Captcha management method and computer system based on multi-modal and reinforcement learning

Abstract

The application relates to the field of communication verification management, in particular to a verification code management method and computer system based on multi-modal and reinforcement learning. After responding to verification information sent by a third-party platform, the method first extracts multi-modal features of the verification information, then calls a preset semantic understanding engine to fuse and analyze the multi-modal features to obtain verification analysis results, finally, according to the verification analysis results, a verification behavior is executed, or the verification analysis results are sent to a user end. Compared with the prior art, the method realizes the automation of the verification behavior and the automation of the verification forwarding, and improves the verification efficiency and security of enterprise-level verification management.

Description

Technical Field

[0001] This invention relates to the field of communication verification management. More specifically, this invention relates to a CAPTCHA management method and computer system based on multimodal and reinforcement learning. Background Technology

[0002] CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) is an interactive security mechanism that distinguishes between a machine and a human user. CAPTCHAs are widely used in website authentication, preventing not only malicious password cracking, vote rigging, and forum spamming, but also effectively preventing hackers from repeatedly attempting to log in to a specific user account using brute-force methods. Currently, to facilitate daily communication for employees (customer service, e-commerce, or sales positions), companies register numerous mobile phone numbers for their respective roles as shared phone numbers. Employees can use these company-assigned numbers to log in to third-party platforms (Taobao, JD.com, or Douyin, etc.) through verification.

[0003] However, the following flaw exists in the process of employees performing the above operations to request and obtain verification codes:

[0004] 1. Inefficient verification management: The existing enterprise verification forwarding process involves employees submitting a verification request, a third-party platform sending the verification code to the enterprise's integrated system, and then manual forwarding it to the corresponding employee. This method leads to a delay in verification response.

[0005] 2. High security risks in verification management: Traditional SMS verification platforms and methods cannot meet enterprise-level privacy protection requirements. During manual forwarding and manual verification, there is a high risk of SMS verification code leakage, and there is a high risk of malicious repeated verification / incorrect input leading to the blacklisting of mobile phone numbers.

[0006] Therefore, existing enterprise-level verification management systems suffer from low verification efficiency and high security risks. Summary of the Invention

[0007] To address the aforementioned technical problems of low verification efficiency and high security risks, this invention discloses a CAPTCHA management method and computer system based on multimodal and reinforcement learning.

[0008] In a first aspect, this invention discloses a CAPTCHA management method based on multimodal and reinforcement learning, comprising:

[0009] In response to verification information issued by a third-party platform, extract the multimodal features of the verification information;

[0010] The preset semantic understanding engine is invoked to fuse and parse the multimodal features, and the verification and parsing results are obtained.

[0011] Based on the verification parsing result, perform the verification action; or, send the verification parsing result to the user terminal.

[0012] Beneficial Effects: Upon receiving verification information from a third-party platform, the method of this invention first extracts the multimodal features of the verification information, then calls a semantic understanding engine to fuse and parse the multimodal features to obtain the verification parsing result, and automatically executes the verification behavior or automatically forwards and distributes the verification code. Compared with existing technologies, the method of this invention automates the verification behavior and verification forwarding, improving the verification efficiency of enterprise-level verification management. The user (employee) is unaware of the initial verification information throughout the process; the entire verification process belongs to the enterprise's internal system verification processing, requiring no manual forwarding or input. This verification method solves the problem of high security risks in enterprise verification management.

[0013] Preferably, the multimodal features are fused and analyzed to obtain the verification analysis results, including:

[0014] Configure the modal weights corresponding to the multimodal features;

[0015] Call the extraction function corresponding to the multimodal features;

[0016] The modal weights and extraction functions are inserted into a preset multimodal fusion function to calculate the recognition confidence.

[0017] The identification confidence and multimodal features are fed into a semantic understanding engine based on the Transformer architecture to obtain the verification results.

[0018] Beneficial effects: The method of this invention uses a semantic understanding engine based on the Transformer architecture to parse multimodal features, which can improve the recognition rate of CAPTCHAs. Compared with existing technologies, the method of this invention can dynamically adjust weights based on the semantic understanding engine and supports the parsing of logical expressions.

[0019] Furthermore, the multimodal fusion function is specifically as follows:

[0020]

[0021] In the formula, Indicates a given under conditions The confidence level of identification, Indicates the first Feature extraction function for each modality express function, Indicates the bias term. Indicates the first Weights of each modality.

[0022] Furthermore, the modal weights include at least text weights, image weights, and speech weights; wherein, the image weight is greater than the speech weight and less than the text weight, and the sum of the text weight, image weight, and speech weight is 1.

[0023] Beneficial effects: The method of the present invention can adjust the verification weight of the initial verification information from multiple dimensions such as text, image and voice, thereby improving the success rate of CAPTCHA parsing.

[0024] Preferably, before extracting the multimodal features of the verification information in response to the verification information issued by the third-party platform, the method of the present invention further includes:

[0025] In response to the user's verification request, a mobile phone number for information verification is allocated to the user from a preset number pool according to a preset number resource scheduling model.

[0026] A verification request is sent to a third-party platform using the phone number as the verification method.

[0027] Compared to existing technologies, the method of this invention can automatically assign mobile phone numbers to users based on their verification needs, eliminating the need for users to use their own mobile phone numbers for verification and login. On the one hand, this achieves the separation of public and private mobile phone numbers, reducing the inconvenience caused to personal information by enterprise operations; on the other hand, it is beneficial to enterprise information management.

[0028] The specific number resource scheduling model is as follows:

[0029]

[0030] In the formula, This indicates the actual allocation parameters of the mobile phone number. Indicates in Find the minimum value within the range and output the corresponding value. A function of value Indicates the first Current task volume for each mobile phone number Indicates the first Credit score for a mobile phone number. Represents the smoothing coefficient. This indicates the total number of mobile phone numbers. This indicates the allocation parameters for the current mobile phone number. This indicates the allocation parameters for the current mobile phone number. This represents taking the square of the 2-norm.

[0031] Compared with existing technologies, the number resource scheduling scheme of the present invention can flexibly allocate available mobile phone numbers to users based on the current task volume, reputation score and total number of numbers in the number pool, thereby improving verification efficiency.

[0032] Preferably, the method further includes performing verification actions based on the verification parsing results; or sending the verification parsing results to the user terminal.

[0033] A preset signal strength monitoring module is used to assess the communication strength of a mobile phone number under the current operator in real time.

[0034] If the communication strength is below the communication threshold, the current operator of the mobile phone number will be switched to the backup operator.

[0035] Preferably, before extracting the multimodal features of the verification information in response to the verification information issued by the third-party platform, the method of the present invention further includes:

[0036] A pre-defined fault prediction model is used to predict the fault probability of the corresponding SIM card in a pre-defined number pool.

[0037] If the probability of failure of the corresponding SIM card is greater than the failure threshold, the self-recovery mechanism of the system or device to which the corresponding SIM card belongs will be triggered.

[0038] Preferably, after sending the verification and parsing results to the user terminal, the method of the present invention further includes:

[0039] Calculate the user's operation anomaly degree based on the user's operation records;

[0040] Operational anomaly rate is used as an indicator parameter for the security audit management of the number pool.

[0041] Secondly, the present invention also discloses a computer system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the CAPTCHA management method based on multimodal and reinforcement learning described in the first aspect is implemented.

[0042] The beneficial effects of this invention are as follows:

[0043] (1) Compared with the prior art, the method of the present invention realizes the automation of verification behavior and verification forwarding, which improves the verification efficiency of enterprise-level verification management. The user (employee) cannot know the initial verification information throughout the process. The entire verification process belongs to the verification processing process of the enterprise internal system. There is no need for manual forwarding or input. This verification method solves the problem of high security risk in enterprise verification management.

[0044] (2) Compared with the prior art, the method of the present invention can adjust the verification weight of the initial verification information from multiple dimensions such as text, image and voice, thereby improving the success rate of CAPTCHA parsing.

[0045] (3) Compared with the prior art, the method of the present invention can automatically assign a mobile phone number to the user terminal according to the user terminal's verification requirements, without requiring the user terminal to use its own mobile phone number for verification and login. On the one hand, it realizes the separation of public and private mobile phone numbers of the user terminal, reducing the trouble caused to personal information by enterprise operations; on the other hand, it is conducive to enterprise information management. Attached Figure Description

[0046] Figure 1 This is a flowchart of the CAPTCHA management method based on multimodal and reinforcement learning in Embodiment 1 of the present invention;

[0047] Figure 2 This is a schematic diagram of the computer system in Embodiment 2 of the present invention;

[0048] Figure 3 This is a framework diagram of the CAPTCHA management system based on multimodal and reinforcement learning in Embodiment 3 of the present invention. Detailed Implementation

[0049] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The specific implementation methods of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment discloses a verification code management method based on multimodal and reinforcement learning, applicable to mobile phone number management and automatic verification code processing in enterprise multi-account registration scenarios. The method includes:

[0052] S10: In response to verification information issued by a third-party platform, extract the multimodal features of the verification information.

[0053] In this embodiment, the third-party platform can refer to domestic platforms such as Taobao, JD.com, Douyin, or Xiaohongshu, which are required for the company's daily operations. If the company has an overseas registration location, the third-party platform can also be an overseas platform, such as YouTube, Google, or Amazon. After the user submits a verification request through the company's internal system, the company system dispatches the mobile phone number and submits the verification request to the third-party platform. Multimodal features include text, image, and voice features. Text features involve textual logic expressions such as "1+3=?", "According to the illustration, please enter the 2nd and 4th digits", or "Find the bicycle in the image", etc. Image features involve images automatically generated by the third-party platform after receiving the verification request, such as "nine-square grid", "six-view view", or "color difference image", etc. Voice features involve language or semantics. Multimodal features can be extracted using a CNN convolutional neural network.

[0054] S20: Call the preset semantic understanding engine to fuse and parse the multimodal features and obtain the verification and parsing results.

[0055] In this embodiment, the semantic understanding engine is based on the Transformer architecture to obtain more accurate verification and parsing results.

[0056] S30: Execute the verification action based on the verification parsing result; or, send the verification parsing result to the user terminal.

[0057] In this embodiment, the user terminal generally refers to the communication devices (such as computers, mobile phones, or iPads) configured by the enterprise for its employees. These user terminals are pre-programmed with scripts by the enterprise to automatically input verification text or automatically click verification icons for the user terminal upon receiving the verification parsing results, thereby achieving automatic execution of the verification behavior. In some special scenarios, if the script is not configured, the verification parsing results can also be directly sent to a specific information receiving interface on the user terminal.

[0058] Through steps S10-S30 above, the method of the present invention automates the verification process and the verification forwarding, improving the verification efficiency of enterprise-level verification management. The intermediate process of verification is a confidential internal process within the enterprise system, and employees cannot maliciously enter incorrect verification codes. This approach solves the problem of high security risks in enterprise verification management.

[0059] Preferably, if the verification information is SMS verification information, features can be extracted using a TLS-encrypted direct connection to the operator's gateway (packet loss rate <0.01%). If the verification information is voice verification information, features are extracted using an ASR engine with WER <5% to achieve mixed Chinese and English recognition. If the verification information is image verification information, features are extracted using a ResNet-50 model (anti-distortion attack accuracy >99%).

[0060] Furthermore, the specific process of step S20 above is as follows:

[0061] S21: Configure the modal weights corresponding to the multimodal features.

[0062] S22: Call the extraction function corresponding to the multimodal features.

[0063] S23: Insert the modal weights and extraction functions into the preset multimodal fusion function to calculate the recognition confidence.

[0064] S24: The identification confidence and multimodal features are put into the semantic understanding engine based on the Transformer architecture to obtain the verification results.

[0065] In this embodiment, the modal weights include text weight, image weight, and speech weight; wherein the image weight is greater than the speech weight but less than the text weight, and the sum of the text weight, image weight, and speech weight is 1. By rationally configuring these multiple weights, the accuracy of the verification and parsing results can be further improved.

[0066] As a preferred approach, the text weight is set to 0.6, the image weight to 0.3, and the speech weight to 0.1.

[0067] More specifically, the aforementioned multimodal fusion function is as follows:

[0068]

[0069] In the formula, Indicates a given under conditions The confidence level of identification, Indicates the first Feature extraction function for each modality express function, Indicates the bias term. Indicates the first Weights of each modality. The range is 0-1.

[0070] It should be explained that after the initial configuration of the multi-dimensional weights, the semantic understanding engine will dynamically adjust the weights based on its understanding of the context and specific modal features.

[0071] Therefore, through the design of steps S21-S24 above, relying on the semantic understanding engine of the Transformer architecture and the aforementioned multimodal fusion function, the method of this invention achieves cross-language parsing of text, image, and voice verification codes. Furthermore, it can improve the recognition accuracy in complex scenarios through contextual analysis (such as automatically marking the validity period of "verification code valid for 5 minutes"). Using the method of this invention, the parsing accuracy can reach 99.2%.

[0072] Furthermore, to further improve verification efficiency, before step S10, the method of the present invention further includes:

[0073] S100: In response to the user's verification request, according to the preset number resource scheduling model, it schedules mobile phone numbers from the preset number pool for the user to use for information verification.

[0074] S200: Sends a verification request to a third-party platform in the name of a mobile phone number.

[0075] It should be explained that a number pool refers to a collection containing a large number of numbers (such as phone numbers, SMS verification code numbers, service numbers, etc.), which can be dynamically allocated, recycled, and reused by the system.

[0076] Furthermore, the aforementioned number resource scheduling model is specifically as follows:

[0077]

[0078] In the formula, This indicates the actual allocation parameters of the mobile phone number. Indicates in Find the minimum value within the range and output the corresponding value. A function of value Indicates the first Current task volume for each mobile phone number Indicates the first Credit score for a mobile phone number. Represents the smoothing coefficient. This indicates the total number of mobile phone numbers. This indicates the allocation parameters for the current mobile phone number. This indicates the allocation parameters for the current mobile phone number. This represents taking the square of the 2-norm.

[0079] Through steps S100-S200 above, the method of the present invention can promptly retrieve available mobile phone numbers for verification based on the actual login needs of employees. During the number resource scheduling process, the adopted number resource scheduling model fully considers the current workload, reputation score, and stability of the mobile phone number, preventing logged-in mobile phone numbers from logging out without cause, thus ensuring seamless connection of mobile phone number scheduling. Furthermore, the method of the present invention can also filter out risky mobile phone numbers (mobile phone numbers blacklisted by operators), thereby improving the efficiency of number resource scheduling and further improving the efficiency of the verification process.

[0080] It should be noted that, generally, the smoothing coefficient mentioned above is set to 0.1 to control allocation volatility. This smoothing coefficient can also be adjusted adaptively based on the employee's job category. For example, if the employee's job is customer service, the smoothing coefficient is set to 0.2. If the employee's job is marketing, the smoothing coefficient is set to 0.05. When the smoothing coefficient is small, the scheduling scheme tends to allocate resources based more on the current number of tasks and the number's reputation score, which may result in larger allocation fluctuations. When the smoothing coefficient is large, the scheduling scheme focuses more on maintaining consistency with previous allocation schemes, reducing allocation volatility, and making the allocation of number resources more stable.

[0081] Furthermore, to avoid verification failure due to SIM card communication malfunction during the verification process, the method of the present invention further includes the following steps before step S10:

[0082] A pre-defined fault prediction model is used to predict the fault probability of the corresponding SIM card in a pre-defined number pool.

[0083] If the probability of failure of the corresponding SIM card is greater than the failure threshold, the self-recovery mechanism of the system or device to which the corresponding SIM card belongs will be triggered.

[0084] Specifically, the algorithm for the above fault prediction model is as follows:

[0085]

[0086] In the formula, Indicates the probability of failure. Indicates the first The product of the probabilities of the fault classes. for a subset of Indicates the first The base probability of this type of fault (if it is signal attenuation), Take 0.2; if it is SIM card oxidation, Take 0.5). This indicates the cumulative number of days for the fault indicator.

[0087] Specifically, when When the value is greater than 0.7, the self-recovery mechanism of the corresponding SIM card's system or device is triggered.

[0088] Compared to existing technologies, the above method can calculate the probability of compound faults, and the predicted fault probability is more reliable. Furthermore, the method can also trigger preventative maintenance, predicting the risk of SIM card oxidation 48 hours in advance.

[0089] In step S30 above, to ensure communication quality and reduce delays during the verification code filling process, the method of the present invention further includes:

[0090] The system uses a pre-set signal strength monitoring module to assess the communication strength of a mobile phone number under the current operator in real time.

[0091] If the communication strength is below the communication threshold, the current operator of the mobile phone number will be switched to the backup operator.

[0092] In this embodiment, the intensity monitoring module is used to calculate the RSRP value, triggering a handover latency of less than 500ms for the three networks (China Telecom, China Mobile, and China Unicom). The communication threshold is -110dBm.

[0093] Furthermore, after sending the verification and parsing results to the user terminal, the method of the present invention further includes:

[0094] Calculate the user's operation anomaly degree based on the user's operation records.

[0095] Operational anomaly rate is used as an indicator parameter for the security audit management of the number pool.

[0096] Specifically, the algorithm for calculating the anomaly degree of user-side operations is as follows:

[0097]

[0098] In the formula, Indicates the degree of operational abnormality. This indicates the verification performance score of the company's employees. This represents the mean of abnormal operation data of company employees. This indicates the annotation difference in abnormal operation data of enterprise employees. An indicator function that indicates whether the operator's blacklist has been accessed. Indicates a penalty item. The operator for taking the L2 norm.

[0099] In this embodiment, the penalty term is set to 0.5.

[0100] Through the above technical solution, the method of the present invention calculates the degree of operational anomaly after the verification behavior is completed, and uses it as an indicator parameter for security audit management, thereby realizing audit traceability, timely identifying abnormal verification behavior, and further reducing the security risks of enterprise verification management.

[0101] Furthermore, in order to control communication latency within a relatively reasonable range while maintaining low maintenance costs in enterprise verification management, this invention introduces a multi-dimensional routing decision model to combine maintenance costs and achieve dynamic channel quality scoring. The specific calculation method is as follows:

[0102]

[0103] In the formula, Represents the State-Action Value Function. This indicates the number of times a verification code has been successfully received in the past. This indicates the total number of historical verification requests. This indicates the preset channel cost threshold. This represents the total cost of the historical verification process. This represents the time interval between the current time and the most recent failure. This represents the first weighting coefficient. This represents the second weighting parameter. This represents the third weighting coefficient. e The base of the logarithm of natural numbers. This represents the attenuation coefficient.

[0104] Preferred, .

[0105] Compared to existing technologies, the above method introduces a cost budget constraint. This allows businesses to set cost thresholds (e.g., cost per CAPTCHA < 0.05 yuan) to control verification costs. A time decay factor has also been introduced. This is to improve the dynamic adaptability of the scoring calculation, so as to adjust the sensitivity of operator strategies (such as gradually reducing the weight of "China Mobile" when it is detected that the operator frequently blocks numbers).

[0106] Example 2

[0107] like Figure 2 As shown, this embodiment discloses a computer system including a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the CAPTCHA management method based on multimodal and reinforcement learning described in Embodiment 1 is implemented.

[0108] Example 3

[0109] like Figure 3 As shown in the figure, this embodiment discloses a CAPTCHA management system based on multimodal and reinforcement learning. The system includes a hardware layer, a data processing layer, a decision layer, an application layer, and a security layer.

[0110] The hardware layer is used to load the physical SIM card, monitor the signal strength of the SIM card, and execute a self-recovery mechanism.

[0111] The data processing layer is used to execute steps S10-S30 as described in Embodiment 1.

[0112] The decision-making layer is used to execute the channel quality dynamic scoring process and number resource scheduling process described in Example 1.

[0113] The application layer is used to implement API integration with enterprise systems and automatically populate validation fields.

[0114] The security layer is used to record the historical operation records of the enterprise system and, based on the historical operation records, execute the security audit management process described in Example 1.

[0115] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A CAPTCHA management method based on multimodal and reinforcement learning, characterized in that, include: In response to verification information issued by a third-party platform, the multimodal features of the verification information are extracted; A preset semantic understanding engine is invoked to fuse and parse the multimodal features to obtain the verification and parsing results; Based on the verification parsing results, perform the verification action; Alternatively, the verification and parsing results can be sent to the user's device; The multimodal features are fused and parsed to obtain the verification and parsing results, including: Configure the modal weights corresponding to the multimodal features; Call the extraction function corresponding to the multimodal features; The modal weights and the extraction function are placed into a preset multimodal fusion function to calculate the recognition confidence. The identification confidence and the multimodal features are fed into a semantic understanding engine based on the Transformer architecture to parse and obtain the verification parsing result; The multimodal fusion function is specifically as follows: In the formula, Indicates a given under conditions The confidence level of identification, Indicates the first Feature extraction function for each modality express function, Indicates the bias term. Indicates the first Weights of each modality.

2. The CAPTCHA management method based on multimodal and reinforcement learning according to claim 1, characterized in that, The modal weights include at least text weights, image weights, and speech weights; wherein the image weights are greater than the speech weights and less than the text weights, and the sum of the text weights, image weights, and speech weights is 1.

3. The CAPTCHA management method based on multimodal and reinforcement learning according to claim 1, characterized in that, Before extracting the multimodal features of the verification information issued by the third-party platform, the method further includes: In response to the user's verification request, a mobile phone number for information verification is allocated to the user from a preset number pool according to a preset number resource scheduling model. A verification request is sent to the third-party platform in the name of the mobile phone number.

4. The CAPTCHA management method based on multimodal and reinforcement learning according to claim 3, characterized in that, The number resource scheduling model is specifically as follows: In the formula, This indicates the actual allocation parameters of the mobile phone number. Indicates in Find the minimum value within the range and output the corresponding value. A function of value Indicates the first Current task volume for each mobile phone number Indicates the first Credit score for a mobile phone number. Represents the smoothing coefficient. This indicates the total number of mobile phone numbers. This indicates the allocation parameters for the current mobile phone number. This indicates the allocation parameters for the previous mobile phone number. This represents taking the square of the 2-norm.

5. The CAPTCHA management method based on multimodal and reinforcement learning according to claim 3, characterized in that, Based on the verification parsing results, the verification action is performed; Alternatively, the verification and parsing results can be sent to the user terminal, and the method further includes: A preset signal strength monitoring module is used to evaluate the communication strength of the mobile phone number under the current operator in real time. If the communication strength is lower than the communication threshold, the current operator of the mobile phone number will be switched to an alternative operator.

6. The CAPTCHA management method based on multimodal and reinforcement learning according to claim 1, characterized in that, Before extracting the multimodal features of the verification information issued by the third-party platform, the method further includes: A pre-defined fault prediction model is used to predict the fault probability of the corresponding SIM card in a pre-defined number pool. If the probability of failure of the corresponding SIM card is greater than the failure threshold, the self-recovery mechanism of the system or device to which the corresponding SIM card belongs will be triggered.

7. The CAPTCHA management method based on multimodal and reinforcement learning according to claim 1, characterized in that, After sending the verification parsing result to the user terminal, the method further includes: Calculate the user's operation anomaly degree based on the user's operation records; The operational anomaly level is used as an indicator parameter for the security audit management of the number pool.

8. A computer system, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the CAPTCHA management method based on multimodal and reinforcement learning as described in any one of claims 1-7 is implemented.

Images