishes a quantitative standard for signature anomaly judgment by analyzing historical signature data, including abnormal thresholds of signature duration, number of strokes, pressure change and timestamp features. Deep learning technology is used to comprehensively analyze multi-dimensional features such as signature trajectory, speed and timestamp to construct a signature behavior risk assessment model. The present invention obtains the signature timestamps of each participant in real time, determines whether the difference exceeds a preset threshold, and triggers a signature delay abnormality warning and initiates intervention measures if it exceeds. At the same time, considering factors such as transaction amount, user credit score and transaction frequency, an adaptive signature intervention strategy is adopted to automatically adjust the intervention frequency and intensity through a dynamic weight allocation mechanism. The present invention can effectively identify abnormal signature behavior, improve the security and reliability of electronic signatures, and provide more intelligent and accurate risk management and control means for scenarios such as financial transactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 The present invention is a flowchart of an Internet-based multi-party signature electronic payment method.

[0009] Figure 2 A schematic diagram of an Internet-based multi-party signature electronic payment method and system of the present invention.

[0010] Figure 3 This is another schematic diagram of an Internet-based multi-party signature electronic payment method and system of the present invention.

[0011] Figure 4 It is a structural schematic diagram of an Internet-based multi-party signature electronic payment system of the present invention. DETAILED DESCRIPTION

[0012] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] like Figure 1-4 In this embodiment, a multi-party signature electronic payment system based on the Internet may specifically include:

[0014] S101. Obtain historical signature data, use machine learning algorithms to analyze statistical characteristics of signature behavior, obtain quantitative standards for signature anomaly judgment, and determine abnormal thresholds for signature duration, number of strokes, pressure changes, and timestamp characteristics.

[0015] Obtain pre-stored user historical signature data, and analyze the statistical features of the signature data, wherein the statistical features include signature duration, number of strokes, pressure change and timestamp; use support vector machine algorithm to establish a signature anomaly judgment model, determine the abnormal threshold of each statistical feature through training, and obtain a quantitative standard for signature anomaly judgment; obtain the signature data of the current user, and judge whether the signature duration of the signature data is less than a preset duration threshold, and if so, judge it as a signature anomaly; judge whether the number of strokes of the signature data is less than a preset stroke number threshold, and if so, judge it as a signature anomaly; judge whether the pressure change of the signature data exceeds a preset normal pressure change range, and if so, judge it as a signature anomaly; judge the signature Whether the interval between the timestamp of the data and the timestamp of the last signature is less than a preset time interval threshold, if so, it is judged as a signature abnormality; the result of the abnormality judgment is input into the decision tree model for classification, and the authenticity of the signature data is determined according to the classification result; a convolutional neural network is used to extract features of the signature image of the signature data to obtain the deep features of the signature image; the deep features of the signature image are spliced ​​with the statistical features of the signature data to construct a signature authenticity discrimination model; according to the discrimination result of the signature authenticity discrimination model, the signature abnormality judgment threshold is dynamically adjusted; a personal signature feature model is established for each user, and a personalized judgment threshold is set in combination with the user's personal signature habits when making abnormality judgments.

[0016] Specifically, first, the pre-stored user historical signature data is obtained, and the statistical characteristics of the signature behavior are analyzed. When the user uses an electronic device to sign, the system will record various data in the signature process, such as the signature start time, end time, stroke coordinate sequence, stroke pressure change and other information. By analyzing these data, the user's signature features can be extracted. For example, the historical signature data of a user shows that the signature duration is usually between 3 and 5 seconds, the average number of strokes is 12, and the pressure change range is between 500 and 1000. At the same time, the timestamp of each signature can be recorded for subsequent analysis of the consistency of the signature. Then, the support vector machine algorithm is used to establish a signature anomaly judgment model. In this scenario, the signature data can be divided into two categories: normal signatures and abnormal signatures. Through training, the abnormal threshold of each feature can be determined, and the quantitative standard for signature anomaly judgment can be obtained. For example, according to the user's historical signature data, the threshold of the signature duration can be set to 2 seconds and 6 seconds, that is, if the signature duration is less than 2 seconds or greater than 6 seconds, the signature is considered abnormal. Similarly, the threshold of the number of strokes can be set to 8 and 16, and the threshold of the pressure change range can be set to 300 to 1200. The setting of these thresholds needs to be adjusted according to the actual situation to achieve the best judgment effect. After receiving the user's signature data, an abnormality judgment is performed. First, it is determined whether the signature duration is less than the preset duration threshold. If it is less than, it is determined that the signature is abnormal. For example, if the duration threshold is set to 2 seconds, and the user's signature duration is only 1.5 seconds, the signature is considered abnormal. This is because a too short signature duration may mean that the user did not write seriously and there is a possibility of forgery. Then, it is determined whether the number of strokes is less than the preset stroke number threshold. If it is less than, it is determined that the signature is abnormal. For example, if the stroke number threshold is set to 8 strokes, and the user's signature has only 5 strokes, the signature is considered abnormal. Too few strokes may indicate that the user did not write his signature completely. Next, it is determined whether the pressure change exceeds the preset normal pressure change range. If it exceeds, it is determined that the signature is abnormal. For example, if the pressure change range is set to 500 to 1000, and the pressure change during the user's signing process exceeds this range, the signature is considered abnormal. Abnormal pressure changes may indicate that the user has unnatural pauses or jitters during the signing process, which may be a feature of a forged signature. Finally, determine whether the interval between the timestamp of the current signature and the timestamp of the previous signature is less than the preset time interval threshold. If so, the signature is considered abnormal. For example, if the time interval threshold is set to 5 minutes, and the user signs twice in succession within 2 minutes, the latter signature is considered abnormal. Too short a time interval may indicate that the user has made multiple signature attempts in a short period of time, and there is a risk of forged signatures. The results of the above abnormal judgment are input into the decision tree model for classification, and the authenticity of the signature is determined based on the classification results. Decision tree is a commonly used classification algorithm that achieves classification by constructing a series of judgment rules.In this scenario, the abnormal judgment result can be used as the input feature of the decision tree, such as whether the signature duration is abnormal, whether the number of strokes is abnormal, etc., and the authenticity of the signature is used as the output category. Through training, the decision tree can learn how to judge the authenticity of the signature based on the abnormal judgment result. If it is judged to be a real signature, the signature is accepted; if it is judged to be a forged signature, the signature is rejected and the abnormal feature information of the signature is recorded. For example, through the decision tree model judgment, it is found that the duration and number of strokes of a certain signature are abnormal, and the final classification result is a forged signature. The system will reject the signature and record the duration and number of strokes of the signature for subsequent analysis. In order to further improve the accuracy of signature authenticity judgment, a convolutional neural network is used to extract features from the signature image to obtain the deep features of the signature image. The convolutional neural network is a deep learning algorithm that is good at processing image data. By performing convolution operations and pooling operations on the signature image, feature information in the image can be extracted, such as the shape, thickness, and tilt angle of the strokes. These feature information can reflect the user's signature habits and personalized features. The image depth features are combined with the previously obtained signature behavior statistical features to construct a signature authenticity discrimination model. For example, the feature vector extracted from the signature image through a convolutional neural network can be combined with statistical features such as signature duration and number of strokes to form a more comprehensive feature vector for training the signature authenticity discrimination model. The advantage of this is that image information and behavior information can be combined to more accurately reflect the user's signature features. According to the discrimination results of the signature authenticity discrimination model, the signature abnormality judgment threshold is dynamically adjusted. For example, after the system has been running for a period of time, statistics show that the proportion of real signatures judged as abnormal signatures is relatively high, such as reaching 10%, which means that the current abnormal threshold setting is too strict, resulting in some real signatures being misjudged. In order to improve the system's recall rate, that is, the ability to identify more real signatures, the abnormal judgment threshold can be appropriately relaxed, such as adjusting the lower limit of the signature duration from 2 seconds to 1.5 seconds. On the contrary, if statistics show that the proportion of forged signatures judged as real signatures is relatively high, such as also reaching 10%, it means that the current abnormal threshold setting is too loose, resulting in some forged signatures not being recognized. In order to improve the judgment accuracy of the system, that is, the accuracy of identifying forged signatures, the abnormal judgment threshold can be appropriately tightened, such as adjusting the upper limit of the signature duration from 6 seconds to 5.5 seconds. By dynamically adjusting the abnormal judgment threshold, the system can better adapt to changes in user signing habits and improve the overall performance of the system. In addition, a personal signature feature model can be established for each user, and the normal distribution range of characteristics such as the duration, number of strokes, and pressure changes of the user's signature can be counted. When making abnormal judgments, a personalized judgment threshold can be set based on the user's personal signature habits. For example, for a certain user, the signature duration is usually between 4 and 5 seconds. If the threshold is set according to a unified standard, it may not be possible to accurately judge the abnormality of the signature.Therefore, a personalized signature duration threshold can be set for the user, such as 3.5 seconds and 5.5 seconds. The advantage of this is that it can more accurately capture the subtle changes in the user's signature features and improve the accuracy of signature anomaly detection.

[0017] S102. According to the signature anomaly judgment quantitative standard, obtain the multi-dimensional features of signature trajectory, speed and timestamp, use deep learning technology to perform comprehensive analysis on the multi-dimensional features, and obtain the accuracy rate of abnormal signature recognition.

[0018] Acquire signature time series data, the time series data at least including signature coordinates and timestamp; extract signature trajectory and movement speed from the time series data; calculate the time interval and displacement between adjacent sampling points to obtain the sampling point velocity vector; for the signature trajectory, use discrete Fourier transform to obtain the frequency domain characteristics of the signature; by analyzing the high-frequency and low-frequency components in the frequency domain characteristics, calculate the ratio of high-frequency energy to low-frequency energy, and use it as the signature smoothness feature, the larger the smoothness feature value, the smoother the signature trajectory; according to the signature smoothness feature, use the local polynomial fitting method to calculate the curvature of the signature trajectory, obtain the curvature value of each position of the signature, and form a curvature change sequence, the curvature change is drastic, indicating that the signature trajectory has a large degree of curvature; for the movement speed, calculate the modulus of each velocity vector as the velocity magnitude, and the direction angle as the velocity direction, obtain the velocity scalar sequence and the velocity direction sequence, reflecting the change of velocity during the signing process; use a one-dimensional convolutional neural network to extract the signature trajectory characteristics and velocity respectively. The method comprises the following steps: inputting the coordinate sequence, the speed scalar sequence, and the speed direction sequence into a one-dimensional convolution layer respectively, extracting local features through convolution operation, and then performing feature compression through a pooling layer, flattening and splicing each feature to form a deep feature representation of the signature; connecting the deep feature representation of the signature with the corresponding timestamp information to form a fused feature vector sequence; inputting the fused feature vector sequence into a long short-term memory recurrent neural network, and learning the temporal dependency of the feature sequence through the long short-term memory recurrent neural network; connecting a fully connected layer and a Softmax layer to the output of the last time step of the long short-term memory recurrent neural network to classify and judge the authenticity of the signature; using real signature data and various counterfeit signature data to train the long short-term memory recurrent neural network model, and continuously adjusting the network parameters to improve the classification performance of the model; in the testing stage, calculating the abnormal probability of the signature to be verified according to the model parameters obtained through training; if the abnormal probability exceeds a preset threshold, it is determined to be an abnormal signature, otherwise it is determined to be a real signature.

[0019] Specifically, when acquiring signature timing data on a touch device, it is first necessary to ensure that the device has high-precision touch sensing capabilities and can accurately record the coordinates and timestamp information during the signature process. For example, assuming that the sampling frequency of a touch device is 100Hz, that is, the coordinate point is recorded every 0.01 seconds, so that enough data points can be captured during the signature process for subsequent analysis. When extracting the signature trajectory and movement speed from the timing data, it can be achieved by calculating the displacement and time interval between adjacent sampling points. For example, suppose there are two adjacent sampling points A and B, the coordinates of A are (x1, y1), and the timestamp is t1; the coordinates of B are (x2, y2), and the timestamp is t2. Then the displacement Δs can be calculated by the Euclidean distance formula, and the time interval Δt is t2-t1. The velocity vector v=Δs / Δt, its modulus is the velocity magnitude, and the direction angle is the velocity direction. When using discrete Fourier transform (DFT) to obtain the frequency domain features of the signature, the coordinate sequence of the signature trajectory can be converted into a frequency domain signal. For example, assuming that the signature trajectory consists of N coordinate points, N frequency domain components can be obtained through DFT. The high-frequency component reflects the detailed changes in the signature trajectory, while the low-frequency component reflects the overall trend. The ratio of high-frequency energy to low-frequency energy can be calculated as the signature smoothness feature. If the smoothness eigenvalue of a signature is 0.8, it indicates that its trajectory is relatively smooth; while the smoothness eigenvalue of another signature is 1.5, it indicates that its trajectory is relatively rough. When calculating the curvature of the signature trajectory by the local polynomial fitting method, a sub-trajectory containing k sampling points can be selected, the polynomial can be used to fit the trajectory, and its curvature value can be calculated. For example, k=5 is selected, and a polynomial function is obtained by fitting. The curvature of each sampling point is calculated to form a curvature change sequence. If the curvature of a certain trajectory changes dramatically, it indicates that the signature trajectory has a large degree of curvature, which may reflect the specific writing habits of the signer. For the movement speed, the modulus and direction angle of each velocity vector are calculated to form a velocity scalar sequence and a velocity direction sequence. For example, assuming that the speed scalar sequence in a signature process is [1.2, 1.5, 1.3], and the speed direction sequence is [30°, 45°, 40°]. These sequences reflect the change in speed during the signature process, which helps to distinguish between real signatures and forged signatures. When using a one-dimensional convolutional neural network (1D-CNN) to extract signature trajectory features and speed features, the coordinate sequence, speed scalar sequence, and speed direction sequence can be input into the convolution layer of the 1D-CNN respectively. For example, assuming that the convolution kernel size is 3, local features are extracted through convolution operations, and then feature compression is performed through the pooling layer. Finally, each feature is flattened and spliced ​​to form a deep feature representation of the signature. The deep feature representation of the signature is connected with the corresponding timestamp information to form a fused feature vector sequence.For example, assuming that the deep feature representation is [0.2, 0.5, 0.3] and the timestamp information is [0.01, 0.02, 0.03], the two are connected to form a fused feature vector sequence [0.2, 0.5, 0.3, 0.01, 0.02, 0.03]. The sequence is input into the long short-term memory recurrent neural network (LSTM), and the temporal dependency of the feature sequence is learned through the LSTM unit. For example, assuming that the LSTM network has 128 units, and the input of each time step is a fused feature vector, the LSTM unit can capture the temporal information in the feature sequence, and finally a fully connected layer and a Softmax layer are connected to the output of the last time step to classify and judge the authenticity of the signature. In the model training stage, real signature data and various counterfeit signature data are used for training, and the network parameters are continuously adjusted. For example, assuming that the training set contains 1000 real signatures and 1000 counterfeit signatures, the classification performance of the model is gradually improved through multiple iterative optimizations. In the testing stage, the abnormal probability of the signature to be verified is calculated based on the model parameters obtained through training. For example, if the probability of anomaly of a signature to be verified is 0.07 and exceeds the preset threshold of 0.05, it is determined to be an abnormal signature; otherwise, it is a real signature. Through the above steps, a signature verification model based on deep learning is constructed by comprehensively utilizing signature trajectory features, speed features, and timing information. The model can effectively distinguish between real signatures and abnormal signatures, and improve the accuracy and reliability of signature verification. For example, in practical applications, the model can increase the accuracy of signature verification from 80% to 95%, significantly reducing the risk of forged signatures. In specific implementation, a personalized signature feature model can also be established according to the user's personal signature habits. For example, the normal distribution range of features such as the duration, number of strokes, and pressure changes of a user's signature is counted, and these personalized features are combined in abnormal judgment to set a personalized judgment threshold to further improve the accuracy of signature anomaly detection. For example, the normal signature duration of a user ranges from 2 to 3 seconds. If the duration of a signature is only 1 second, it can be immediately determined to be an abnormal signature. By dynamically adjusting the signature abnormality judgment threshold, the flexibility and adaptability of the model can be further improved. For example, the proportion of real signatures judged as abnormal signatures within a certain period of time is counted. If the proportion is high, the abnormal judgment threshold is appropriately relaxed; the proportion of forged signatures judged as real signatures is counted. If the proportion is high, the abnormal judgment threshold is appropriately tightened. In this way, the recall rate of signature abnormality judgment can be improved while taking into account the judgment accuracy. In summary, through the application of multi-level feature extraction and deep learning models, combined with personalized signature features and dynamic threshold adjustment, an efficient and accurate signature verification system can be built to effectively prevent the risk of forged signatures.

[0020] S103: Establish a signature behavior risk assessment model based on the abnormal signature recognition accuracy, quantify the risk level of the signature behavior, and obtain a decision basis for the signature intervention strategy.

[0021] A pre-established abnormal signature feature library is obtained, wherein the abnormal signature feature library contains multiple abnormal signature features; a signature image to be identified is obtained; a pattern matching algorithm is used to match the signature image to be identified with the abnormal signature features in the abnormal signature feature library to obtain an abnormal signature identification result; based on the abnormal signature identification result, a confusion matrix is ​​used to calculate the accuracy, precision and recall rate of abnormal signature identification to obtain an identification accuracy value; the identification accuracy value is used as a feature, a support vector machine algorithm is used to construct a signature behavior risk assessment model to obtain signature behavior risk assessment model parameters; based on the signature behavior risk assessment model parameters, a weighted average method is used to calculate the signature behavior risk assessment model. risk degree, and obtain a risk degree value; according to the risk degree value, use the CART decision tree algorithm to generate a signature intervention strategy decision tree to obtain an intervention strategy decision basis; determine whether the risk degree of the signature behavior exceeds a preset threshold, and if so, trigger the signature intervention strategy; determine the signature intervention measures according to the intervention strategy decision basis; according to the signature intervention measures, guide the user to adjust the signature behavior through a graphical user interface; obtain the user's adjusted signature, and calculate the risk degree of the adjusted signature; determine whether the risk degree of the adjusted signature is lower than the preset risk threshold, and if so, end the signature intervention process, if not, return to the signature behavior adjustment step until the risk is reduced to an acceptable level.

[0022] Specifically, according to the pre-established abnormal signature feature library, a pattern matching algorithm is used to identify abnormal signatures in the signature image to obtain abnormal signature recognition results. For example, it is assumed that the feature library contains common forged signature features, such as discontinuous strokes, sudden changes in speed, abnormal pressure distribution, etc. When the user signs on the touch device, the system will capture the image data of the signature in real time and extract the corresponding features. Through the pattern matching algorithm, these features are compared with the abnormal features in the feature library. If it is found that the user's signature features are highly matched with a certain type of abnormal features in the feature library, such as the speed of a certain stroke suddenly increases and the pressure distribution is abnormally concentrated, the system determines that the signature is an abnormal signature. Based on the abnormal signature recognition results, the confusion matrix is ​​used to calculate the accuracy, precision and recall rate of abnormal signature recognition to obtain the recognition accuracy value. For example, it is assumed that in a test, the system recognizes 100 signature samples, of which 80 are real signatures and 20 are forged signatures. Through the confusion matrix, it can be counted that the system correctly recognizes 75 real signatures, correctly recognizes 15 forged signatures, mistakenly recognizes 5 real signatures as forged, and mistakenly recognizes 5 forged signatures as real. Based on these data, we can calculate that the accuracy is (75+15) / 100=0.9, the precision is 15 / (15+5)=0.75, and the recall is 15 / 20=0.75. The final recognition accuracy value is 0.9. Using the recognition accuracy value as a feature, the support vector machine algorithm is used to construct a signature behavior risk assessment model to obtain the risk assessment model parameters. For example, assume that in addition to the accuracy, other features such as the duration and number of strokes of the signature are also considered. Take these features as input and use the support vector machine algorithm for training to obtain the model parameters. Assume that the model parameters are the weight vector W and the bias b. These parameters are continuously optimized through training data so that the model can better predict the risk of signature behavior. According to the risk assessment model parameters, the weighted average method is used to calculate the risk level of the signature behavior and obtain the risk level value. For example, assume that the feature vector of a signature includes an accuracy of 0.9, a duration of 2.5 seconds, and a number of strokes of 15. The weighted average risk value is calculated based on the model parameters. Assuming that the weight vector W is [0.6, 0.3, 0.1] and the bias b is -1, the risk level is 0.6*0.9+0.3*2.5+0.1*15-1=0.54+0.75+1.5-1=1.79. According to the risk level, the CART decision tree algorithm is used to generate a signature intervention strategy decision tree to obtain the decision basis for the intervention strategy. For example, assume that the node splitting basis of the decision tree is the risk level value and the threshold is set to 1.5. If the risk level value is greater than 1.5, it is judged as high risk and intervention measures need to be taken; if it is less than 1.5, it is judged as low risk and no intervention is required. The decision tree generated by the CART algorithm can clarify the intervention strategies at different risk levels.If the risk level of the signature behavior exceeds the preset threshold, the signature intervention strategy is triggered, and the signature intervention measures are determined based on the decision basis. For example, assuming that the preset threshold is 1.5, and the risk level of a certain signature is 1.79, the system triggers the intervention strategy if it exceeds the threshold. According to the decision tree, specific intervention measures are determined, such as prompting the user to re-sign, providing signature samples, etc. According to the signature intervention measures, the user is guided to adjust the signature behavior through the graphical user interface to reduce the risk of the signature behavior. For example, the system displays the user's abnormal signature and prompts "The signature speed is uneven, please refer to the sample for adjustment." At the same time, the interface displays a standard signature sample, and the user can refer to the sample for adjustment. The user re-signs through the interface, and the system evaluates the risk of the adjusted signature in real time until the risk is reduced to an acceptable level. This can not only effectively identify and prevent forged signatures, but also improve the authenticity and reliability of signatures by guiding users to adjust their signature behavior. Through multi-level feature extraction and risk assessment, combined with dynamic intervention strategies, the system can flexibly respond to different situations and significantly improve the accuracy of signature verification and the standardization of user signatures.

[0023] S104, acquiring the signature timestamps of each participant in real time from the Internet, and judging whether the difference between the signature timestamps exceeds a preset threshold according to the abnormal thresholds of the signature duration, number of strokes, pressure change and timestamp characteristics and the signature behavior risk assessment model.

[0024] Acquire signature data of multiple parties, the signature data including signature duration, number of strokes, pressure change and timestamp features; preprocess the signature data, remove noise and outliers, and extract key features; use support vector machine algorithm to build a signature behavior risk assessment model, and use historical signature data for training and optimization; input the preprocessed signature feature data into the signature behavior risk assessment model to obtain an abnormality score for each feature; compare the abnormality score for each feature with a preset abnormal threshold, and mark the features exceeding the threshold as abnormal features; judge the signature behavior risk level as one of the three levels of low, medium and high according to the proportion of the number of abnormal features to the total number of features; extract the timestamp information of the signature data, and calculate the time difference between adjacent signature timestamps; compare the time difference with a preset reasonable time threshold, and mark the time difference exceeding the threshold as abnormal; assign weights to the signature behavior risk level and the time difference abnormality respectively, and obtain a first comprehensive risk score of the signature by weighted average; compare the first comprehensive risk score with a preset risk level threshold.

[0025] Specifically, the signature data of each participant is obtained in real time through the Internet, including signature duration, number of strokes, pressure change, and timestamp. For example, when a user uses the electronic signature system, the system records the signature process and captures the signature duration of 3.2 seconds, the number of strokes is 12, the pressure change range is 0.2 to 0.8 Newtons, and the timestamp is 2023-10-01 14:30:00. These data will be transmitted to the backend server in real time for further analysis. The acquired signature data is preprocessed to remove noise and outliers and extract key features. For example, during the preprocessing process, the system found that there was an abrupt peak in the pressure change data of a certain signature, which may be caused by equipment failure or external interference. Through the smoothing filter algorithm, these outliers are removed to ensure the accuracy of the data. The extracted key features include average signature speed, uniformity of pressure distribution, and stroke coherence. The support vector machine (SVM) algorithm is used to build a signature behavior risk assessment model, and historical signature data is used for training and optimization. Assume that the system has collected 1,000 historical signature samples, of which 500 are real signatures and 500 are forged signatures. Through feature extraction, these samples are converted into multidimensional feature vectors and input into the SVM model for training. SVM effectively distinguishes real signatures from forged signatures by finding the optimal hyperplane to form a risk assessment model. The preprocessed signature feature data is input into the trained SVM model to calculate the abnormality score of each feature. For example, the feature vector of a signature includes an average speed of 0.8 m / s, a pressure distribution uniformity of 0.7, and a stroke continuity of 0.9. These features are input into the SVM model, and the model calculates the abnormality score of each feature based on the weights and biases obtained from the training, such as a speed abnormality score of 0.2, a pressure distribution abnormality score of 0.1, and a stroke continuity abnormality score of 0.05. The abnormality score of each feature is compared with the preset abnormality threshold, and features that exceed the threshold are marked as abnormal features. Assuming the preset abnormal threshold is 0.15, the speed abnormality score of 0.2 will be marked as an abnormal feature, while the pressure distribution and stroke continuity scores do not exceed the threshold and are considered normal features. Count the number of abnormal features, and divide the signature behavior risk level into three levels: low, medium, and high according to the proportion of the number of abnormal features to the total number of features. For example, a signature has 10 features, 2 of which are marked as abnormal features, and the abnormal feature ratio is 20%. According to the preset classification standard, the abnormal ratio of 20% corresponds to the medium risk level. Extract the timestamp information of the signature and calculate the time difference between adjacent signature timestamps. For example, the user signed twice at 14:30:00 and 14:35:00, respectively, with a time difference of 5 minutes. Compare the time difference with the preset reasonable time threshold, and mark the time difference that exceeds the threshold as abnormal. Assuming the preset reasonable time threshold is 10 minutes, the time difference of 5 minutes is considered normal, and if the time difference exceeds 10 minutes, it is marked as abnormal.Different weights are assigned to the risk level of the signature behavior and the abnormal time difference, and the first comprehensive risk score of the signature is calculated by weighted average. For example, assume that the weight of the risk level is 0.7 and the weight of the abnormal time difference is 0.3. The risk level of a certain signature is medium (score 0.5) and the time difference is normal (score 0), then the first comprehensive risk score is 0.7*0.5+0.3*0=0.35. The first comprehensive risk score is compared with the preset risk level threshold to determine the validity and risk level of the signature, which serves as the basis for subsequent signature verification and business review. Assuming that the preset risk level threshold is 0.4, the first comprehensive risk score of 0.35 is lower than the threshold, the system determines that the signature is valid and the risk is low, and allows the business process to continue; if the score is higher than the threshold, it indicates that the risk is high and further review or intervention measures are required. In this way, the system can monitor and analyze signature behavior in real time, detect abnormal situations in a timely manner, and improve the accuracy of signature verification and business security.

[0026] S105: If the difference between the signature timestamps exceeds a preset threshold, a signature delay abnormality warning is triggered and signature intervention measures are initiated.

[0027] The current signature timestamp and the last signature timestamp are obtained, and the difference between the current signature timestamp and the last signature timestamp is calculated to obtain the signature timestamp difference value; it is determined whether the signature timestamp difference value exceeds a preset threshold; if the signature timestamp difference value exceeds the preset threshold, a signature delay abnormality warning is triggered; according to pre-configured rules, signature intervention measures are initiated, and the signature intervention measures include adjusting the priority of the signature request, increasing the resource allocation of the signature server, and batch processing of the signature requests; historical signature data is obtained, and the historical signature data is trained using a support vector machine algorithm to establish a classification model for signature delay abnormalities; the current signature request data is input into the classification model to determine whether the current signature request belongs to a delay abnormality; according to the classification results, the signature intervention measures are dynamically adjusted to optimize the signature efficiency; historical data of signature delay abnormalities are obtained, and the historical data is analyzed using a decision tree algorithm to obtain the key factors that cause signature delay abnormalities; according to the key factors, the signature process is optimized, the deployment architecture of the signature system is adjusted, and the fault tolerance and scalability of the signature system are improved.

[0028] Specifically, obtaining the current signature timestamp and the previous signature timestamp and calculating the difference between the two timestamps are the basic steps of signature behavior monitoring. For example, a user completed a signature at 14:30:00 and signed the next signature at 14:35:00, and the timestamp difference is 5 minutes. The purpose of this step is to capture the time interval of the signature behavior, so as to determine whether there is an abnormal delay. Comparing the signature timestamp difference value with the preset threshold is the key link to determine the abnormal signature delay. Assuming that the reasonable time threshold preset by the system is 10 minutes, the time difference of 5 minutes is within the normal range and will not trigger an abnormal warning. However, if the time difference is 15 minutes, which exceeds the preset threshold, the system will determine that the current signature has an abnormal delay and automatically trigger a signature delay abnormal warning. When the system triggers the signature delay abnormal warning, the corresponding signature intervention measures are automatically initiated according to the pre-configured rules. For example, the system can adjust the priority of the signature request and upgrade the delayed signature request to a high priority to ensure that it can be processed quickly. In addition, the system can also increase the resource allocation of the signature server, such as temporarily increasing CPU and memory resources, to increase the speed of signature processing. For batch signature requests, the system can use parallel processing to reduce the processing time of a single request. While executing signature intervention measures, the support vector machine (SVM) algorithm is used to analyze signature delay anomalies in real time. The SVM algorithm establishes a classification model for signature delay anomalies by training historical signature data. Assume that the system collects 1,000 historical signature samples, of which 500 are normal signatures and 500 are delayed signatures. Through feature extraction, these samples are converted into multidimensional feature vectors and input into the SVM model for training. The SVM effectively distinguishes normal signatures from delayed signatures by finding the optimal hyperplane to form a classification model. The current signature request data is used as input. After classification by the SVM model, it can be judged in real time whether the signature request belongs to the delay anomaly. For example, the feature vector of a signature includes parameters such as request time, server load, and network delay. After being input into the SVM model, the model judges whether the signature request belongs to the delay anomaly based on the weights and biases obtained through training. Based on the classification results, the system can dynamically adjust signature intervention measures, such as further optimizing resource allocation or adjusting the processing order of signature requests. Automatically adjusting the relevant parameters in the signature process based on the real-time analysis results of the SVM algorithm is an important means to optimize signature efficiency. For example, the system can dynamically adjust the timeout of the signature request from the default 30 seconds to 20 seconds based on the real-time analysis results to speed up the signature processing. At the same time, the system can also adjust the cache time of the signature result from 5 minutes to 10 minutes to reduce the processing time of repeated signatures. In addition, the system can also limit the number of concurrent signature requests to avoid delays caused by server overload.Using the decision tree algorithm to analyze the historical data of abnormal signature delays is an effective method to mine the key factors that cause abnormal signature delays. The decision tree algorithm recursively splits the data set to find the main factors that affect signature delays. For example, the system found through analysis that poor network conditions are one of the main factors leading to signature delays, followed by excessive server load and excessive complexity of signature keys. According to these key factors, the system can optimize the signature process, such as enabling backup network channels when the network conditions are poor, or automatically expanding capacity when the server load is too high. Adjusting the deployment architecture of the signature system is an important measure to improve the system's fault tolerance and scalability. For example, the system can adopt a distributed architecture to distribute signature requests to multiple servers for processing to avoid delays caused by single point failures. At the same time, the system can also introduce a load balancing mechanism to ensure load balancing of each server and improve overall processing efficiency. Through these multi-level and multi-dimensional analysis and optimization measures, the system can effectively alleviate abnormal signature delays and ensure that signatures can be completed in a timely manner. Real-time monitoring and dynamic adjustment of signature parameters not only improve signature efficiency, but also improve the intelligence and reliability of the system, providing solid data support for business decisions. Fundamentally avoiding the occurrence of abnormal signature delays and ensuring the reliability and timeliness of signatures are the key to ensuring the smooth progress of business processes. For example, in actual applications, the electronic signature system of a financial institution successfully reduced the signature delay rate from 5% to 1% through the above method, greatly improving customer experience and business efficiency. Through real-time analysis and dynamic adjustment, the system can automatically increase server resources during peak hours to ensure that signature requests can be responded to quickly, avoiding business interruptions caused by delays. For another example, during a promotional activity on an e-commerce platform, the number of signature requests surged. The system found through decision tree algorithm analysis that excessive server load was the main delay factor. The system immediately started the expansion mechanism to increase server resources, and at the same time adjusted the priority of signature requests to ensure that signature requests for important orders were processed first, effectively avoiding signature delays and ensuring the smooth progress of promotional activities.

[0029] S106. Comprehensively assess the payment risk level based on transaction amount, user credit score and transaction frequency factors, adopt an adaptive signature intervention strategy, and automatically adjust the intervention frequency and intensity according to the importance of different factors through a dynamic weight allocation mechanism.

[0030] Acquire transaction data, the transaction data including transaction amount, user credit score and transaction frequency; use preset initial weight value to calculate the second comprehensive risk score through weighted average model; map the second comprehensive risk score to the corresponding risk level according to preset risk level threshold, the risk level including low risk, medium risk and high risk; obtain pre-configured signature intervention strategy library, select corresponding intervention strategy from the signature intervention strategy library according to the risk level; use linear regression model to calculate weight coefficient of each risk factor, the training data of linear regression model includes historical transaction data and its corresponding risk label; use the risk factor weight coefficient to update the weight value in the weighted average model; acquire new transaction data in real time, extract transaction amount, user credit score and transaction frequency, and calculate the risk factor of the second comprehensive risk score through weighted average model; frequency, input the updated weighted average model, calculate the second comprehensive risk score of the current transaction; for the second comprehensive risk score of the current transaction, compare it with the adjusted risk level threshold to determine the risk level of the current transaction; determine the signature intervention probability and the intervention intensity numerical range according to the risk level of the current transaction, and the intervention intensity numerical range determines the signature complexity; if signature intervention is required, generate a random number between 0 and 1, and if the random number is less than the signature intervention probability, trigger the signature process; adjust the signature complexity according to the intervention intensity numerical range; continuously collect transaction data over a period of time, and count the number of transactions at each risk level and the occurrence rate of risk events; use the rule engine to analyze statistical data, and dynamically adjust the risk level threshold and signature intervention strategy parameters based on the analysis results.

[0031] Specifically, obtaining transaction amount, user credit score and transaction frequency data is the first step in risk management. For example, when an e-commerce platform processes a transaction, the system first extracts the transaction amount of 5,000 yuan, the user's credit score of 85 points, and the transaction frequency of 10 times in the past month. These data are the basis for subsequent risk assessment. Using the pre-set initial weight values, the second comprehensive risk score is calculated through the weighted average model. Assume that the initial weights are: transaction amount weight 0.4, user credit score weight 0.3, and transaction frequency weight 0.3. Then the second comprehensive risk score = 0.4*5000+0.3*85+0.3*10=2000+25.5+3=2028.5. This score reflects the comprehensive risk level of the current transaction. According to the pre-set risk level threshold, the second comprehensive risk score is mapped to the corresponding risk level. Assume that the risk level threshold is: low risk (0-1000), medium risk (1001-2000), high risk (2001 and above). The second comprehensive risk score of the current transaction is 2028.5, which is a high risk level. Obtain the pre-configured signature intervention strategy library and select the corresponding intervention strategy from the strategy library according to the risk level of the current transaction. The strategy library contains signature intervention strategies of different complexity, such as SMS verification code, fingerprint recognition, and face recognition. For high-risk transactions, the system selects the most complex signature method - a composite signature method of face recognition plus digital certificate to ensure the security of the transaction. The weight of the risk factor is calculated using a linear regression model. The training data of the linear regression model includes historical transaction data and its corresponding risk labels. Assume that the system has collected 1,000 historical transaction data, each of which includes the transaction amount, user credit score and transaction frequency, as well as the corresponding risk label (low, medium, high). Through linear regression model training, the system obtains new weight coefficients: transaction amount weight 0.5, user credit score weight 0.2, and transaction frequency weight 0.3. Use the calculated risk factor weight values ​​to update the weights in the weighted average model. The new second comprehensive risk score = 0.5*5000+0.2*85+0.3*10=2500+17+3=2520. The update of risk factor weights makes the calculation result of the second comprehensive risk score more accurate and reflects the latest risk trends. Get new transaction data in real time, extract transaction amount, user credit score and transaction frequency. For example, the system monitors a new transaction in real time, the amount is 3,000 yuan, the user credit score is 90 points, and the transaction frequency is 5 times. Enter these data into the weighted average model with updated weights to calculate the second comprehensive risk score of the current transaction = 0.5*3000+0.2*90+0.3*5=1500+18+1.5=1519.5. For the second comprehensive risk score of the current transaction, compare it with the adjusted risk level threshold to determine the risk level of the current transaction.Assume that the adjusted risk level threshold is: low risk (0-1500), medium risk (1501-2500), high risk (2501 and above). Then the second comprehensive risk score of the current transaction is 1519.5, which belongs to the medium risk level. Determine the signature intervention probability and intervention intensity numerical range according to the current risk level. For example, the signature intervention probability of medium-risk transactions is 30%, and the intervention intensity numerical range is 1-5. The intervention intensity numerical range determines the signature complexity. If the preset intensity is 3, the signature method of fingerprint recognition plus SMS verification code is adopted. If signature intervention is required, a random number between 0 and 1 is generated. For example, the system generates a random number of 0.25, which is less than the signature intervention probability of 30%, and the signature process is triggered. Adjust the signature complexity according to the preset intervention intensity numerical range, and adopt the signature method of fingerprint recognition plus SMS verification code. Continuously collect transaction data over a period of time, and count the number of transactions at each risk level and the occurrence rate of risk events. For example, the system counted 10,000 transactions in a month, of which 6,000 were low-risk transactions, 3,000 were medium-risk transactions, and 1,000 were high-risk transactions. The risk event occurrence rates were: low risk 1%, medium risk 5%, and high risk 10%. The rule engine is used to analyze the statistical data. Based on the analysis results of the rule engine, the risk level threshold and signature intervention strategy parameters are dynamically adjusted. For example, if the high-risk level risk event occurrence rate exceeds the preset threshold of 8%, the risk level threshold is adjusted so that more transactions are assessed as high risk. The new risk level threshold may become: low risk (0-1200), medium risk (1201-2200), high risk (2201 and above). Through this dynamic adjustment mechanism, the system can more accurately identify high-risk transactions and take corresponding signature intervention measures in a timely manner, thereby effectively reducing the probability of risk events and ensuring the security and reliability of transactions. This method not only improves the efficiency of risk management, but also improves the intelligence and adaptability of the system, providing solid data support for business decision-making.

[0032] It will be apparent to those skilled in the art that the present application is not limited to the details of the exemplary embodiments described above, and that the present application can be implemented in other specific forms without departing from the spirit or essential features of the present application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present application. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An Internet-based multi-party signature electronic payment method, characterized in that: The method comprises: Obtain historical signature data, use machine learning algorithms to analyze the statistical characteristics of signature behavior, obtain quantitative standards for signature anomaly judgment, and determine the abnormal thresholds of signature duration, number of strokes, pressure changes, and timestamp characteristics; According to the signature abnormality judgment quantitative standard, the multi-dimensional features of signature trajectory, speed and timestamp are obtained, and the multi-dimensional features are comprehensively analyzed using deep learning technology to obtain the abnormal signature recognition accuracy; According to the abnormal signature recognition accuracy, a signature behavior risk assessment model is established to quantify the risk level of the signature behavior and obtain a decision-making basis for the signature intervention strategy; Acquire the signature timestamps of each participant in real time from the Internet, and determine whether the difference between the signature timestamps exceeds a preset threshold based on the abnormal thresholds of the signature duration, number of strokes, pressure change and timestamp characteristics and the signature behavior risk assessment model; If the difference between the signature timestamps exceeds a preset threshold, a signature delay abnormality warning is triggered and signature intervention measures are initiated; Based on the transaction amount, user credit score and transaction frequency factors, the payment risk level is comprehensively assessed, and an adaptive signature intervention strategy is adopted. Through a dynamic weight allocation mechanism, the intervention frequency and intensity are automatically adjusted according to the importance of different factors.

2. The method according to claim 1, characterized in that The method of obtaining historical signature data, analyzing the statistical characteristics of signature behavior using a machine learning algorithm, obtaining a quantitative standard for judging signature anomalies, and determining the abnormal thresholds of signature duration, number of strokes, pressure changes, and timestamp characteristics includes: Obtaining pre-stored historical signature data of a user, and analyzing statistical features of the signature data, wherein the statistical features include signature duration, number of strokes, pressure change, and timestamp; The support vector machine algorithm is used to establish a signature anomaly judgment model. The abnormal threshold of each statistical feature is determined through training to obtain the quantitative standard for signature anomaly judgment. Obtain the signature data of the current user, and determine whether the signature duration of the signature data is less than a preset duration threshold. If so, determine that the signature is abnormal; Determine whether the number of strokes of the signature data is less than a preset stroke number threshold, and if so, determine that the signature is abnormal; Determine whether the pressure change of the signature data exceeds a preset normal pressure change range, and if so, determine that the signature is abnormal; Determine whether the interval between the timestamp of the signature data and the timestamp of the last signature is less than a preset time interval threshold, and if so, determine that the signature is abnormal; Inputting the result of the abnormality judgment into a decision tree model for classification, and determining the authenticity of the signature data according to the classification result; Using a convolutional neural network to extract features from the signature image of the signature data to obtain deep features of the signature image; The depth features of the signature image are combined with the statistical features of the signature data to construct a signature authenticity discrimination model; Dynamically adjusting the signature abnormality judgment threshold according to the judgment result of the signature authenticity judgment model; A personal signature feature model is established for each user, and a personalized judgment threshold is set based on the user's personal signature habits when making abnormal judgments.

3. The method according to claim 1, characterized in that The method of obtaining the multi-dimensional features of signature trajectory, speed and timestamp according to the signature abnormality judgment quantitative standard, and using deep learning technology to comprehensively analyze the multi-dimensional features to obtain the abnormal signature recognition accuracy rate includes: Acquire signature timing data, wherein the timing data at least includes signature coordinates and a timestamp; Extracting signature trajectory and motion speed from the time series data; Calculate the time interval and displacement between adjacent sampling points to obtain the velocity vector of the sampling points; For the signature trajectory, a frequency domain feature of the signature is obtained by using discrete Fourier transform; By analyzing the high-frequency and low-frequency components in the frequency domain features, the ratio of high-frequency energy to low-frequency energy is calculated and used as a signature smoothness feature. The larger the smoothness feature value, the smoother the signature trajectory. According to the signature smoothness feature, a local polynomial fitting method is used to calculate the curvature of the signature trajectory, and the curvature value of each position of the signature is obtained to form a curvature change sequence. The drastic change of the curvature indicates that the signature trajectory has a large curvature. For the movement speed, the modulus of each velocity vector is calculated as the velocity magnitude, and the direction angle is calculated as the velocity direction, to obtain a velocity scalar sequence and a velocity direction sequence, which reflect the change of the velocity during the signing process; Using a one-dimensional convolutional neural network to extract the signature trajectory features and speed features respectively; The coordinate sequence, velocity scalar sequence, and velocity direction sequence are input into the one-dimensional convolution layer respectively, and local features are extracted through convolution operation. Then, the features are compressed through the pooling layer, and each feature is flattened and concatenated to form a deep feature representation of the signature. Connecting the deep feature representation of the signature with the corresponding timestamp information to form a fused feature vector sequence; Inputting the fused feature vector sequence into a long short-term memory recurrent neural network, and learning the temporal dependency of the feature sequence through the long short-term memory recurrent neural network; A fully connected layer and a Softmax layer are connected to the output of the last time step of the LSTM RNN to classify and judge the authenticity of the signature; Using real signature data and various counterfeit signature data to train the long short-term memory recurrent neural network model, continuously adjusting network parameters, and improving the classification performance of the model; According to the model parameters obtained through training, the abnormal probability of the signature to be verified is calculated; If the abnormal probability exceeds a preset threshold, it is determined to be an abnormal signature, otherwise it is determined to be a true signature.

4. The method according to claim 1, characterized in that The method of establishing a signature behavior risk assessment model based on the abnormal signature recognition accuracy, quantifying the risk level of the signature behavior, and obtaining a decision basis for the signature intervention strategy includes: Acquire a pre-established abnormal signature feature library, wherein the abnormal signature feature library contains a plurality of abnormal signature features; Obtaining a signature image to be recognized; Using a pattern matching algorithm, the signature image to be identified is matched with the abnormal signature features in the abnormal signature feature library to obtain an abnormal signature identification result; According to the abnormal signature recognition result, the accuracy, precision and recall of the abnormal signature recognition are calculated using a confusion matrix to obtain a recognition accuracy value; The recognition accuracy value is used as a feature, and a support vector machine algorithm is used to construct a signature behavior risk assessment model to obtain signature behavior risk assessment model parameters; According to the signature behavior risk assessment model parameters, a weighted average method is used to calculate the risk level of the signature behavior to obtain a risk level value; According to the risk level value, a CART decision tree algorithm is used to generate a signature intervention strategy decision tree to obtain a decision basis for the intervention strategy; Determine whether the risk level of the signature behavior exceeds a preset threshold, and if so, trigger a signature intervention strategy; Determine signature intervention measures based on the decision basis of the intervention strategy; According to the signature intervention measures, guiding the user to adjust the signature behavior through a graphical user interface; Obtain the user's adjusted signature and calculate the risk level of the adjusted signature; Determine whether the risk level of the adjusted signature is lower than the preset risk threshold. If so, end the signature intervention process. If not, return to the signature behavior adjustment step until the risk is reduced to an acceptable level.

5. The method according to claim 1, characterized in that The method of obtaining the signature timestamps of each participant in real time from the Internet and judging whether the difference between the signature timestamps exceeds a preset threshold according to the abnormal thresholds of the signature duration, the number of strokes, the pressure change and the timestamp characteristics and the signature behavior risk assessment model includes: Acquire signature data of multiple parties, the signature data including signature duration, number of strokes, pressure change and timestamp features; Preprocessing the signature data to remove noise and outliers and extract key features; The support vector machine algorithm is used to build a signature behavior risk assessment model, and historical signature data is used for training and optimization; Inputting the preprocessed signature feature data into the signature behavior risk assessment model to obtain an abnormality score for each feature; The abnormality score of each feature is compared with the preset abnormality threshold, and the features exceeding the threshold are marked as abnormal features; According to the ratio of the number of abnormal features to the total number of features, the risk level of the signature behavior is judged as one of the three levels: low, medium and high; Extracting timestamp information of the signature data, and calculating the time difference between adjacent signature timestamps; Compare the time difference with a preset reasonable time threshold, and mark the time difference exceeding the threshold as abnormal; Assigning weights to the risk level of the signature behavior and the abnormality of the time difference respectively, and obtaining a first comprehensive risk score of the signature by weighted average; The first comprehensive risk score is compared with a preset risk level threshold.

6. The method according to claim 1, characterized in that If the difference between the signature timestamps exceeds a preset threshold, a signature delay abnormality warning is triggered and signature intervention measures are initiated, including: Obtaining a current signature timestamp and a previous signature timestamp, calculating a difference between the current signature timestamp and the previous signature timestamp, and obtaining a signature timestamp difference value; Determine whether the signature timestamp difference value exceeds a preset threshold; If the signature timestamp difference value exceeds the preset threshold, a signature delay abnormality warning is triggered; Initiate signature intervention measures according to pre-configured rules, wherein the signature intervention measures include adjusting the priority of signature requests, increasing resource allocation of signature servers, and batch processing of signature requests; Acquire historical signature data, train the historical signature data using a support vector machine algorithm, and establish a classification model for abnormal signature delays; Inputting the current signature request data into the classification model to determine whether the current signature request belongs to a delay abnormality situation; According to the classification results, dynamically adjust the signature intervention measures to optimize the signature efficiency; Acquire historical data of abnormal signature delays, and use a decision tree algorithm to analyze the historical data to obtain key factors that lead to abnormal signature delays; Based on the key factors, the signature process is optimized, the deployment architecture of the signature system is adjusted, and the fault tolerance and scalability of the signature system are improved.

7. The method according to claim 1, characterized in that The payment risk level is comprehensively assessed based on the transaction amount, user credit score and transaction frequency factors, and an adaptive signature intervention strategy is adopted. Through a dynamic weight allocation mechanism, the intervention frequency and intensity are automatically adjusted according to the importance of different factors, including: Acquiring transaction data, wherein the transaction data includes transaction amount, user credit score and transaction frequency; Using the preset initial weight values, a second comprehensive risk score is calculated through a weighted average model; According to a preset risk level threshold, mapping the second comprehensive risk score to a corresponding risk level, wherein the risk level includes low risk, medium risk and high risk; Obtaining a preconfigured signature intervention strategy library, and selecting a corresponding intervention strategy from the signature intervention strategy library according to the risk level; A linear regression model is used to calculate the weight coefficient of each risk factor, wherein the training data of the linear regression model includes historical transaction data and its corresponding risk labels; Updating the weight value in the weighted average model using the risk factor weight coefficient; Acquire new transaction data in real time, extract transaction amount, user credit score and transaction frequency, input the updated weighted average model, and calculate the second comprehensive risk score of the current transaction; The second comprehensive risk score of the current transaction is compared with the adjusted risk level threshold to determine the risk level of the current transaction; Determine the signature intervention probability and the intervention strength numerical range according to the risk level of the current transaction, and the intervention strength numerical range determines the signature complexity; If signature intervention is required, a random number between 0 and 1 is generated. If the random number is less than the signature intervention probability, the signature process is triggered; adjusting the complexity of the signature according to the numerical range of the intervention intensity; Continuously collect transaction data over a period of time, and count the number of transactions at each risk level and the rate of risk events; The rule engine is used to analyze statistical data and dynamically adjust risk level thresholds and signature intervention strategy parameters based on the analysis results.

8. An Internet-based multi-party signature electronic payment system, characterized in that: The system comprises: A signature anomaly judgment module is used to establish a quantitative standard for signature anomalies; Abnormal signature recognition module, used to analyze multi-dimensional features to improve recognition accuracy; Risk assessment module, used to build models to quantify the degree of risk; Real-time monitoring module, used to obtain timestamps to determine abnormalities; The intervention strategy module is used to take adaptive intervention measures according to the risk level.

Images