A financial live broadcast compliance risk control method and system based on differential game

Abstract

The application relates to a financial live broadcast compliance risk control method and system based on differential game, and belongs to the intelligent director field. The method comprises the following steps: inputting a financial live broadcast stream original signal and performing compressed sensing state mapping to obtain a state vector and a noise covariance matrix; performing implicit intention inference based on the state vector and the noise covariance matrix to obtain estimated intention parameters and uncertainty at the current moment; performing zero-sum differential game based on the state vector and the intention parameters to obtain an optimal intervention signal; injecting the optimal intervention signal into the live broadcast stream in real time to obtain a new state vector, constructing a meta-learning update model, and obtaining an updated value function and a meta-learning rate; and performing safety supervision and intervention rollback. By introducing the implicit intention inference, the differential game and the meta-learning mechanism, active identification and low-perception confrontation of induced behaviors are realized, and the foresight, adaptability and safety of the financial live broadcast compliance risk control are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent broadcasting technology, specifically relating to a financial live streaming compliance risk control method and system based on differential game theory. Background Technology

[0002] Current compliance and risk control in financial live streaming primarily relies on static, post-event or rule-driven methods such as keyword filtering, manual review, and image recognition of inappropriate content. For example, this involves matching bullet comments and voice messages using a pre-set sensitive word library, or identifying enticing elements like QR codes and red envelopes in the video using object detection models, supplemented by manual spot checks and account bans. While these technologies can intercept obvious violations to some extent, they are inherently reactive and struggle to address the constantly evolving enticing strategies employed by streamers, especially those using subtle language, emotional manipulation, and visual metaphors to circumvent rules in a gray area. Furthermore, existing methods lack the ability to model and dynamically engage with the streamer's intentions, failing to provide precise and subtle proactive intervention before or in the early stages of enticing behavior. This results in delayed risk control, high false positive rates, and a compromised user experience. As financial live streaming becomes increasingly large-scale and complex, static rule systems are no longer sufficient to meet the combined requirements of real-time performance, accuracy, and concealment. There is an urgent need for an intelligent risk control method capable of inferring the streamer's intentions online, dynamically engaging with their behavior, and adaptively optimizing intervention strategies. Summary of the Invention

[0003] To address the aforementioned problems in the existing technology, this invention provides a financial live streaming compliance risk control method and system based on differential game theory.

[0004] The objective of this invention can be achieved through the following technical solutions: A financial live streaming compliance risk control method based on differential game theory, the implementation of which includes the following steps: Step S1: Input the raw signal of the financial live stream and perform compressed sensing state mapping to obtain the state vector. and noise covariance matrix ; Step S2: Based on the state vector and the noise covariance matrix Perform implicit intent inference to obtain the estimated intent parameters at the current time. and its uncertainties; Step S3: Based on the state vector and the intent parameters To obtain the optimal intervention signal, a zero-sum differential game is performed. ; Step S4: Transfer the optimal intervention signal Injected into the live stream in real time, a new state vector is obtained. And construct a meta-learning update model to obtain the updated value function. Learning rate ; Step S5: Based on the new state vector Conduct safety monitoring and intervention rollback.

[0005] Preferably, the compressed sensing state mapping in step S1 specifically includes: The raw signals of the financial live stream are collected, including video frame sequences, audio waveforms, text streams, and audience behavior data; Align the original signals of the financial live stream to the same time axis, and construct an original feature vector for each time point. ; The original feature vector The state vector is obtained by compressing it into a low-dimensional state space. Mathematically described ,in, A fixed sparse random projection matrix with dimensions d×N, where d is the target state dimension and N is the original feature dimension, and the elements in the matrix are selected with probability p. The rest are 0, and p is the online estimate. sparsity, For measuring noise; Based on measurement noise Real-time estimation of the noise covariance matrix .

[0006] Preferably, the implicit intent inference in step S2 specifically includes: The intent parameters are updated in real time using a variational Bayesian filter. The probability estimate, mathematically described as ,in, Let be the probability density function. This represents the intention parameters after observing the states from time 0 to k. The probability distribution, This represents the posterior distribution at the previous time step. For temperature parameters, Indicates the current state Assuming the streamer uses When an intention is made, action 'a' is selected according to a certain probability distribution, and these actions are used to predict the next state. And calculate the log-expected value of the predicted probability; simultaneously through the intent parameters The uncertainty is obtained from the covariance matrix.

[0007] Preferably, the zero-sum differential game in step S3 specifically refers to: Treating the interaction between the risk control process and the livestreamer as a zero-sum differential game, the optimal intervention signal is obtained by approximately solving the HJB equation of the zero-sum differential game. Mathematically described ,in, The intervention energy cost matrix is ​​diagonally positive definite, where each diagonal element corresponds to the cost of different components of the intervention signal. Let be a value function, representing the cumulative risk value starting from state s, assuming both parties adopt optimal strategies. This is the intervention effect matrix, which represents how the intervention changes the state at the next moment.

[0008] Preferably, the construction of the meta-learning update model in step S4 specifically involves: The optimal intervention signal Injected into the live stream in real time, the system continues to collect live data to obtain a new state vector after intervention. and actual risk loss ; The updated value function is obtained based on the meta-learning update rate. and the updated meta-learning rate Mathematically described ,in, This represents the update amount of the value function. The learning rate is the value function. For the risk loss predicted before intervention, for Regarding value functions gradient, The meta-learning rate, For the expectation of the meta-learning task, For the loss of yuan, This indicates an assignment operation; The updated value function and the meta-learning rate Used for the next round of intervention.

[0009] Preferably, the safety supervision and intervention rollback in step S5 specifically refers to: Set intervention safety indicators Mathematically described ,in, To intervene in the proportion that is noticed by the audience, The upper limit of the allowed detection rate, The square of the weighted energy of the intervention signal is summed according to the perceived weight. To sense the upper limit of energy; when the intervention safety index When the value falls below a preset threshold, an intervention rollback is performed, reducing the intervention intensity to 0 and restoring the value function and meta-learning rate of the previous version, and then re-intervening.

[0010] A financial live streaming compliance and risk control system based on differential game theory is used to execute the aforementioned financial live streaming compliance and risk control method based on differential game theory, including a perception mapping module, an implicit intent inference module, a zero-sum differential game module, a meta-learning update module, and a security supervision module. The perception mapping module is used to input the raw signal of the financial live broadcast stream and perform compressed perception state mapping to obtain a state vector. and noise covariance matrix ; The implicit intent inference module is used to infer based on the state vector. and the noise covariance matrix Perform implicit intent inference to obtain the estimated intent parameters at the current time. and its uncertainties; The zero-sum differential game module is used to base the state vector and the intent parameters To obtain the optimal intervention signal, a zero-sum differential game is performed. ; The meta-learning update module is used to update the optimal intervention signal. Injected into the live stream in real time, a new state vector is obtained. And construct a meta-learning update model to obtain the updated value function. Learning rate ; The safety monitoring module is used to base on the new state vector. Conduct safety monitoring and intervention rollback.

[0011] The beneficial effects of this invention are as follows: (1) By mapping multimodal live signals into low-dimensional state vectors through compressed sensing, the computational complexity of real-time processing is significantly reduced, laying the foundation for intervention decision-making.

[0012] (2) Based on the variational Bayesian filter, the implicit intention parameters of the anchor and their uncertainty are inferred online, which enables the system to capture abnormal strategy tendencies before the inducement behavior becomes explicit, thus greatly improving the early warning lead time.

[0013] (3) By solving the HJB equation of the zero-sum differential game, the optimal intervention signal can be obtained, which can effectively reduce the risk exposure of the audience while maintaining the concealment of the intervention and avoid interfering with the normal live broadcast experience.

[0014] (4) Introducing a meta-learning mechanism to update the value function and learning rate online, enabling the system to continuously adapt to new induction methods and form a continuously evolving defense capability. At the same time, the robustness and security of the system operation are ensured through security supervision and intervention rollback mechanisms. Attached Figure Description

[0015] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart illustrating the steps of a financial live-streaming compliance and risk control method based on differential game theory, as proposed in this invention. Detailed Implementation

[0017] To better understand the invention, various aspects of the invention will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are merely illustrative of exemplary embodiments of the invention and are not intended to limit the scope of the invention in any way. Throughout the specification, the expression "and / or" includes any and all combinations of one or more of the associated listed items. As used herein, the terms "approximately," "about," and similar terms are used as expressions of approximation, not as expressions of degree, and are intended to describe inherent deviations in measured or calculated values ​​that will be recognized by those skilled in the art. Furthermore, the order in which the steps are described in this invention does not necessarily indicate the order in which these steps occur in actual operation, unless otherwise expressly defined or deduced from the context.

[0018] It should also be understood that expressions such as "comprising," "including," "having," "containing," and / or "comprising" are open-ended rather than closed-ended expressions in this specification, indicating the presence of the stated features, elements, and / or components, but not excluding the presence of one or more other features, elements, components, and / or combinations thereof. Furthermore, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features, not just individual elements in the list. Additionally, when describing embodiments of the invention, the word "may" is used to mean "one or more embodiments of the invention." And the term "exemplary" is intended to refer to examples or illustrations.

[0019] Unless otherwise specified, all terms used herein (including engineering and technical terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that, unless expressly stated herein, terms defined in common dictionaries shall be interpreted as having the meaning consistent with their meaning in the context of the relevant art, and not in an idealized or overly formalized sense.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] Example 1: Please see Figure 1 A compliance and risk control method for financial live streaming based on differential game theory includes: Step S1: Input the raw signal of the financial live stream and perform compressed sensing state mapping to obtain the state vector. and noise covariance matrix ; Step S2: Assume there is a hidden intent parameter vector behind the anchor's guiding behavior, which determines how the anchor will choose actions (such as what to say, what images to show, etc.) in different states. This vector cannot be directly observed, but can be inferred by observing how the state evolves over time. Therefore, based on the state vector... and the noise covariance matrix Perform implicit intent inference to obtain the estimated intent parameters at the current time. and its uncertainties; Step S3: Based on the state vector and the intent parameters To obtain the optimal intervention signal, a zero-sum differential game is performed. ; Step S4: Transfer the optimal intervention signal Real-time injection into the live stream and construction of a meta-learning update model to obtain the updated value function. Learning rate ; Step S5: Based on the new state vector Conduct safety monitoring and intervention rollback.

[0022] In this embodiment, the compressed sensing state mapping is specifically as follows: S101: Collect the original signal of the financial live stream, including video frame sequence, audio waveform, text stream (the audio is transcribed into text in real time through speech recognition, and the bullet screen text is extracted at the same time), and audience behavior data (reward records, number of likes, changes in the number of online users, etc.). S102: Align the original signals of the financial live stream to the same time axis, and construct an original feature vector for each time point. For video frame sequences, key information is extracted from the images. For example, a pre-trained simple object detection model is used to identify whether specific objects such as "K-line charts", "QR codes", and "red envelope animations" appear in the images, and their positions, sizes, and durations of appearance are recorded. High-dimensional pixels are converted into low-dimensional semantic features (e.g., "K-line charts appear" is 1, otherwise it is 0). For audio waveforms, acoustic features such as short-time energy, zero-crossing rate, and fundamental frequency (pitch) are calculated. At the same time, the text obtained through speech recognition is merged with the bullet screen text to form text features (e.g., keyword hit count, sentence length, etc.). For audience behavior data, the reward amount is normalized to the [0,1] interval, and the bullet screen sentiment score is (-1,1), etc. S103: Due to the original feature vector The space is high-dimensional and contains a lot of redundant information, so it needs to be compressed into a low-dimensional state space (before compression, the state space needs to be compressed). (Normalization) yields the state vector. Mathematically described ,in, A fixed sparse random projection matrix with dimensions d×N, where d is the target state dimension and N is the original feature dimension, and the elements in the matrix are selected with probability p. The rest are 0, and p is the online estimate. sparsity, For measuring noise; S104: Based on measurement noise The noise covariance matrix is ​​estimated in real time using an online Kalman filter method. .

[0023] In this embodiment, the implicit intent inference specifically refers to: The intent parameters are updated in real time using a variational Bayesian filter. The probability estimate, mathematically described as ,in, Let be the probability density function. This represents the intention parameters after observing the states from time 0 to k. The probability distribution, This represents the posterior distribution at the previous time step. Temperature is a parameter that controls the degree of confidence in the transition probability. Indicates the current state Assuming the streamer uses When an intention is made, action 'a' is selected according to a certain probability distribution, and these actions are used to predict the next state. It also calculates the log-expected value of the predicted probability; the larger this value, the stronger the current intent parameter. The better it can explain the observed state transitions; and simultaneously through the intended parameters The uncertainty is obtained from the covariance matrix. Example: Observing that after the streamer performs the same action at three consecutive time points, the audience donations increase significantly, and the state transition pattern is highly consistent with the historical pattern of "high induction intensity", then... The mean of the "induction intensity" dimension will rise from the initial 0.5 to 0.85, and the confidence interval will narrow. At the same time, the "concealment" dimension may also be high because the streamer did not use obvious prohibited words.

[0024] In this embodiment, the zero-sum differential game is specifically: The interaction between the risk control process and the livestreamer is viewed as a zero-sum differential game, and the risk control process selects intervention signals. (For example, adding noise of a specific frequency to the audio) to try to minimize the viewer's risk exposure, while the broadcaster tries to maximize the risk exposure. Both parties make decisions simultaneously within a continuous time frame, and the intervention signal is energy-constrained. Therefore, the optimal intervention signal is obtained by approximately solving the Hamilton-Jacobi-Bellman (HJB) equation of the zero-sum differential game. Mathematically described ,in, The intervention energy cost matrix is ​​diagonally positive definite. Each diagonal element corresponds to the cost of a different component of the intervention signal, i.e., the intrusiveness of that component to the viewer's perception. For example, audio noise with a frequency between 2kHz and 4kHz (where the human ear is most sensitive) has a high cost, while noise at 19kHz (almost inaudible) has a low cost. These costs are pre-calibrated through psychophysical experiments and fine-tuned online based on viewer feedback: if viewers comment with keywords such as "the screen is too flickering," it indicates that the intervention has been detected, and the cost of the corresponding dimension automatically increases, forcing the intervention to be more covert. Let be a value function, representing the cumulative risk value (the lower the better) starting from state s, assuming both parties adopt optimal strategies. The intervention impact matrix represents how the intervention changes the state at the next moment. It is obtained by linearizing the audience response model R(s,u) around u=0. R(s,u) needs to be pre-learned. Example: For a broadcaster with a high inductive intent parameter, the calculated... It is possible to inject narrowband noise into the audio that is out of phase with the broadcaster's voice fundamental frequency to counteract its wake-up effect; for broadcasters with high concealment, the system may rely more on visual perturbations, because these broadcasters rely more on visual inducement.

[0025] In this embodiment, the construction of the meta-learning update model is specifically as follows: S401: The optimal intervention signal Injected into the live stream in real time, the system continues to collect live data to obtain a new state vector after intervention. and actual risk loss (For example, the reduction in negative comments within a certain period after the intervention). S402: Obtain the updated value function based on the meta-learning update rate. and the updated meta-learning rate Mathematically described ,in, This represents the update amount of the value function. The learning rate is the value function. For the risk loss predicted before intervention, for Regarding value functions The gradient (obtained through policy gradient or adjoint method). The meta-learning rate, For the expectation of the meta-learning task, The meta-loss measures the prediction error in the subsequent short term using the updated value function. This indicates an assignment operation; S403: The updated value function... and the meta-learning rate Used for the next round of intervention.

[0026] In this embodiment, the security monitoring and intervention rollback specifically refers to: To ensure that the intervention does not disrupt the normal live broadcast or cause viewer backlash, security monitoring is necessary, and regular assessments should be conducted to determine if the intervention has been detected; intervention security indicators should be established. Mathematically described ,in, The proportion of interventions that are noticed by the audience is estimated through bullet screen keywords, manual spot checks, etc. The upper limit of the allowed detection rate, The square of the weighted energy of the intervention signal is summed according to the perceptual weight (e.g., audio energy multiplied by human ear sensitivity weight, video energy multiplied by human eye sensitivity weight, etc.). To sense the upper limit of energy; when the intervention safety index If the value is below the preset threshold, it indicates that the intervention is too obvious or the energy is too high. At this time, the intervention is rolled back, the intervention intensity is reduced to 0, and the value function and meta-learning rate of the previous version are restored, and the intervention is carried out again.

[0027] Example 2: A financial live streaming compliance and risk control system based on differential game theory includes a perception mapping module, an implicit intent inference module, a zero-sum differential game module, a meta-learning update module, and a security supervision module. The perception mapping module is used to input the raw signal of the financial live broadcast stream and perform compressed perception state mapping to obtain a state vector. and noise covariance matrix ; The implicit intent inference module is used to infer based on the state vector. and the noise covariance matrix Perform implicit intent inference to obtain the estimated intent parameters at the current time. and its uncertainties; The zero-sum differential game module is used to base the state vector and the intent parameters To obtain the optimal intervention signal, a zero-sum differential game is performed. ; The meta-learning update module is used to update the optimal intervention signal. Injected into the live stream in real time, a new state vector is obtained. And construct a meta-learning update model to obtain the updated value function. Learning rate ; The safety monitoring module is used to base on the new state vector. Conduct safety monitoring and intervention rollback.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A compliance risk control method for financial live streaming based on differential game theory, characterized in that, Includes the following steps: Step S1: Input the raw signal of the financial live stream and perform compressed sensing state mapping to obtain the state vector. and noise covariance matrix ; Step S2: Based on the state vector and the noise covariance matrix Perform implicit intent inference to obtain the estimated intent parameters at the current time. and its uncertainties; Step S3: Based on the state vector and the intent parameters To obtain the optimal intervention signal, a zero-sum differential game is performed. ; Step S4: Transfer the optimal intervention signal Injected into the live stream in real time, a new state vector is obtained. And construct a meta-learning update model to obtain the updated value function. Learning rate ; Step S5: Based on the new state vector Conduct safety monitoring and intervention rollback.

2. The financial live streaming compliance risk control method based on differential game theory according to claim 1, characterized in that, The compressed sensing state mapping in step S1 specifically refers to: The raw signals of the financial live stream are collected, including video frame sequences, audio waveforms, text streams, and audience behavior data; Align the original signals of the financial live stream to the same time axis, and construct an original feature vector for each time point. ; The original feature vector The state vector is obtained by compressing it into a low-dimensional state space. Mathematically described ,in, A fixed sparse random projection matrix with dimensions d×N, where d is the target state dimension and N is the original feature dimension, and the elements in the matrix are selected with probability p. The rest are 0, and p is the online estimate. sparsity, For measuring noise; Based on measurement noise Real-time estimation of the noise covariance matrix .

3. The financial live streaming compliance risk control method based on differential game theory according to claim 1, characterized in that, The implicit intent inference in step S2 specifically refers to: The intent parameters are updated in real time using a variational Bayesian filter. The probability estimate, mathematically described as ,in, Let be the probability density function. This represents the intention parameters after observing the states from time 0 to k. The probability distribution, This represents the posterior distribution at the previous time step. For temperature parameters, Indicates the current state Assuming the streamer uses When an intention is made, action 'a' is selected according to a certain probability distribution, and these actions are used to predict the next state. And calculate the log-expected value of the predicted probability; simultaneously through the intent parameters The uncertainty is obtained from the covariance matrix.

4. The financial live streaming compliance risk control method based on differential game theory according to claim 1, characterized in that, The zero-sum differential game mentioned in step S3 is specifically as follows: Treating the interaction between the risk control process and the livestreamer as a zero-sum differential game, the optimal intervention signal is obtained by approximately solving the HJB equation of the zero-sum differential game. Mathematically described ,in, The intervention energy cost matrix is ​​diagonally positive definite, where each diagonal element corresponds to the cost of different components of the intervention signal. Let be a value function, representing the cumulative risk value starting from state s, assuming both parties adopt optimal strategies. This is the intervention effect matrix, which represents how the intervention changes the state at the next moment.

5. The financial live streaming compliance risk control method based on differential game theory according to claim 1, characterized in that, The construction of the meta-learning update model in step S4 is specifically as follows: The optimal intervention signal Injected into the live stream in real time, the system continues to collect live data to obtain a new state vector after intervention. and actual risk loss ; The updated value function is obtained based on the meta-learning update rate. and the updated meta-learning rate Mathematically described ,in, This represents the update amount of the value function. The learning rate is the value function. For the risk loss predicted before intervention, for Regarding value functions gradient, The meta-learning rate, For the expectation of the meta-learning task, For the loss of yuan, This indicates an assignment operation; The updated value function and the meta-learning rate Used for the next round of intervention.

6. The financial live streaming compliance risk control method based on differential game theory according to claim 5, characterized in that, The safety supervision and intervention rollback in step S5 specifically refers to: Set intervention safety indicators Mathematically described ,in, To intervene in the proportion that is noticed by the audience, The upper limit of the allowed detection rate, The square of the weighted energy of the intervention signal is summed according to the perceived weight. To sense the upper limit of energy; when the intervention safety index When the value falls below a preset threshold, an intervention rollback is performed, reducing the intervention intensity to 0 and restoring the value function and meta-learning rate of the previous version, and then re-intervening.

7. A financial live-streaming compliance and risk control system based on differential game theory, characterized in that, The system is applied to the financial live streaming compliance risk control method based on differential game theory as described in any one of claims 1-6, including a perception mapping module, an implicit intent inference module, a zero-sum differential game module, a meta-learning update module, and a security supervision module; The perception mapping module is used to input the raw signal of the financial live broadcast stream and perform compressed perception state mapping to obtain a state vector. and noise covariance matrix ; The implicit intent inference module is used to infer based on the state vector. and the noise covariance matrix Perform implicit intent inference to obtain the estimated intent parameters at the current time. and its uncertainties; The zero-sum differential game module is used to base the state vector and the intent parameters To obtain the optimal intervention signal, a zero-sum differential game is performed. ; The meta-learning update module is used to update the optimal intervention signal. Injected into the live stream in real time, a new state vector is obtained. And construct a meta-learning update model to obtain the updated value function. Learning rate ; The safety monitoring module is used to base on the new state vector. Conduct safety monitoring and intervention rollback.

Images