Anti-fraud auditing system, method and equipment for gift giving behavior of live broadcast platform

Abstract

The invention relates to an anti-fraud auditing system, method and device for a gift giving behavior of a live broadcast platform, and relates to the technical field of anti-fraud auditing and risk management and control. The system comprises a self-adaptive access module, a data acquisition and front management and control module, a risk identification and punishment suggestion module and a man-machine collaborative auditing module. Data interaction with multiple live broadcast platforms is realized based on the adaptive access module; a gift giving relation pair is constructed based on the data acquisition and front management and control module, multi-modal risk control feature data is collected, and transaction funds are pre-frozen; performing cross-modal cross attention fusion, violation tendency score and comprehensive risk index calculation based on a risk identification and punishment suggestion module, and automatically matching suggested punishment levels and generating differentiated pre-freezing and unfreezing strategies; and based on a man-machine collaborative auditing module, performing a final auditing decision and feeding back optimization risk identification and punishment suggestions. The system has efficient, accurate, extensible and self-optimized anti-fraud auditing capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of anti-fraud auditing and risk control, in particular to an anti-fraud auditing system, method and equipment for live broadcast platform gift-giving behavior. BACKGROUND

[0002] With the rapid development of the network live broadcast industry, live broadcast fraud behavior presents a trend of scale, complexity and concealment, which poses a serious threat to platform security and user rights. However, the current anti-fraud auditing technology applied to the gift-giving behavior of the live broadcast industry still has many key defects, which restricts the effectiveness of risk identification and prevention.

[0003] Firstly, in terms of risk identification, the analysis granularity of traditional methods is relatively coarse, usually only detecting single transactions or independent user behavior, lacking in-depth mining of long-term interactive relationships between users, making it difficult to discover systematic and concealed fraud behavior in a timely manner. At the same time, traditional methods mostly rely on a single type of structured data (such as transaction amount, frequency), failing to fully integrate multi-modal information such as text and voice on live broadcast platforms, limiting the accuracy of fraud behavior identification, and existing risk intervention mostly occurs after the fact, lacking real-time interception and pre-control capabilities, making it difficult to avoid financial losses and fraud escape. Secondly, in terms of punishment decision, traditional methods still highly rely on manual experience or static rules, lacking unified and quantitative judgment standards, resulting in inconsistent and lagging punishment results, and failing to effectively incorporate audit results and disposal feedback into the risk identification optimization process, limiting the continuous improvement of automated risk control capabilities.

[0004] The above defects result in that the existing technology has low risk identification efficiency and limited accuracy in dealing with large-scale, multi-platform and high-complexity live broadcast gift-giving behavior, making it difficult to form sustainable and intelligent risk identification and anti-fraud capabilities, restricting its intelligent and sustainable development. SUMMARY

[0005] Therefore, it is necessary to provide an anti-fraud auditing system, method and equipment for live broadcast platform gift-giving behavior in view of the above technical problems.

[0006] An anti-fraud auditing system for live broadcast platform gift-giving behavior, the system comprising: An adaptive access module for providing a standard interface to map the standard audit request of the system in real time to an API call recognizable by the target live broadcast platform, realizing data interaction between the system and the target live broadcast platform; A data acquisition and pre-control module for acquiring the gift-giving behavior of the target live broadcast platform according to the adaptive access module, constructing a gift-giving relationship pair formed by the ID combination of the gift-giving parties, collecting multi-modal risk control feature data of the gift-giving relationship pair within a preset period, and pre-freezing all transaction funds within the period. The risk identification and punishment suggestion module, which is internally provided with a fraud behavior knowledge base, an intelligent scoring engine and a punishment strategy matrix, is used to perform cross-modal cross-attention fusion on the multi-modal risk control feature data according to the intelligent scoring engine, to perform parallel matching calculation on the fused risk control features and various types of violation labels in the fraud behavior knowledge base, to obtain violation tendency scores of each violation label, to select the violation label with the highest score as the main violation reason, and to calculate a comprehensive risk index according to the main violation reason, the cumulative transaction amount and the historical cumulative violation times of the gift-giving relationship pair, to match a suggested punishment level in the punishment strategy matrix and to automatically generate a differentiated pre-freezing and thawing strategy, and finally to generate a to-be-reviewed task by combining the gift-giving relationship pair, the main violation reason, the comprehensive risk index and the suggested punishment level; The human-computer collaborative review module is used to receive the to-be-reviewed task for human-computer collaborative review, and to feed back the final review result and the final punishment level of the gift-giving behavior of the gift-giving relationship pair within a preset period to the target live broadcast platform through the adaptive access module, and to feed back the final review result and the final punishment level to the risk identification and punishment suggestion module, so as to realize adaptive optimization of risk identification and punishment suggestion.

[0007] In one of the embodiments, the adaptive access module includes a standard interface and an interface mapping unit. The standard interface includes: a data reporting and launching interface, which is used for the target live broadcast platform to report the gift-giving behavior and the multi-modal risk control feature data in real time; an information query and tracing interface, which is used for the human-computer collaborative review module to call the deep information of the target live broadcast platform on demand to support the final review result; a risk disposal and execution interface, which is used to send the final review result and the final punishment level to the target live broadcast platform and execute the corresponding punishment behavior after the review is completed; and an event notification and reach interface, which is used to send a notification or an alarm SMS to the target live broadcast platform. The interface mapping unit is used to realize dynamic mapping of the standard interface to the API of the target live broadcast platform, including interface address mapping and request parameter mapping; wherein the interface address mapping is used to point the standard interface to the actual API of the target live broadcast platform, and the request parameter mapping is used to map the standard interface parameter name to the actual parameter name of the target live broadcast platform.

[0008] In one of the embodiments, the multi-modal risk control feature data includes transaction data, user portrait data and text data of the gift-giving parties; wherein the transaction data includes the cumulative transaction amount and the transaction frequency of the gift-giving relationship pair within a preset period; the user portrait data includes the historical violation records, the account level and the real-name authentication information of the gift-giving parties; and the text data includes the chat records of the gift-giving parties and the text information extracted from the audio and video calls of the gift-giving parties.

[0009] In one of the embodiments, the intelligent scoring engine performs cross-modal cross-attention fusion on the multi-modal risk control feature data, performs parallel matching calculation of the fused risk control features and various types of violation labels in the fraud behavior knowledge base, and obtains violation tendency scores of each violation label, including: The intelligent scoring engine first encodes the transaction data, user portrait data and text data in the multi-modal risk control feature data to obtain the multi-modal risk control feature vectors of the gift relationship pair, including transaction features, portrait features and text features; Secondly, a cross-modal cross-attention mechanism is used, any one of the multi-modal risk control feature vectors is taken as a query vector Q, and the other is taken as a key vector K and a value vector V. The cross-attention weight between any two multi-modal risk control feature vectors is calculated according to the query vector Q, the key vector K and the value vector V, and the multi-modal risk control feature vectors after attention weighting are obtained according to the cross-attention weight, including transaction attention weighted portrait features, transaction attention weighted text features, portrait attention weighted text features, portrait attention weighted transaction features, text attention weighted transaction features and text attention weighted portrait features; The multi-modal risk control feature vectors and the multi-modal risk control feature vectors after attention weighting are spliced, dimensionally reduced and residual connected to obtain the fused risk control features of the gift relationship pair. The fused risk control features and the feature templates of various types of violation labels in the fraud behavior knowledge base are calculated in parallel correlation matching. According to the matching degree, the violation tendency scores of each violation label are obtained. Wherein, the violation label is composed of various types of fraud behaviors and corresponding feature templates.

[0010] In one of the embodiments, according to the comprehensive risk index calculated according to the main violation reason of the gift relationship pair, the cumulative transaction amount and the historical cumulative violation times, the recommended penalty level is matched in the penalty strategy matrix and the differentiated pre-freezing unfreezing strategy is automatically generated, including: The confidence of the main violation reason is mapped into a reason confidence risk score, and the cumulative transaction amount and the historical cumulative violation times of the gift relationship pair in the preset period are respectively mapped into an amount risk score, a gift giver historical risk score and a receiver historical risk score according to a configurable, non-linear grading mapping table; The reason confidence risk score, the amount risk score, the gift giver historical risk score and the receiver historical risk score are weighted and summed to obtain the comprehensive risk index of the gift relationship pair, which is expressed as: ; Among them, 、 、 and represent the weight coefficients of each risk score; By calculating the comprehensive risk index comparing with the preset penalty level threshold table of the penalty strategy matrix, obtaining the recommended penalty level of the gift relationship pair in the preset period, and adjusting the pre-freezing proportion and duration of all transaction funds of the gift relationship pair in the preset period according to the interval of the comprehensive risk index comparing with the preset penalty level threshold table of the penalty strategy matrix, obtaining the recommended penalty level of the gift relationship pair in the preset period, and adjusting the pre-freezing proportion and duration of all transaction funds of the gift relationship pair in the preset period according to the interval of the comprehensive risk index

[0011] In one of the embodiments, the man-machine collaborative review module comprises: The task management module is configured to store the to-be-reviewed tasks output by the risk identification and penalty suggestion module, and to perform task priority ranking and distribution according to the size of the comprehensive risk index or the severity of the main violation reason; The man-machine review decision interface comprises: a pre-processing conclusion abstract area configured to display the main violation reason and the recommended penalty level of the to-be-reviewed task; an associated evidence one-key tracing area configured to access the multi-modal risk control feature data of the gift relationship pair corresponding to the to-be-reviewed task, including transaction data, user portrait data and text data of the gift parties; and an audit decision operation area configured to manually review and output the final review result and the final penalty level of the gift behavior of the gift relationship pair in the preset period corresponding to each to-be-reviewed task in combination with the information provided by the pre-processing conclusion abstract area and the associated evidence one-key tracing area. The feedback module is configured to feed back the final review result and the final penalty level to the target live streaming platform through the adaptive access module, and to feed back the final review result and the final penalty level to the risk identification and penalty suggestion module, so as to realize adaptive optimization of risk identification and penalty suggestion by comparing and analyzing the deviation of risk identification and penalty level.

[0012] In one of the embodiments, the feedback module is further configured to, when the final review result is no risk, feed back the no-risk result to the data acquisition and pre-control module and the risk identification and penalty suggestion module, and to unfreeze the currently pre-frozen transaction funds.

[0013] A live streaming platform gift behavior anti-fraud review method, which is implemented based on the live streaming platform gift behavior anti-fraud review system described above, comprises the following steps: Through adaptive access between the system and the target live streaming platform, the gift behavior of the target live streaming platform is obtained, the gift relationship pair formed by the ID combination of the gift parties is constructed, the multi-modal risk control feature data of the gift relationship pair in the preset period is collected, and all transaction funds in the period are pre-frozen; According to the cross-modal cross-attention fusion of the multi-modal risk control feature data by the intelligent scoring engine, the fused risk control features and various types of violation labels in the fraud behavior knowledge base are matched and calculated in parallel to obtain the violation tendency scores of each violation label, the violation label with the highest score is selected as the main violation reason, and the comprehensive risk index is calculated according to the main violation reason of the gift-giving relationship pair, the cumulative transaction amount and the historical cumulative violation times, the recommended punishment level is matched in the punishment strategy matrix, and the differentiated pre-freezing and thawing strategy is automatically generated, and finally the to-be-reviewed task is generated by combining the gift-giving relationship pair, the main violation reason, the comprehensive risk index and the recommended punishment level; The final review result and the final punishment level of the gift-giving relationship pair in the preset period are fed back to the target live broadcast platform, and the final review result and the final punishment level are fed back to the risk identification and punishment suggestion module, so that the adaptive optimization of risk identification and punishment suggestion is realized.

[0014] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor realizes the following steps when executing the computer program: Through adaptive access between the system and the target live broadcast platform, the gift-giving behavior of the target live broadcast platform is obtained, the gift-giving relationship pair formed by the combination of the gift-giving parties ID is constructed, the multi-modal risk control feature data of the gift-giving relationship pair in the preset period is collected, and all transaction funds in the period are pre-frozen; According to the cross-modal cross-attention fusion of the multi-modal risk control feature data by the intelligent scoring engine, the fused risk control features and various types of violation labels in the fraud behavior knowledge base are matched and calculated in parallel to obtain the violation tendency scores of each violation label, the violation label with the highest score is selected as the main violation reason, and the comprehensive risk index is calculated according to the main violation reason of the gift-giving relationship pair, the cumulative transaction amount and the historical cumulative violation times, the recommended punishment level is matched in the punishment strategy matrix, and the differentiated pre-freezing and thawing strategy is automatically generated, and finally the to-be-reviewed task is generated by combining the gift-giving relationship pair, the main violation reason, the comprehensive risk index and the recommended punishment level; The final review result and the final punishment level of the gift-giving relationship pair in the preset period are fed back to the target live broadcast platform, and the final review result and the final punishment level are fed back to the risk identification and punishment suggestion module, so that the adaptive optimization of risk identification and punishment suggestion is realized.

[0015] Compared with the traditional anti-fraud auditing system, the anti-fraud auditing system, method and device for live broadcast platform gift-giving behavior can realize seamless connection and data interaction with multiple live broadcast platforms based on the adaptive access module, support rapid access of multiple live broadcast platforms, and guarantee system universality and scalability; the gift-giving relationship pair is constructed based on the data acquisition and pre-control module, and multi-modal risk control feature data is collected, avoiding misjudgment of single transaction or independent user behavior, accurately identifying long-term and systematic fraud behavior, improving the accuracy and efficiency of risk identification, and ensuring the safety of funds by pre-freezing transaction funds of the gift-giving relationship pair, improving the initiative of risk control; the multi-modal risk control feature data is cross-modal cross-attention fusion based on the risk identification and penalty suggestion module, which is conducive to learning the complex nonlinear relationship between features of different modalities under different fraud behaviors, improving the comprehensiveness and explainability of violation identification, accurately pointing out the main violation reasons, further calculating a comprehensive risk index by multi-factor quantization and matching the recommended penalty level, providing reliable auditing decision basis for artificial auditing, reducing the dependence on artificial experience, avoiding inconsistent or lagging penalty standards, improving the auditing efficiency, and automatically generating a differentiated pre-freezing and thawing strategy based on the comprehensive risk index to avoid one-size-fits-all long-term freezing, realizing a closed loop of risk perception-feedback correction-dynamic thawing, thereby reducing the user experience of false positives while ensuring the effectiveness of risk prevention and control; finally, the final auditing decision is made based on the human-computer collaborative auditing module and the risk identification and penalty suggestion are fed back to realize sustainable optimization of intelligent risk identification and violation penalty, so that the system has efficient, accurate, scalable and self-optimizing anti-fraud auditing capabilities in the multi-live broadcast platform, massive data and complex behavior scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 FIG. 1 is a structural schematic diagram of an anti-fraud auditing system for live broadcast platform gift-giving behavior in an embodiment; Figure 2 FIG. 2 is a flowchart of an anti-fraud auditing method for live broadcast platform gift-giving behavior in an embodiment; Figure 3 FIG. 3 is an internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0018] In an embodiment, as shown in FIG. 1, the anti-fraud auditing system for live broadcast platform gift-giving behavior comprises an adaptive access module, a data acquisition and pre-control module, a risk identification and penalty suggestion module, and a human-computer collaborative auditing module. Figure 1As shown, a live streaming platform gift-giving fraud prevention auditing system is provided, including an adaptive access module, a data acquisition and pre-control module, a risk identification and penalty recommendation module, and a human-computer collaborative auditing module.

[0019] 1. The adaptive access module is used to provide a standard interface to map the standard auditing request of the system in real time to the API call recognizable by the target live streaming platform, realize the data interaction between the system and the target live streaming platform. Specifically, it includes a standard interface and an interface mapping unit.

[0020] The standard interface includes: (1) a data submission and delivery interface (Data Submission Protocol), which is used for the target live streaming platform to report gift-giving behavior and multi-modal risk control feature data in real time, and the standard definition is: define the standard interface such as SubmitGiftEvent (gift event), require the input parameters to include the donor ID, the recipient ID, the gift value, the timestamp, etc., and ensure the consistency of data format and content. (2) An information query and traceability interface (Data Retrieval Protocol) is used for the human-computer collaborative auditing module to call the deep information of the target live streaming platform on demand to support the final auditing result, and the standard definition is: define the interface such as GetUserInfo (get user information), QueryChatHistory (query chat record), FetchAVRecord (query audio and video record), and shield the differences of data storage and access of each platform. A risk disposal and execution interface (Penalty Execution Protocol) is used to send the final auditing result and the final penalty level to the target live streaming platform and execute the corresponding penalty behavior after the auditing is completed, and the standard definition is: define the atomized interface such as BanAccount (ban account), UnbanAccount (unban account), FreezeFunds (freeze funds), DeductFunds (deduct funds), and ensure that the disposal operation can be automatically executed. An event notification and reach interface (Message Notification Protocol) is used to send notification or alarm SMS to the target live streaming platform, and the standard definition is: define the interface such as SendSystemMessage (send system message), SendSms (send SMS), and ensure the unified standard of notification reach.

[0021] The interface mapping unit is used to realize the dynamic mapping of the standard interface to the target live broadcast platform API, so that new live broadcast platforms can be accessed through visual configuration without writing code. Specifically, it includes interface address mapping and request parameter mapping. Among them, the interface address mapping is used to point the standard interface to the actual API of the target live broadcast platform, for example, pointing the BanAccount interface to the actual ban API of live broadcast platform A. The request parameter mapping is used to map the standard interface parameter name to the actual parameter name of the target live broadcast platform, for example: standard interface parameters: {"userId": "123", "reason": "investment inducement"}, live broadcast platform A interface parameters: {"uid": "123", "block_reason_text": "investment inducement"}, authentication and header information: configure API Key, Token and other authentication information to realize safe calling.

[0022] It should be understood that the adaptive access module can ensure that the system can quickly access any live broadcast platform without modifying the core logic, and realize the risk monitoring and disposal ability of multi-platform generalization, strong expansibility and easy maintenance.

[0023] 2、Data acquisition and pre-control module, used to obtain the gift-giving behavior of the target live broadcast platform according to the adaptive access module, construct a gift-giving relationship pair formed by the combination of the IDs of the gift-giving parties, collect the multi-modal risk control feature data of the gift-giving relationship pair within a preset period (such as one day), and pre-freeze all transaction funds within the period.

[0024] Among them, the gift-giving relationship pair is formed by the combination of the unique gift-giving party ID and the unique receiving party ID as the basic unit of risk analysis. The system obtains the gift-giving behavior record through real-time stream processing or batch interface, analyzes and stores each record, generates the gift-giving relationship pair, and establishes an index in the database, which is convenient for subsequent efficient query and risk calculation. This method discards the traditional isolated transaction analysis mode, can focus on the fund flow and interactive behavior between specific gift-giving pairs, and realize accurate identification of potential fraud relationship chains.

[0025] The multi-modal risk control feature data of the gift-giving relationship pair within the preset period includes transaction data, user portrait data and text data of the gift-giving parties; among them, the transaction data includes the cumulative transaction amount and transaction frequency of the gift-giving relationship pair within the preset period; the user portrait data includes the historical violation records, account level and real name authentication information of the gift-giving parties; the text data includes the chat records of the gift-giving parties and the text information extracted from the audio-video calls of the gift-giving parties by using the automatic speech recognition (ASR) technology.

[0026] It should be understood that the pre-freezing of all transaction funds in the data acquisition and pre-control module can ensure that the funds cannot be transferred or withdrawn before the audit, and the freezing state is associated with the subsequent audit results. When the audit is completed, the final audit results can be used to determine automatic unfreezing or perform corresponding penalty operations. This mechanism changes the risk management from after-the-fact tracking to pre-control, effectively ensuring platform fund safety and risk controllability.

[0027] 3. Risk identification and penalty suggestion module, which has a fraud behavior knowledge base, an intelligent scoring engine and a penalty strategy matrix.

[0028] The fraud behavior knowledge base is maintained and expanded by the system administrator through the background management system. Each record in the knowledge base defines a specific fraud behavior pattern and its feature template, such as "impersonating an official", "investment inducement", "cashback fraud", "seduction behavior", etc. Each behavior pattern is bound to a feature template (such as gift amount threshold, interaction frequency, private message keywords, historical violation records, etc.), forming a "violation tag" that can be recognized by the algorithm, which is used for subsequent automatic scoring and screening.

[0029] The penalty strategy matrix is used to map the severity of the violation, the number of historical violations and related business indicators (such as gift amount, user level, etc.) to specific penalty measures. This matrix allows each violation tag to correspond to one or more penalty levels. For example: S+ level: permanently ban the account and freeze the violation earnings; A level: freeze today's earnings; B level: temporarily freeze earnings (24-72 hours). The system can dynamically adjust the penalty strategy matrix through the configuration interface to achieve flexible management and update of the penalty rules, while ensuring consistency between the algorithm processing and the manual audit results.

[0030] The running logic of the risk identification and penalty suggestion module includes: First, the transaction data, user portrait data and text data in the multi-modal risk control feature data are respectively encoded according to the intelligent scoring engine to obtain a multi-modal risk control feature vector of the gift relationship pair, including transaction features, portrait features and text features, for example, the natural language processing (NLP) technology is used to analyze the keywords (such as "investment", "add WeChat", "emergency" and the like), semantic relationship and context in the text data and perform feature encoding to generate text features; secondly, a cross-modal cross-attention mechanism is used, any one of the multi-modal risk control feature vectors is taken as a query vector Q, the other is taken as a key vector K and a value vector V, the cross-attention weight between any two multi-modal risk control feature vectors is calculated according to the query vector Q, the key vector K and the value vector V, and the multi-modal risk control feature vector after attention weighting is obtained according to the cross-attention weight, including transaction attention weighted portrait features, transaction attention weighted text features, portrait attention weighted text features, portrait attention weighted transaction features, text attention weighted transaction features and text attention weighted portrait features; the multi-modal risk control feature vector and the multi-modal risk control feature vector after attention weighting are spliced, dimensionally reduced and residual connected to obtain a fusion risk control feature of the gift relationship pair, the fusion risk control feature and the feature templates of various types of violation labels in the fraud behavior knowledge base are matched and calculated in parallel according to the relevance, the violation tendency scores of each violation label are obtained according to the matching degree, and the violation label with the highest score is selected as the main violation reason; wherein the violation label is composed of various types of fraud behaviors and corresponding feature templates. It can be understood that this step not only can determine whether the risk exists, but also can clearly point out the specific violation reason, providing a direct basis for artificial auditing and punishment decision.

[0031] Then, the confidence of the main violation reason is mapped to a reason confidence risk score, for example, when the system determines that the confidence of the "investment inducement" reason is 90%, the risk score is 90 points, and this method ensures that the punishment decision can directly reflect the reliability of automatic violation identification. The cumulative transaction amount and the historical cumulative violation times of the gift relationship pair within a preset period are respectively mapped to an amount risk score, a historical risk score of the gift giver and a historical risk score of the recipient according to a configurable, nonlinear grading mapping table, so as to reflect the nonlinear contribution of the amount and repeated violation behavior to the risk. For example, cumulative transaction amount 0-1000 yuan → amount risk score 10; cumulative transaction amount 1001-5000 yuan → amount risk score 30; cumulative transaction amount 5001-10000 yuan → amount risk score 85, the mapping table supports dynamic configuration, which can be adjusted according to the platform strategy and real-time risk control demand, ensuring that the risk weight of different amount intervals is reasonable. For example, the more the historical cumulative violation times, the higher the score, so as to enhance the attention to repeated violation behaviors, and this quantitative method can be directly combined with other factors to provide a reliable basis for comprehensive punishment decision.

[0032] After standardizing the scoring of all key factors, the risk scores for reason confidence, amount, donor's historical risk score, and recipient's historical risk score are weighted and summed to obtain the comprehensive risk index of the gift-giving relationship, expressed as: ; in, , , and The weighting coefficients for each risk score can be configured by the system administrator in the backend, and their sum is 100%. This design gives the system great flexibility, allowing operators to dynamically adjust the importance of each factor according to the regulatory focus at different times (e.g., whether to focus on cracking down on large-scale fraud or on cleaning up the streamer's past behavior), thereby achieving refined risk management.

[0033] Comprehensive risk index The penalty level is compared with a pre-defined penalty level threshold table in the penalty strategy matrix, and the corresponding suggested penalty level is matched based on the comparison result. For example, CRI>90 → S+ level penalty, 80≤CRI≤89 → S level penalty, 65≤CRI≤79 → A+ level penalty, and other ranges → mapped to levels A, B1, B2, etc., according to the strategy. This penalty level threshold table can be dynamically configured to adapt to different risk control strategies or business changes. Through this quantitative and configurable mapping method, the system can achieve standardization and interpretability of penalty level decisions while ensuring automation.

[0034] Furthermore, based on the comprehensive risk index Within a given timeframe, the system automatically adjusts the pre-freezing ratio and duration of all transaction funds related to gift-giving relationships within a preset period, generating differentiated pre-freezing and unfreezing strategies. For example, To minimize risk, most of the pre-frozen funds will be quickly unfrozen, with only a small amount remaining pre-frozen as a risk control observation period. For medium-risk cases, the pre-frozen funds will be partially unfrozen, and the frozen amount will decrease proportionally to the freezing period. For example, only 50% will be unfrozen, and the remaining funds will have an extended observation period. This is considered high-risk; therefore, all data will remain frozen until further manual review or subsequent actions confirm that no abnormalities have been detected.

[0035] Finally, by combining gift-giving relationships, main reasons for violations, comprehensive risk index, and suggested penalty level, a task awaiting review is generated.

[0036] It should be understood that the cross-modal cross-attention fusion of the multi-modal risk control feature data based on the risk identification and punishment recommendation module is beneficial to learn the complex nonlinear relationship between the features of different modalities under different fraud behaviors, improve the comprehensiveness and explainability of the violation identification, accurately point out the main violation reason, further calculate the comprehensive risk index through multi-factor quantization and match the recommended punishment level, and convert the generation of the punishment recommendation from relying on experience and artificial judgment to data-driven, clear logic, and traceable automatic decision, which can provide reliable audit decision basis for artificial audit, reduce the dependence on artificial experience, avoid inconsistent or lagging punishment standards, not only improve the anti-fraud audit efficiency, but also enhance the fairness and explainability of the punishment recommendation. And the generation of the differentiated pre-freezing and thawing strategy according to the comprehensive risk index avoids the "one-size-fits-all" long-term freezing in the prior art, but realizes the closed loop of risk perception-feedback correction-dynamic thawing, so as to reduce the user experience of false positives while ensuring the effectiveness of risk prevention and control.

[0037] 4. A man-machine collaborative audit module for receiving a to-be-audited task for man-machine collaborative audit, and feeding back the final audit result and the final punishment level of the gifting behavior in a preset period to the target live broadcast platform through the adaptive access module, and feeding back the final audit result and the final punishment level to the risk identification and punishment recommendation module, to realize adaptive optimization of risk identification and punishment recommendation. Specifically, it includes: A task management module for storing the to-be-audited task output by the risk identification and punishment recommendation module, and performing task priority sorting and distribution according to the size of the comprehensive risk index or the severity of the main violation reason, so that high-risk tasks are placed at the top to ensure that the most urgent events are processed first, and realize intelligent scheduling of audit resources. Further, the audit personnel can receive the to-be-audited task according to the permission and expertise, and the system can automatically distribute the task based on the preset rules (such as automatically distributing the financial fraud task to the financial audit group), to improve the distribution efficiency.

[0038] The man-machine review decision interface includes: a preprocessing conclusion abstract area for displaying main violation reasons and recommended penalty levels of the to-be-reviewed task, providing a reference for manual review; a related evidence one-key tracing area for accessing multi-modal risk control feature data of the to-be-reviewed task corresponding to the gift-giving relationship pair, including transaction data, user portrait data, and text data of the gift-giving parties, and realizing efficient evidence review; and a review decision operation area for manually reviewing and outputting final review results and final penalty levels of the gift-giving relationship pair in the preset period, in combination with information provided by the preprocessing conclusion abstract area and the related evidence one-key tracing area. Specifically, if it is confirmed that there is no risk, the feedback module feeds back the no-risk result to the data acquisition and pre-control module and the risk identification and penalty suggestion module, and unfreezes the current pre-frozen transaction funds. If it is confirmed that there is a violation, the system suggestion can be adopted or the penalty level can be manually adjusted, and a note can be added. If it is determined that there is a false positive, the task is closed, and no operation is performed.

[0039] The feedback module is configured to feed back the final review results and the final penalty levels to the target live streaming platform through the adaptive access module, and feed back the final review results and the final penalty levels to the risk identification and penalty suggestion module, realize closed-loop feedback, and realize adaptive optimization of risk identification and penalty suggestion by comparing and analyzing the deviation of risk identification and penalty level.

[0040] It should be understood that the final review decision and feedback of optimized risk identification and penalty suggestion based on the man-machine collaborative review module can realize sustainable optimization of intelligent risk identification and violation penalty, The above-mentioned anti-fraud review system for gift-giving behavior of a live streaming platform is based on the collaborative operation of the adaptive access module, the data acquisition and pre-control module, the risk identification and penalty suggestion module, and the man-machine collaborative review module, and can have efficient, accurate, scalable, and self-optimizing anti-fraud review capabilities in multiple live streaming platforms, massive data, and complex behavior scenarios.

[0041] In one embodiment, as shown in Figure 2 An anti-fraud review method for gift-giving behavior of a live streaming platform is provided, which is realized based on the above-mentioned anti-fraud review system for gift-giving behavior of a live streaming platform, and includes the following steps: Step 1: Through adaptive access between the system and the target live streaming platform, the gift-giving behavior of the target live streaming platform is acquired, a gift-giving relationship pair formed by the ID combination of the gift-giving parties is constructed, multi-modal risk control feature data of the gift-giving relationship pair in a preset period is collected, and all transaction funds in the period are pre-frozen.

[0042] Step 2, according to the intelligent scoring engine, the multi-modal risk control feature data is cross-modal cross-attention fusion, the fused risk control features are matched with various types of violation labels in the fraud behavior knowledge base in parallel to calculate the violation tendency score of each violation label, and the violation label with the highest score is selected as the main violation reason. According to the comprehensive risk index calculated according to the main violation reason, the cumulative transaction amount and the historical cumulative violation times of the gift-giving relationship pair, the suggested punishment level is matched in the punishment strategy matrix and the differentiated pre-freezing unfreezing strategy is automatically generated. Finally, the to-be-audited task is generated by combining the gift-giving relationship pair, the main violation reason, the comprehensive risk index and the suggested punishment level.

[0043] Step 3, receiving the to-be-audited task for human-computer collaborative auditing, feeding back the final audit result and the final punishment level of the gift-giving relationship pair in the preset period to the target live broadcast platform, and feeding back the final audit result and the final punishment level to the risk identification and punishment suggestion module, realizing adaptive optimization of risk identification and punishment suggestion.

[0044] In one embodiment, a computer device can be provided, which can be a terminal, and an internal structure diagram thereof can be as shown in Figure 3 The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement an anti-fraud auditing method for gift-giving behavior of a live broadcast platform. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0045] Those skilled in the art can understand that Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0046] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the following steps: By adaptive access between the system and the target live broadcast platform, the gift-giving behavior of the target live broadcast platform is obtained, the gift-giving relationship pair formed by the ID combination of the gift-giving parties is constructed, the multi-modal risk control feature data of the gift-giving relationship pair in the preset period is collected, and all transaction funds in the period are pre-frozen; According to the cross-modal cross-attention fusion of the multi-modal risk control feature data by the intelligent scoring engine, the fused risk control features and various types of violation labels in the fraud behavior knowledge base are matched and calculated in parallel to obtain the violation tendency scores of each violation label, the violation label with the highest score is selected as the main violation reason, and the comprehensive risk index calculated according to the main violation reason of the gift-giving relationship pair, the cumulative transaction amount and the historical cumulative violation times is matched with the recommended punishment level in the punishment strategy matrix and the differentiated pre-freezing unfreezing strategy is automatically generated, and finally the to-be-reviewed task is generated by combining the gift-giving relationship pair, the main violation reason, the comprehensive risk index and the recommended punishment level. The to-be-reviewed task is received for human-machine collaborative review, and the final review result and the final punishment level of the gift-giving behavior of the gift-giving relationship pair in the preset period are fed back to the target live broadcast platform, and the final review result and the final punishment level are fed back to the risk identification and punishment suggestion module, realizing adaptive optimization of risk identification and punishment suggestion.

[0047] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0048] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are within the scope of the present application.

Claims

1. A live broadcast platform gift-giving behavior anti-fraud auditing system, characterized in that, The system comprises: An adaptive access module for providing a standard interface to map standard audit requests of the system in real time into API calls identifiable by a target live broadcast platform, so as to realize data interaction between the system and the target live broadcast platform; A data acquisition and pre-pipeline control module for acquiring gift-giving behaviors of the target live broadcast platform according to the adaptive access module, constructing a gift-giving relationship pair formed by an ID combination of both parties of gift-giving, collecting multi-modal risk control feature data of the gift-giving relationship pair within a preset period, and pre-freezing all transaction funds within the period; A risk identification and penalty suggestion module internally provided with a fraud behavior knowledge base, an intelligent scoring engine and a penalty strategy matrix, for performing cross-modal cross-attention fusion on the multi-modal risk control feature data according to the intelligent scoring engine, performing parallel matching calculation on the fused risk control features and various types of violation labels in the fraud behavior knowledge base to obtain violation tendency scores of each violation label, selecting the violation label with the highest score as the main violation reason, and calculating a comprehensive risk index according to the main violation reason of the gift-giving relationship pair, the cumulative transaction amount and the historical cumulative violation times, matching a recommended penalty level in the penalty strategy matrix and automatically generating a differentiated pre-freezing unfreezing strategy, and finally generating a to-be-audited task by combining the gift-giving relationship pair, the main violation reason, the comprehensive risk index and the recommended penalty level; A man-machine collaborative audit module for receiving the to-be-audited task for man-machine collaborative audit, and feeding back the final audit result and the final penalty level of the gift-giving behavior of the gift-giving relationship pair within the preset period to the target live broadcast platform through the adaptive access module, and feeding back the final audit result and the final penalty level to the risk identification and penalty suggestion module, so as to realize adaptive optimization of risk identification and penalty suggestion.

2. The anti-fraud review system for live broadcast platform gift-giving behavior according to claim 1, wherein, The adaptive access module comprises a standard interface and an interface mapping unit; The standard interface comprises a data reporting and delivery interface for the target live broadcast platform to report gift-giving behaviors and multi-modal risk control feature data in real time, an information query and traceability interface for the man-machine collaborative audit module to call deep information of the target live broadcast platform on demand to support the final audit result, a risk disposal and execution interface for sending the final audit result and the final penalty level to the target live broadcast platform and executing corresponding penalty behaviors after the audit is completed, and an event notification and reach interface for sending notification or alarm SMS to the target live broadcast platform; The interface mapping unit is used to realize dynamic mapping of the standard interface to the API of the target live broadcast platform, including interface address mapping and request parameter mapping; wherein the interface address mapping is used to point the standard interface to the actual API of the target live broadcast platform, and the request parameter mapping is used to map the standard interface parameter name to the actual parameter name of the target live broadcast platform.

3. The anti-fraud review system for live broadcast platform gift-giving behavior according to claim 1, characterized in that, The multi-modal risk control feature data includes transaction data, user portrait data, and text data of both gift-giving parties; the transaction data includes cumulative transaction amount and transaction frequency of the gift-giving relationship pair in a preset period; the user portrait data includes historical violation records, account level, and real-name authentication information of both gift-giving parties; and the text data includes chat records of both gift-giving parties and text information extracted from audio and video calls of both gift-giving parties.

4. The anti-fraud auditing system for live broadcast platform gift-giving behavior according to claim 3, characterized in that, The intelligent scoring engine performs cross-modal cross-attention fusion on the multi-modal risk control feature data, performs parallel matching calculation on the fused risk control features and various types of violation labels in the fraud behavior knowledge base, and obtains violation tendency scores of each violation label, including: The intelligent scoring engine first performs feature encoding on the transaction data, user portrait data, and text data in the multi-modal risk control feature data, respectively, to obtain multi-modal risk control feature vectors of the gift-giving relationship pair, including transaction features, portrait features, and text features; Secondly, a cross-modal cross-attention mechanism is used, any one of the multi-modal risk control feature vectors is taken as a query vector Q, and the other is taken as a key vector K and a value vector V, cross-attention weights between any two multi-modal risk control feature vectors are calculated according to the query vector Q, the key vector K, and the value vector V, and the multi-modal risk control feature vectors after attention weighting are obtained according to the cross-attention weights, including transaction attention weighted portrait features, transaction attention weighted text features, portrait attention weighted text features, portrait attention weighted transaction features, text attention weighted transaction features, and text attention weighted portrait features; The multi-modal risk control feature vectors and the multi-modal risk control feature vectors after attention weighting are spliced, dimensionally reduced, and residual connected to obtain fused risk control features of the gift-giving relationship pair, parallel correlation matching calculation is performed on the fused risk control features and feature templates of various types of violation labels in the fraud behavior knowledge base, and violation tendency scores of each violation label are obtained according to the matching degree; wherein the violation labels are composed of various types of fraud behaviors and corresponding feature templates.

5. The anti-fraud review system for live broadcast platform gift-giving behavior according to claim 1, wherein, According to the comprehensive risk index calculated based on the main violation reason of the gift-giving relationship pair, the cumulative transaction amount, and the historical cumulative violation times, the recommended penalty level is matched in the penalty strategy matrix, and a differentiated pre-freezing unfreezing strategy is automatically generated, including: The confidence of the main violation reason is mapped to a reason confidence risk score, and the cumulative transaction amount and the historical cumulative violation times of the gift-giving relationship pair in a preset period are respectively mapped to an amount risk score, a gift-giving party historical risk score, and a receiving party historical risk score according to a configurable, non-linear grading mapping table; The reason confidence risk score, the amount risk score, the gift-giving party historical risk score, and the receiving party historical risk score are weighted and summed to obtain the comprehensive risk index of the gift-giving relationship pair, represented as: ； wherein, , , and represent the weight coefficients of each risk score; By comparing the comprehensive risk index with the preset penalty level threshold table of the penalty strategy matrix, a recommended penalty level of the gift relationship pair in a preset period is obtained, and according to the interval of the comprehensive risk index , the pre-freezing proportion and the time length of all transaction funds of the gift relationship pair in the preset period are automatically adjusted to generate a differentiated pre-freezing and thawing strategy.

6. The anti-fraud review system for live broadcast platform gift-giving behavior according to claim 1, wherein, The human-computer collaborative review module includes: The task management module is used to store the tasks to be reviewed output by the risk identification and penalty suggestion module, and to prioritize and allocate tasks according to the magnitude of the comprehensive risk index or the severity of the main reasons for violation. The human-machine review decision interface includes: a preprocessing conclusion summary area, used to display the main reasons for violations and the suggested penalty level for the task under review; a one-click tracing area for related evidence, used to access the multimodal risk control feature data of the gift-giving relationship pair corresponding to the task under review, including transaction data, user profile data, and text data of both parties; and a review decision operation area, used to combine the information provided by the preprocessing conclusion summary area and the one-click tracing area for related evidence, and manually review and output the final review result and final penalty level of the gift-giving behavior of the gift-giving relationship pair corresponding to each task under review within a preset period. The feedback module is used to feed back the final review results and final penalty level to the target live streaming platform through the adaptive access module, and at the same time feed back the final review results and final penalty level to the risk identification and penalty suggestion module. By comparing and analyzing the deviation between risk identification and penalty level, the risk identification and penalty suggestion are adaptively optimized.

7. The anti-fraud review system for live broadcast platform gift-giving behavior according to claim 6, characterized in that, The feedback module is also used to provide a risk-free result to the data acquisition and pre-control module and the risk identification and penalty suggestion module when the final review result is risk-free, so as to unfreeze the currently pre-frozen transaction funds.

8. A live broadcast platform gift-giving behavior anti-fraud auditing method, characterized in that, The method is implemented based on the anti-fraud verification system for gift-giving behavior on live streaming platforms as described in any one of claims 1-7, and includes the following steps: By adaptively accessing the system and the target live streaming platform, the gift-giving behavior of the target live streaming platform is obtained, a gift-giving relationship pair formed by the combination of the IDs of the gift-givers is constructed, the multimodal risk control feature data of the gift-giving relationship pair is collected within a preset period, and all transaction funds within the period are pre-frozen. The intelligent scoring engine performs cross-modal cross-attention fusion on the multimodal risk control feature data, and performs parallel matching calculations on the fused risk control features and various violation tags in the fraud behavior knowledge base to obtain the violation tendency score of each violation tag. The violation tag with the highest score is selected as the main violation reason. Based on the main violation reason, cumulative transaction amount and historical cumulative number of violations of the gift-giving relationship pair, a comprehensive risk index is calculated. The suggested penalty level is matched in the penalty strategy matrix and a differentiated pre-freeze and unfreeze strategy is automatically generated. Finally, by combining the gift-giving relationship pair, the main violation reason, the comprehensive risk index and the suggested penalty level, a task to be reviewed is generated. The system receives the pending task and conducts a human-machine collaborative review. It then feeds back the final review result and final penalty level of the gift-giving relationship for the gift-giving behavior within a preset period to the target live streaming platform. At the same time, it feeds back the final review result and final penalty level to the risk identification and penalty suggestion module, thereby achieving adaptive optimization of risk identification and penalty suggestion. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the method of claim 8.

Images