Handwritten signature verification method and electronic equipment

Abstract

The invention discloses a handwritten signature verification method and electronic equipment. The method is applied to the electronic equipment and comprises the following steps: acquiring dynamic behavior data of a current signature of a user; performing multi-level comparison on the dynamic behavior data and a preset dynamic behavior reference model, and calculating a risk score based on a comparison result; and dividing risk levels based on the risk scores, and executing risk control measures corresponding to the risk levels. According to the method, the dynamic behavior characteristics in the writing process of the user are obtained and established, and a multi-level comparison and quantitative risk assessment mechanism is introduced, so that the authenticity of the signature is effectively identified, identity false use risks such as non-personal operation, signing for others and unauthorized swiping are accurately discriminated, and the anti-fraud capability, the calculation efficiency and the service decision reliability of a verification system are improved.

Description

Technical Field

[0001] This invention relates to the field of behavioral feature recognition technology, and in particular to a handwritten signature verification method and electronic device. Background Technology

[0002] Among related technologies, online handwritten signature verification technology has poor security, and signatures are easily misused. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a handwritten signature verification method and electronic device that can effectively identify identity fraud risks such as unauthorized signatures and impersonation, as well as the lack of a fast filtering mechanism for unnatural writing behavior, resulting in low risk control efficiency and response delay.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A handwritten signature verification method, applied to an electronic device, the method comprising: Obtain dynamic behavior data of the user's current signature; The dynamic behavior data is compared with a preset dynamic behavior benchmark model at multiple levels, and a risk score is calculated based on the comparison results. The dynamic behavior benchmark model is used to obtain a risk score based on dynamic behavior. Risk levels are determined based on the risk scores, and risk control measures corresponding to the risk levels are implemented.

[0005] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the various steps of the handwritten signature verification method described above.

[0006] The beneficial effects of this invention are as follows: By identifying dynamic behavioral data during the user's signature process and inputting it into a dynamic behavior benchmark model for real-time, multi-level comparison, a risk score is calculated based on the comparison results, and differentiated risk control measures are implemented. This bypasses the reliance on static signature images in traditional solutions, utilizing dynamic behavioral features during the writing process for identity verification, thereby improving the accuracy of identifying identity fraud. Unlike traditional authentication methods that rely solely on static signature image comparison or simple template matching, this invention achieves a shift from "static morphological verification" to "dynamic behavioral risk control" through multi-level comparison involving natural writing complexity, statistical feature analysis, and fine comparison of feature vector sequences. While effectively improving the ability to identify identity fraud such as proxy signing and fraudulent transactions, the multi-level processing architecture optimizes system computational efficiency, enhancing risk control response speed and decision reliability in real-time scenarios. Attached Figure Description

[0007] Figure 1 A flowchart illustrating the steps of a handwritten signature verification method provided in an embodiment of the present invention; Figure 2 A multi-level risk assessment architecture diagram of a handwritten signature verification method provided in this embodiment of the invention; Figure 3 This invention provides a flowchart of user behavior feature modeling and real-time verification for a handwritten signature verification method. Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0008] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0009] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0010] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0011] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0012] In POS and mobile electronic ticketing systems, users typically need to handwrite their signatures after completing a transaction on their mobile phone, tablet, or POS terminal. However, current electronic ticketing systems only save static signature images (such as PNG / JPEG) for archiving. Online handwritten signature verification technology aims to solve the authentication problem of "signature authenticity." However, when applied to POS, mobile payment, and other electronic ticketing scenarios, this technology reveals its inherent limitations: 1. Inability to effectively identify the risk of "the user and the imposter": Related technologies typically compare the signature to be verified against one or more preset "static templates." Even if the signature itself is "genuine" (i.e., a convincing imitation), the system cannot determine whether the signature was signed by the legitimate user under normal circumstances. For risks such as "fraudulent signature signing" or "signature signed without the user's authorization" after a phone is stolen, these technologies lack proactive identification capabilities, resulting in a deficiency in transaction risk detection.

[0013] 2. Lack of a rapid filtering mechanism for simple forgery: The comparison process in related technologies is often quite complex. For "non-natural handwriting" signatures generated by simple means such as mechanical copying and screen dragging, a complex calculation method is still required to draw a conclusion, resulting in waste of system resources and response delay.

[0014] 3. The template is rigid and fails to reflect the long-term stability of user behavior: It relies on static templates generated by a few signature registrations, which makes it difficult to capture and adapt to subtle changes in users' writing habits. It lacks stability and cannot form a dynamic baseline that represents the user's "behavioral biometrics".

[0015] Therefore, there is a need for a method that can surpass traditional signature verification, proactively identify abnormal user identities at the behavioral level, and perform risk quantification and rating.

[0016] To address the aforementioned issues, this application provides a handwritten signature verification method. This method constructs a dynamic behavioral benchmark model reflecting a user's unique writing habits and employs a multi-level progressive comparison and quantitative risk scoring mechanism to achieve dual and efficient verification of both the authenticity of the signature and the identity of the writer.

[0017] The handwritten signature verification method of this invention is described in detail below, with reference to the appendix. Figure 1 This includes steps 110 to 130.

[0018] Step 110: Obtain dynamic behavioral data of the user's current signature. For example, in a mobile payment scenario, when a user completes a transaction signature on a POS terminal or mobile phone screen, the system uses the signature component of the electronic device to lightweightly and in real-time collect dynamic behavioral data of the current signature trajectory.

[0019] Step 120: Perform multi-level comparisons between the dynamic behavior data and the pre-set dynamic behavior benchmark model, and calculate the risk score based on the comparison results. The dynamic behavior benchmark model is used to obtain the risk score based on dynamic behavior. For example, the acquired dynamic behavior data is compared with the previously pre-set dynamic behavior benchmark model layer by layer, and the corresponding risk score is calculated based on the comparison results. The dynamic behavior benchmark model obtains the risk score through dynamic behavior.

[0020] Step 130: Classify risk levels based on risk scores and implement corresponding risk control measures. For example, the system maps the calculated risk scores to specific risk levels according to preset threshold ranges and triggers differentiated business responses.

[0021] In this way, by constructing and utilizing a user-based dynamic behavioral benchmark model, a fundamental shift has been achieved in signature verification from traditional "static image morphological comparison" to "dynamic biometric behavioral risk control." This approach effectively penetrates the surface of static forgery, accurately identifying identity fraud such as unauthorized signatures and fraudulent transactions, thus elevating signature verification from the traditional "morphological authenticity identification" to a higher dimension of "behavioral identity confirmation." The method introduces a multi-level progressive analysis framework, significantly improving the accuracy of identifying fraudulent activities such as identity theft and forged signatures, as well as the system's anti-fraud capabilities. Furthermore, by optimizing computational resource allocation through layered processing and using a pre-filtering layer to efficiently intercept low-quality, simple forgeries, it avoids unnecessary complex calculations, thereby optimizing overall system resource utilization and response efficiency. This makes the method particularly suitable for high-frequency business scenarios such as mobile payments and electronic invoice signing, which have stringent requirements for security and real-time performance, providing a solid technical foundation for achieving reliable, efficient, and intelligent proactive risk control.

[0022] In one embodiment of this application, step 121 is included before step 120.

[0023] Step 121: Based on the dynamic behavior data corresponding to users' historical valid signatures, construct a natural writing complexity threshold model, a single statistical feature model, and a feature vector sequence model. For example, the system constructs the required models based on the collected historical valid signature data of users. These models include: a natural writing complexity threshold model, a single statistical feature model, and a feature vector sequence model. This is for subsequent comparisons. Step 120 includes steps 122 to 126.

[0024] Step 122: Collect dynamic behavior data corresponding to the user's new signature to obtain the signature writing complexity, signature statistical features, and signature feature vector sequence. For example, when a user completes a new transaction signature, the system synchronously collects the dynamic behavior data of their handwriting and obtains the signature writing complexity value, signature statistical features, and signature feature vector sequence of the current signature.

[0025] Step 123: Perform a first-level comparison based on the signature writing complexity and natural writing complexity threshold model. If the signature writing complexity is lower than the threshold set by the natural writing complexity threshold model, the risk level is determined to be high-risk. For example, if the calculated signature writing complexity is 5, while the user's corresponding set signature writing complexity threshold is 15, the signature is determined to be too smooth and does not conform to the characteristics of natural writing, and is directly marked as high-risk, terminating subsequent comparisons.

[0026] Step 124: If the first-level comparison is passed, a second-level comparison is performed based on at least one signature statistical feature of the new signature and the single-statistic feature model to calculate the statistical feature compliance rate. For example, if the first-level comparison is passed, a second-level comparison is performed on at least one signature statistical feature of the new signature and the single-statistic feature model to calculate the statistical feature compliance rate.

[0027] Step 125: Perform a third-level comparison between the signature feature vector sequence of the new signature and the feature vector sequence model to calculate the sequence similarity. For example, extract the feature vector sequence of the current signature and compare it with the data stored in the feature vector sequence model to calculate the sequence similarity.

[0028] Step 126: Generate the final risk score by combining the statistical feature compliance rate and the sequence similarity. For example, the final required risk score is generated based on the calculated statistical feature compliance rate and the sequence similarity.

[0029] In this way, by pre-extracting three benchmarks from historical behavior—a "natural writing complexity threshold," a "single-item statistical feature model," and a "feature vector sequence model"—a rigorous and progressively in-depth analysis process can be executed when faced with new signatures: first, obvious anomalies are quickly screened out using thresholds; then, macro-level habitual conformity is assessed using statistical features; and finally, dynamic behavior consistency is accurately verified through sequence comparison. This structured, multi-level verification mechanism not only improves the accuracy and robustness of identity theft detection but also effectively optimizes the overall system's processing efficiency and real-time response capabilities by placing the most computationally complex time-series comparison at the back end of the process, providing a reliable and efficient identity verification solution for scenarios with high security requirements.

[0030] In one embodiment of this application, step 121 includes steps 1201 to 1203.

[0031] Step 1201: Establish a natural handwriting complexity threshold model based on the statistical values ​​of handwriting curve complexity corresponding to the user's historical valid signatures. For example, the system collects the user's historical valid signatures, calculates the handwriting curve complexity value for each signature, such as based on indicators like the number of inflection points or the degree of curvature change, and then performs statistical analysis on all historical complexity values. This statistical value can be selected as the historical minimum value, or the historical mean and standard deviation can be calculated and then a dynamic threshold is determined according to certain rules, such as subtracting a certain multiple of the standard deviation from the mean, to serve as the user's natural handwriting complexity threshold. This model is used for rapid filtering in subsequent comparisons, aiming to efficiently identify handwriting that does not conform to natural handwriting habits.

[0032] Step 1202: Construct a single statistical feature model based on the mean and standard deviation of the global statistical features corresponding to the user's historical valid signatures. For example, for the collected historical signatures, the system extracts multiple global statistical features (such as total signature completion time, total handwriting trajectory length, average writing speed, etc.), and calculates the mean of each feature in all historical signatures. ) and standard deviation ( By combining these statistics, a single statistical feature model describing the normal fluctuation range of each feature can be constructed. If it falls within the compliant range... If the condition is met, it is considered compliant. The compliance rate of the statistical characteristics is calculated as the output result for this item.

[0033] Step 1203: Construct a feature vector sequence model based on the refined behavioral feature sequences corresponding to the user's historical valid signatures. For example, the system further extracts one or more temporal behavioral feature sequences from historical signatures (such as velocity-time variation sequence v(t), acceleration sequence a(t), and orientation angle variation sequence). (etc.). These sequences can capture the dynamic rhythm, changes in intensity, and transitional habits of a user's writing process. After these temporal sequences are organized and cleaned, they serve as a feature vector sequence template library for the user, providing a standard reference for subsequent fine-grained temporal comparisons and enabling in-depth identity verification from the perspective of dynamic behavior.

[0034] This approach clarifies the construction process of three benchmark models, extracting and quantifying users' writing habits from three dimensions: handwriting morphological complexity, macro-level writing statistical features, and fine-grained temporal behavioral sequences, forming a multi-dimensional, multi-level dynamic behavioral benchmark model. This model construction method expands user signatures from a single image form to a multi-dimensional dynamic feature set, providing a solid and structured data foundation for subsequent high-precision, high-efficiency multi-level comparisons.

[0035] In one embodiment of this application, step 110 includes steps 111 and 112.

[0036] Step 111: Collect the coordinate sequence of the handwriting trajectory during the signature process and the high-precision timestamp sequence corresponding to each coordinate point in the coordinate sequence. For example, when a user makes a handwritten signature on the touch screen of an electronic device, the system captures the movement trajectory of the signature in real time at a preset sampling frequency. Each sampling records a data point containing two-dimensional planar position information, i.e., a coordinate sequence (x, y). At the same time, the system records the precise time point at which each coordinate point was collected, forming a high-precision timestamp sequence t that corresponds one-to-one with the coordinate point. These two sets of original sequences together constitute the underlying data foundation for signature dynamic behavior analysis.

[0037] Step 112: Based on the coordinate sequence and timestamp sequence, calculate a behavioral feature sequence that includes at least a velocity change sequence and an acceleration change sequence. For example, the system uses the coordinate sequence and the timestamp sequence to calculate the displacement difference and time difference between adjacent coordinate points, thereby obtaining the instantaneous velocity of each sampling point and forming a velocity change sequence describing the change of writing speed over time during the signature process. Further, based on the velocity change sequence, the system again calculates the rate of change of velocity to obtain an acceleration change sequence reflecting changes in writing rhythm. Depending on actual needs and algorithm support, other behavioral feature sequences such as orientation angle change sequence and curvature sequence can be further derived and calculated based on the coordinate sequence and timestamp sequence to more comprehensively characterize the user's dynamic writing features.

[0038] In this way, through standardized numerical calculations, core temporal features (such as velocity and acceleration sequences) that can quantitatively characterize the user's writing dynamics are extracted from the original trajectory. This standardized mapping from raw data to feature space not only ensures the structured and comparable nature of the input data upon which subsequent modeling and comparison depend, but also establishes from the source the system's ability to capture and analyze the user's subtle, unique, and difficult-to-imitate dynamic writing behaviors (such as fluctuations in pen stroke rhythm and transient acceleration patterns). This forms the key data preprocessing foundation for the entire multi-level risk identification framework to achieve the paradigm shift from static image comparison to dynamic behavior risk control.

[0039] In one embodiment of this application, step 124 includes steps 1241 to 1243.

[0040] Step 1241: Obtain the signature feature vector sequence corresponding to each signature feature vector, and perform a third-level comparison with the feature vector sequence model corresponding to the signature feature vector to obtain the single sequence similarity. For example, the system extracts the signature feature vector sequence (such as velocity sequence, acceleration sequence, orientation angle change sequence, etc.) corresponding to the feature vector of the current new signature, and compares it one by one with the corresponding type of reference sequence template stored in the dynamic behavior benchmark model. For each type of sequence, the system uses an algorithm that can effectively handle the time sequence length and local velocity differences to calculate the matching degree between the current sequence and the reference sequence, and quantifies the matching degree as a single sequence similarity score between 0 and 1. For example, the matching result of the velocity sequence may obtain a similarity score of 0.66, the acceleration sequence 0.83, and the orientation angle change sequence 0.50.

[0041] Step 1242: Calculate the sequence similarity based on a weighted fusion of the single sequence similarity scores for each signature feature vector. For example, after obtaining the single sequence similarity scores for each signature feature vector sequence (such as velocity, acceleration, and change in direction angle) corresponding to various feature vectors, the system fuses these scores according to a preset weight configuration. The weights can be set based on the differences in the importance of different types of temporal features to identity recognition. Through weighted fusion, the system ultimately generates a sequence similarity score that comprehensively reflects the matching degree of multiple dynamic behavioral features. For example, if an average weight is used, averaging the three scores yields a comprehensive sequence similarity of approximately 0.663.

[0042] Step 1243: Based on the comprehensive sequence similarity score and the statistical feature compliance rate, a risk score is calculated through weighted fusion. For example, the system performs a secondary fusion of the obtained comprehensive sequence similarity score and the calculated statistical feature compliance rate. According to a preset fusion strategy, such as assigning different weights to the two results to reflect their relative importance in the final identity determination, the system combines the two to generate a single, comprehensive final risk score. This score aims to quantify the degree of consistency between the current signature behavior and the user's historical behavioral habits across multiple dimensions.

[0043] In this way, by independently calculating the matching scores of multiple temporal features and fusing them at multiple levels, a comprehensive evaluation mechanism capable of fully capturing subtle features of users' dynamic handwriting is constructed. This mechanism not only achieves multi-level information aggregation from single sequences to multiple sequences and from temporal details to macro-statistics, but also enables the scoring system to adapt to the needs of different security scenarios through flexible weight configuration. The resulting quantitative risk score provides a core decision-making basis for the system to achieve accurate and reliable risk level determination and differentiated risk control responses.

[0044] In one embodiment of this application, step 1243 includes step 1244.

[0045] Step 1244:

[0046] Wherein, S_final is the risk score, S_dtw is the comprehensive sequence similarity score, R_dtw is the statistical feature compliance rate, and w1 and w2 are preset weight coefficients, satisfying w1 + w2 = 1. For example, in a specific implementation scenario, S_final represents the final calculated risk score, which is a comprehensive quantitative output value; S_dtw represents the comprehensive sequence similarity score obtained from step 1242, reflecting the overall matching degree between the current signature and the user's historical behavior template in multiple dynamic temporal features; R_dtw represents the statistical feature compliance rate obtained from step 123, reflecting the conformity ratio between the current signature and the user's historical habit range in multiple global statistical features; w1 and w2 are preset weight coefficients, used to adjust the contribution ratio of temporal similarity and statistical compliance rate in the final score, respectively, and satisfying the constraint condition w1 + w2 = 1 to ensure the normalization and interpretability of the score range. For example, the system can assign empirical values ​​(e.g., w1 = 0.7, w2 = 0.3) to w1 and w2 respectively, to reflect the prior knowledge that dynamic temporal features have a higher discriminative power for identity determination. The system can then substitute S_dtw (e.g., 0.663) and R_dtw (e.g., 0.75) into the above formula to calculate the final risk score S_final (e.g., 0.7 × 0.663 + 0.3 × 0.75 = 0.6891).

[0047] In this way, by introducing a clear weighted fusion formula, the two heterogeneous but complementary indicators—temporal similarity and statistical compliance rate—obtained from multi-level comparisons are integrated into a single-dimensional risk score in a configurable and interpretable manner. This formulaic approach not only standardizes and reproducibly makes the scoring calculation process, but more importantly, by adjusting the weight coefficients w1 and w2, the system can flexibly adapt to different emphases on dynamic behavioral characteristics and statistical habitual characteristics in different business scenarios. Thus, while maintaining methodological consistency, it enables the customization and optimization of risk identification strategies.

[0048] In one embodiment of this application, step 130 includes steps 131 to 133.

[0049] Step 131: If the risk score is greater than or equal to the first preset risk threshold, it is determined to be low risk, the new signature is confirmed to be the user's signature, and the current action is authorized to proceed normally. For example, suppose the first preset risk threshold is 0.8. When the system calculates the risk score of the current signature to be 0.85, since this score is greater than or equal to the threshold of 0.8, the system determines that the signature action is low risk and confirms that it was performed by the user. Subsequently, the system will allow the business process associated with this signature to continue to execute normally without any additional intervention.

[0050] Step 132: If the risk score is less than the first preset risk threshold but greater than or equal to the second preset risk threshold, it is determined to be of medium risk. The new signature and related behavior are marked as requiring review, and an additional identity verification process is triggered. For example, the second preset risk threshold is set to 0.5. If the current risk score is 0.65, which is less than 0.8 but greater than or equal to 0.5, the system determines that the signature behavior is of medium risk and is suspicious. At this time, the system will mark the signature and its related transactions or operations as "pending review" and automatically trigger one or more additional identity verification processes, such as sending an SMS verification code to the user's registered mobile phone number for secondary confirmation, or transferring the transaction to the manual risk control console for further review.

[0051] Step 133: If the risk score is less than the second preset risk threshold, it is judged as high risk, a real-time security alert is generated, and the behavior is rejected or blocked. For example, if the current risk score is 0.3, which is below 0.5, the system judges the signature behavior as high risk, which is highly likely to be impersonation or fraud. The system will immediately generate a real-time security alert, which can be notified to risk control personnel, system administrators, or relevant security platforms. At the same time, the system will refuse to execute the current operation associated with the signature and take blocking measures when necessary, such as suspending the service to directly block the service, to prevent possible further losses.

[0052] In this way, by setting clear risk threshold ranges, continuous risk scores are mapped to discrete "low," "medium," and "high" risk levels, with clear and progressive risk control response actions configured for each level. This mechanism not only achieves direct and automated conversion of risk assessment results into specific business operations but also constructs a flexible response strategy: highly trustworthy behaviors are smoothly approved to ensure user experience; questionable behaviors are subject to additional verification steps to balance security and efficiency; and high-threat behaviors are decisively intercepted and immediately alerted to control risk. This decision-making system based on quantitative scoring and tiered thresholds makes the entire risk control process more objective, consistent, and manageable, improving the reliability, response speed, and overall security level of identity verification in scenarios such as electronic invoices and mobile payments.

[0053] In one embodiment of this application, steps 140 to 142 are also included.

[0054] Step 140: Based on the user's historical valid signatures and corresponding dynamic behavior data, construct a dynamic behavior baseline model for the user. For example, the system collects and selects a set of historical valid signature data generated by the user in a trusted scenario as training samples. Using these sample data, the system executes the model building logic in Step 121 above to generate an initial, personalized natural writing complexity threshold model, a single statistical feature model, and a feature vector sequence model for the user. These three sub-models together constitute the user's dynamic behavior baseline model, serving as a benchmark for subsequent comparisons and risk assessments.

[0055] Step 141: After a user completes a valid signature for the first time, an initial baseline model is generated based on the data from the valid signature. For example, when a user successfully completes a transaction and generates a valid signature for the first time in the system, since there is no available historical data, the system uses the dynamic behavior data of this signature as the sole benchmark. Based on this signature data, the system calculates its writing complexity, statistical characteristics, and feature sequence, and uses this as a basis to initially set a complexity threshold, establish the mean and standard deviation of individual statistical characteristics, and use this feature sequence as a sequence template. Thus, the initial baseline model for the user is generated, providing a comparison basis for subsequent initial verification.

[0056] Step 142: After each subsequent valid signature by the user, the dynamic behavior baseline model is retrained and updated based on the dynamic behavior data corresponding to the most recent preset number of signatures in the user's historical valid signatures. For example, the system sets a rolling time window or a fixed number of windows. When a user completes a new valid transaction that the system determines as "signature of the user" based on their risk score, the system includes the dynamic behavior data of this new signature in the historical dataset, and may remove the oldest data to maintain the window size. Subsequently, based on this updated dataset consisting of the most recent preset number of signatures, the system re-executes the model building process described in Step 140, thereby generating an updated dynamic behavior baseline model that reflects changes in the user's recent writing habits.

[0057] In this way, through a complete model lifecycle management process including initialization and periodic retraining, the adaptive construction and dynamic evolution of the user behavior baseline model are achieved. The model initialization mechanism ensures the system's availability in the early stages of user use; while the update mechanism enables the model to continuously absorb the latest valid user signature data, thereby continuously learning and fine-tuning its behavioral feature baseline. This ability to continuously evolve not only allows the model to adapt to subtle changes in individual users' writing habits over time, enhancing the system's long-term robustness and user experience, but also ensures the model's timeliness and discriminative power in anti-fraud identification, effectively countering new attack methods that may be employed by attackers attempting to imitate and learn user signature patterns over a long period.

[0058] In one embodiment of this application, step 120 further includes steps 126 and 127.

[0059] Step 126: Based on the acquired original coordinate sequence and timestamp sequence, calculate the basic behavioral sequence, which includes at least a velocity change sequence and an acceleration change sequence. For example, after obtaining the original handwriting trajectory coordinate sequence (x, y) of the user's signature and its corresponding high-precision timestamp sequence (t), the system calculates the position difference and time difference point by point to obtain the velocity change sequence describing the change in writing speed over time; and further, by differentiating the velocity sequence, obtains the acceleration change sequence reflecting the change in writing rhythm. These calculated sequences constitute the basic behavioral sequence for subsequent feature extraction and are the source for deriving higher-level dynamic features.

[0060] Step 127: Based on the original coordinate sequence, timestamp sequence, and basic behavior sequence, extract dynamic behavior features for comparison. For example, the system comprehensively utilizes the original acquired coordinate sequence, timestamp sequence, and obtained basic behavior sequences such as velocity sequence and acceleration sequence to extract a set of structured features that can be directly used for comparison with the benchmark model.

[0061] Step 127 includes steps 1271 to 1273.

[0062] Step 1271: At least one scalar statistical feature for comparison with a single statistical feature model, and at least one time-series feature sequence for comparison with a feature vector sequence model.

[0063] Step 1272: The single statistical feature model is obtained based on the dynamic behavior data corresponding to the user's historical valid signatures.

[0064] Step 1273: The feature vector sequence model is obtained based on the refined behavioral feature sequence corresponding to the user's historical valid signatures.

[0065] For example, the system performs the following operations: 1. Extract scalar statistical features: Calculate one or more individual numerical features from the above sequence. For example, calculate the total completion time of the signature, the total length of the handwriting trajectory, the average writing speed, the curve complexity, etc. These scalar values ​​will be compared with the normal range of the corresponding features in the individual statistical feature model.

[0066] 2. Extracting Temporal Feature Sequences: Directly select or derive complete temporal data from the base sequence as the comparison sequence. For example, the calculated velocity change sequence and acceleration change sequence can be directly used as the temporal feature sequence; or, the orientation angle change sequence and curvature sequence can be calculated based on the coordinate sequence. These complete sequence data will be used for fine-grained temporal alignment with the corresponding sequence template in the feature vector sequence model.

[0067] In this way, the raw spatiotemporal data is transformed into a well-structured feature set that can be directly used for multi-level comparisons. By clearly distinguishing between scalar statistical features and temporal feature sequences, this process provides suitable input data formats for subsequent multi-level comparisons, ensuring that the comparison process can be carried out efficiently and accurately, and forming a crucial and operable bridge between raw behavioral data and risk score calculation.

[0068] See attached document Figure 2 The following details the application embodiments of this application. This application can apply the above-described solution to high-security scenarios such as electronic invoices, mobile payments, and financial contracts that require real-time risk identification of signature behavior, especially in commercial payment fields such as POS transactions and mobile payments where identity theft risks exist. Taking a certain payment transaction scenario as an example, the steps include: S1. The user completes the transaction and signs an electronic signature on the POS terminal or mobile device. The system acquires the user's dynamic behavioral data (such as coordinate sequence, timestamp sequence, etc.) in real time through the signature acquisition component. This is equivalent to step 110 above.

[0069] S2. The system first performs a first-level complexity threshold filtering. The system calculates the handwriting curve complexity of the current signature and compares it with a preset complexity threshold for the user's natural handwriting. (See attached image) Figure 2 As shown in the flowchart on the left, if the calculated complexity value is lower than a preset threshold (indicating that the handwriting is too smooth and may be mechanically copied), the system directly determines the signature as high-risk and terminates the subsequent process. If the complexity value is higher than the threshold, it proceeds to the next stage. This is equivalent to the first-level comparison in step 120 above.

[0070] S3. Next, the system performs a second-level single-feature statistical comparison. For signatures that pass the first-level filtering, the system extracts multiple global statistical features (such as completion time, total handwriting length, average speed, etc.) and compares them one by one with the corresponding normal range of features in the user's historical behavior model to calculate the statistical feature compliance rate. This is equivalent to the second-level comparison in step 120 above.

[0071] S4. Then, the system performs a third-level fine-grained comparison of the feature vector sequences. The system further extracts one or more temporal behavioral feature sequences (such as velocity sequences, acceleration sequences) from the current signature, and uses algorithms such as Dynamic Time Warping (DTW) to perform a fine-grained comparison with the corresponding sequence templates stored in the user's baseline model, calculating the sequence similarity. This is equivalent to the third-level comparison in step S120 above.

[0072] S5. The system integrates multi-level comparison results to generate risk scores and levels. The system weights and fuses the statistical feature compliance rate obtained from the second level with the sequence similarity obtained from the third level to calculate a comprehensive risk score, and maps it to a specific risk level according to a preset threshold range (e.g., S≥0.8 for low risk, 0.5≤S<0.8 for medium risk, and S<0.5 for high risk). This is equivalent to the score fusion in step 120 and the risk level classification in step 130 above.

[0073] S6. Finally, the system executes control measures corresponding to the risk level. Based on the output risk level, the system automatically triggers differentiated business responses: if it is low risk, the transaction is approved normally; if it is medium risk, the transaction is marked as requiring review and may trigger secondary verification (such as SMS verification code); if it is high risk, a real-time alarm is immediately generated, and the current transaction is rejected. This is equivalent to executing risk control measures in step 130 above.

[0074] Through the above application examples, this invention realizes a handwritten signature verification method, completing a closed-loop process from dynamic behavioral data collection, multi-layered progressive feature comparison, quantitative risk assessment to tiered risk handling. This solution effectively solves the problems of traditional static signature verification, such as its inability to identify identity fraud, lack of rapid filtering mechanisms, and low response efficiency. It improves the accuracy of identity fraud identification in scenarios such as electronic payment and invoice signing. Simultaneously, by optimizing system performance and real-time response capabilities through tiered processing, it provides reliable, efficient, and intelligent proactive risk control for high-security businesses.

[0075] See attached document Figure 3 The following details the application embodiments of this application. This application's solution is applicable to scenarios such as financial payments and electronic contracts that require full lifecycle management of signature behavior, particularly in POS or mobile transactions where continuous learning of user habits and real-time identification of fraudulent activity are necessary. Taking a user's signature verification in a payment transaction as an example, the system executes the following complete process: S11. During system initialization or the user's first use, the system collects multiple valid signature data from the user in normal transactions through the signature behavior collection module, and extracts its dynamic behavioral characteristics (such as statistical values, time series, and handwriting complexity). Based on this historical data, the system constructs a personalized baseline behavioral characteristic model for the user. This model includes three sub-models: a natural handwriting complexity threshold model (setting a complexity filtering threshold), a single statistical feature model (based on the mean), and a single statistical feature model. with standard deviation Define the habitual range) and the feature vector sequence model (store time-series templates such as velocity and acceleration). This model can be continuously updated with subsequent valid user signature data, forming a dynamically evolving user behavior baseline. This is equivalent to steps 121 and 140 to 142 above.

[0076] S12. When a user executes a new transaction signature, the system initiates a real-time verification process. First, the handwriting curve complexity of the signature is calculated, and a complexity threshold from the benchmark model is used for the first level of rapid filtering. If the complexity is below the threshold, it is directly judged as high-risk and the process ends; if it passes, the system proceeds to the second level of single-feature statistical comparison, checking whether the signature conforms to the user's habitual range in terms of macro-statistical characteristics such as completion time and total length, and calculating the compliance rate. Subsequently, the system proceeds to the third level of fine comparison of feature vector sequences, using the Dynamic Time Warping (DTW) algorithm to match the temporal sequences of the signature, such as speed and acceleration, with the sequence template in the benchmark model, and calculate the sequence similarity. This is equivalent to steps 110 to 124 above.

[0077] S13. The system weighted and fused the compliance rate of the second-level statistical features with the sequence similarity of the third level to calculate a comprehensive risk score S. Based on preset thresholds (e.g., S≥0.6 for low risk, 0.5≤S<0.6 for medium risk, and S<0.5 for high risk), the system outputs a clear risk level. This is equivalent to steps 125, 130 to 133 above.

[0078] S14. The system implements differentiated controls based on risk level: (1) Low risk: The dynamic behavior is highly matched with the person, and it is determined to be the person's signature. The process ends normally, the system generates an electronic ticket and completes the transaction.

[0079] (2) Medium risk: The dynamic behavior shows certain deviations and is suspicious. The system triggers an additional secondary verification process, such as sending an SMS verification code or starting facial recognition. If the secondary verification is successful, the process continues; if it fails, it is transferred to the high-risk handling branch.

[0080] (3) High risk: If the dynamic behavior is seriously inconsistent with the user's habits, the system will immediately generate a security alarm, enter the advanced monitoring and auditing branch, record the high-risk event and notify the risk control personnel, and at the same time block the transaction, and the process will end abnormally. This is equivalent to step 130 above.

[0081] Through the above application examples, this application implements a complete closed loop for handwritten signature verification, from self-learning modeling of user behavior baselines to multi-level dynamic behavior comparison, and finally to intelligent risk-based tiered handling based on quantitative scoring. This not only fundamentally changes the traditional authentication model that relies solely on static image comparison, achieving an upgrade to "dynamic biometric risk control," but also ensures the system's long-term adaptability through a modeling update mechanism. It optimizes the system's response efficiency and resource utilization in high-concurrency scenarios such as real-time payments, providing a reliable, accurate, and scalable proactive security protection system for various high-security electronic services.

[0082] In summary, this invention constructs a handwritten signature verification method. By collecting and modeling the dynamic behavioral characteristics of users' daily signatures, a dynamic behavioral benchmark model is constructed, consisting of a natural handwriting complexity threshold model, a single statistical feature model, and a feature vector sequence model. This achieves closed-loop behavioral authentication of user identity, from habit learning to real-time comparison. The system can automatically adapt corresponding comparison strategies and judgment thresholds according to different risk scenarios. For different imitation methods and risk levels, the system can execute differentiated comparison levels and response measures.

[0083] Meanwhile, during the signature verification process, the system employs a multi-level comparison strategy based on the behavioral characteristics of the current signature and the dynamic behavioral benchmark model. It performs complexity filtering, statistical compliance analysis, and fine-grained sequence matching according to different feature types. Based on the varying contributions of different behavioral features to identity verification, the system integrates statistical feature compliance rate and sequence similarity with preset weights to arrive at a final risk score. This method differs from traditional single-image template matching; through multi-dimensional feature-based comprehensive decision-making, it can more accurately identify genuine signatures from unauthorized signatures, significantly improving the accuracy of identity verification.

[0084] Furthermore, during the signature verification process, the system also considers the natural evolution of user habits and the business continuity of model updates, establishing a model adaptive update mechanism. When a user completes a new valid signature, the system can incorporate this data into the training set according to a preset strategy, regenerating a baseline model that is more adapted to the user's current habits, thus achieving a balance between immediate response and long-term adaptability in signature verification.

[0085] This method is applicable to businesses that require real-time identity verification in high-security scenarios such as electronic invoices and mobile payments. By constructing a complete signature verification system from behavior modeling and multi-layer comparison to risk decision-making, it effectively improves the accuracy of identifying identity fraud and the system's anti-fraud capabilities, thereby enhancing the level of proactive business risk prevention and control.

[0086] Please refer to Figure 4 The present invention also provides an electronic device 300, including a memory 301 and a processor 302, and a computer program stored on the memory 301 and running on the processor 302. When the processor 302 executes the computer program, it implements the various steps of the handwritten signature verification method described above.

[0087] The beneficial effects of the electronic device of the present invention are the same as those of the method described above, and will not be repeated here.

[0088] The above are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention's specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A handwritten signature verification method, characterized in that, Applied to electronic devices, the method includes: Obtain dynamic behavior data of the user's current signature; The dynamic behavior data is compared with a preset dynamic behavior benchmark model at multiple levels, and a risk score is calculated based on the comparison results. The dynamic behavior benchmark model is used to obtain a risk score based on dynamic behavior. Risk levels are determined based on the risk scores, and risk control measures corresponding to the risk levels are implemented.

2. The handwritten signature verification method according to claim 1, characterized in that, Before performing multi-level comparisons between the dynamic behavior data and the preset dynamic behavior benchmark model, the following steps are included: Based on the dynamic behavior data corresponding to the user's historical valid signatures, a natural writing complexity threshold model, a single statistical feature model, and a feature vector sequence model are constructed. The step of performing multi-level comparisons between the dynamic behavior data and a preset dynamic behavior benchmark model includes: Collect dynamic behavior data corresponding to the user's new signature to obtain signature writing complexity, signature statistical features, and signature feature vector sequence; A first-level comparison is performed based on the signature writing complexity and the natural writing complexity threshold model. If the signature writing complexity is lower than the threshold set by the natural writing complexity threshold model, the risk level is determined to be a high-risk comparison process. If the first-level comparison is passed, a second-level comparison is performed based on at least one signature statistical feature of the new signature and the single statistical feature model to calculate the statistical feature compliance rate. The signature feature vector sequence of the new signature is compared with the feature vector sequence model at the third level to calculate the sequence similarity. The final risk score is generated by combining the statistical characteristics compliance rate and the sequence similarity.

3. The handwritten signature verification method according to claim 2, characterized in that, The process of collecting dynamic behavior data corresponding to the user's new signature yields signature writing complexity, signature statistical features, and a signature feature vector sequence, including: The natural writing complexity threshold model is set based on the statistical values ​​of the handwriting curve complexity corresponding to the user's historical valid signatures; The single statistical feature model is constructed based on the mean and standard deviation of the global statistical features corresponding to the user's historical valid signatures; The feature vector sequence model is constructed based on the refined behavioral feature sequence corresponding to the user's historical valid signatures.

4. The handwritten signature verification method according to claim 1, characterized in that, The process of obtaining dynamic behavior data of the user's current signature includes: Collect the coordinate sequence of the handwriting trajectory during the signing process and the high-precision timestamp sequence corresponding to each coordinate point in the coordinate sequence; Based on the coordinate sequence and the timestamp sequence, a behavioral feature sequence including at least a velocity change sequence and an acceleration change sequence is calculated.

5. The handwritten signature verification method according to claim 2, characterized in that, The step of performing a third-level comparison between the signature feature vector sequence of the new signature and the feature vector sequence model to calculate the sequence similarity includes: Obtain the signature feature vector sequence corresponding to each signature feature vector, and perform a third-level comparison with the feature vector sequence model corresponding to the signature feature vector to obtain the single sequence similarity. The sequence similarity is calculated by weighted fusion of the single sequence similarity of each signature feature vector; The risk score is obtained by weighted fusion calculation based on the comprehensive sequence similarity score and statistical feature compliance rate.

6. The handwritten signature verification method according to claim 5, characterized in that, The risk score, calculated through weighted fusion based on comprehensive sequence similarity score and statistical feature compliance rate, includes: Wherein, S_final is the risk score, S_dtw is the comprehensive sequence similarity score, R_dtw is the statistical feature compliance rate, and w1 and w2 are preset weight coefficients, satisfying w1 + w2 = 1.

7. The handwritten signature verification method according to claim 1, characterized in that, The process of classifying risk levels based on the risk score and implementing risk control measures corresponding to the risk levels includes: If the risk score is greater than or equal to the first preset risk threshold, it is determined to be low risk, the new signature is confirmed to be the signature of the person in question, and the current action is authorized to be executed normally. If the risk score is less than the first preset risk threshold and greater than or equal to the second preset risk threshold, it is determined to be of medium risk, the new signature and associated behavior are marked as requiring review, and an additional identity verification process is triggered. If the risk score is less than the second preset risk threshold, it is determined to be high risk, a real-time security alarm is generated, and the behavior is rejected or blocked.

8. The handwritten signature verification method according to claim 1, characterized in that, Also includes: Based on the dynamic behavior data corresponding to the user's historical valid signatures, a dynamic behavior baseline model of the user is constructed. After the user completes a valid signature for the first time, an initial baseline model is generated based on the data from the valid signature. After each subsequent valid signature by the user, the dynamic behavior baseline model is retrained and updated based on the dynamic behavior data corresponding to the most recent preset number of signatures in the user's historical valid signatures.

9. The handwritten signature verification method according to claim 1, characterized in that, Before performing multi-level comparisons between the dynamic behavior data and the preset dynamic behavior benchmark model, the method further includes: Based on the collected original coordinate sequence and timestamp sequence, a basic behavior sequence including at least velocity change sequence and acceleration change sequence is calculated; Based on the original coordinate sequence, timestamp sequence, and basic behavior sequence, dynamic behavior features for comparison are extracted. The dynamic behavioral features used for comparison include: At least one scalar statistical feature for comparison with a single statistical feature model, and at least one time-series feature sequence for comparison with a feature vector sequence model; The single statistical feature model is obtained based on the dynamic behavior data corresponding to the user's historical valid signatures; The feature vector sequence model is obtained based on the refined behavioral feature sequence corresponding to the user's historical valid signatures.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement each step of the handwritten signature verification method according to any one of claims 1 to 9.

Images