A dynamic face recognition multi-modal fusion anti-cheating method and system

Abstract

The application relates to the technical field of face recognition, in particular to a dynamic face recognition multi-modal fusion anti-cheating method and system, which comprises the following steps: obtaining personnel configuration requirements, determining an anti-cheating level, a function characteristic switch and a threshold according to the personnel configuration requirements, and generating an anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold. The application improves the accuracy and authenticity of face recognition results.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of face recognition, and in particular to a dynamic face recognition multi-modal fusion anti-cheating method and system. BACKGROUND

[0002] With the rapid development of information technology, biometric recognition technology as an efficient, convenient and highly unique identity authentication method has been widely used in many fields. Among them, face recognition technology has become one of the research hotspots and mainstream application technologies in the field of biometric recognition due to its non-contact, intuitive and high user acceptance and other significant advantages.

[0003] In the field of public security, face recognition technology is widely used in the security check of transportation hubs such as airports, railway stations and subway stations. By quickly comparing the facial features of passengers with the information in the database, efficient and accurate personnel identity verification is achieved, which effectively improves the security and prevention ability of public places and effectively combats various illegal and criminal activities. In the field of financial payment, many banks and third-party payment platforms have introduced face recognition technology as an important supplementary means for user identity verification. When users handle account opening, transfer, payment and other businesses, they only need to perform face scanning through the camera to complete identity authentication, which greatly simplifies the operation process and improves the security and convenience of transactions. In addition, in the fields of access control attendance system, intelligent security monitoring, social entertainment application and other fields, face recognition technology also plays an indispensable role, which has profoundly changed people's way of life and work

[0004] Although the face recognition technology has made significant progress and has been widely applied, its security is also facing increasingly severe challenges in the actual application process, especially the continuous emergence of various cheating means, which has brought huge security risks to the traditional face recognition system. Among them, static attack means and dynamic attack means are the most common, static attack refers to static picture attack, which is the most common and easy to implement cheating means. Attackers obtain the photos of target users, print them out using printing equipment, or display them using electronic devices, trying to cheat the face recognition system. Because the traditional face recognition algorithm mainly focuses on the extraction and comparison of facial features in static images during early design, it lacks an effective judgment mechanism for whether the input is a real face, so it is easy to be broken through by static picture attack. This attack method is low in cost and easy to operate, which brings potential risks to some application scenarios with low security requirements. Dynamic attack means is video playback attack, which has stronger cheating. Attackers will record the face video of the target user in advance, and when identity verification is needed, they will play the video to simulate real face dynamics. Some traditional face recognition systems based on action instructions (such as blinking, opening mouth, turning head, etc.) can resist static picture attack to some extent, but for carefully recorded video playback attack, it is often difficult to effectively identify. Because video playback can naturally simulate these actions, the system mistakenly believes that it detects real face dynamic changes, resulting in illegal identity verification.

[0005] Therefore, how to build a more efficient, accurate and strong anti-attack dynamic face recognition anti-cheating system has become the key to ensuring the safe application of face recognition technology and promoting its further expansion of application boundaries. SUMMARY

[0006] To solve at least one of the above technical problems, the present application provides a dynamic face recognition multi-modal fusion anti-cheating method and system.

[0007] In a first aspect, the present application provides a dynamic face recognition multi-modal fusion anti-cheating method, which adopts the following technical solution:

[0008] A dynamic face recognition multi-modal fusion anti-cheating method applied to a configuration platform, comprising:

[0009] Obtain personnel configuration requirements, and determine an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements;

[0010] Generate an anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value.

[0011] By adopting the technical scheme, the personnel configuration requirement is the starting point of the entire anti-cheating process, which provides key information for subsequent strategy making. Different personnel configuration requirements reflect the differences in business scenarios and the personalized needs of users. Determining the anti-cheating level based on these requirements can preliminarily divide the intensity of coping with cheating. The setting of the function characteristic switch gives the system flexibility, and some anti-cheating functions can be enabled or disabled according to actual needs, avoiding resource waste. The determination of the threshold further refines the rules. The anti-cheating verification scheme is generated by comprehensively considering these factors, making the scheme more targeted, and the anti-cheating strategy can be adjusted according to different personnel configuration, effectively covering various potential cheating scenarios, reducing the possibility of cheating behavior, and improving the security and reliability of the system.

[0012] In a second aspect, the application provides a dynamic face recognition multi-modal fusion anti-cheating method, which adopts the following technical scheme:

[0013] A dynamic face recognition multi-modal fusion anti-cheating method applied to a mobile phone terminal, comprising:

[0014] Receiving an anti-cheating verification scheme sent from a configuration platform, the anti-cheating verification scheme being obtained by acquiring personnel configuration requirements, determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements, and generating the anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value;

[0015] Collecting a face recognition request of a user and determining face video information according to the face recognition request;

[0016] Determining a first calculation scheme and a first threshold standard according to the anti-cheating verification scheme, the first calculation scheme including inter-frame difference calculation and / or Euler angle variance calculation;

[0017] Calculating and processing the face video information according to the first calculation scheme to obtain a first detection result;

[0018] Judging whether the first detection result meets the first threshold standard, if yes, generating normal detection passing information, and if not, generating verification exception information.

[0019] By adopting the technical scheme, the anti-cheating checking scheme from the configuration platform is received, so that the mobile terminal can dynamically adapt to different anti-cheating strategies. The scheme is generated based on personnel configuration requirements, ensuring the pertinence and effectiveness of the anti-cheating measures. The face recognition request of the user is collected and the face video information is determined, providing a necessary data basis for subsequent anti-cheating detection. According to the anti-cheating checking scheme, a first calculation scheme is selected, such as inter-frame difference calculation or / and Euler angle variance calculation. These calculation schemes can effectively capture abnormal behaviors in the video, such as rapid blinking or unnatural head rotation. The first detection result is obtained by calculation processing, and is compared with a preset first threshold standard, which can quickly judge whether the user has cheating behavior. If the detection is passed, normal detection pass information is generated, ensuring the smooth experience of normal users; if the detection is abnormal, checking abnormal information is generated, so as to timely intercept potential cheating behaviors, thereby improving the security and user experience of the system.

[0020] In a possible implementation manner, the calculation processing on the face video information according to the first calculation scheme to obtain the first detection result comprises:

[0021] texture detection analysis is performed on the face video information to obtain a texture detection result;

[0022] It is judged whether the texture detection result is a non-screen texture. If yes, video gray frames and inter-frame absolute difference are determined based on the face video information;

[0023] It is judged whether the inter-frame absolute difference is not less than a preset absolute difference. If not, periodic peak values of the face video information are calculated based on the video gray frames;

[0024] key points are extracted from the face video information, and Euler angles of the obtained face key points are calculated to obtain key point Euler angles corresponding to different time nodes in the face video information;

[0025] The key point Euler angles are sorted in time sequence to obtain Euler angle time sequence curves corresponding to different positions of the face of the personnel;

[0026] The periodic peak values and the Euler angle time sequence curves are sorted to obtain the first detection result.

[0027] In a possible implementation manner, the texture detection analysis on the face video information to obtain the texture detection result comprises:

[0028] The pre-sequence video information and the subsequent video information are determined based on the face video information. The pre-sequence video information is environment video information shot by the personnel, and the subsequent video information is face video information;

[0029] determine environment information and first position information of the face according to the pre-video information;

[0030] determine reflection information of the shooting device and second position information according to the face video information;

[0031] determine a light position group of different positions in the environment where the person is located according to the environment information;

[0032] match the first position information with the light position group to obtain a standard light intensity;

[0033] determine reflection light intensity of the reflection of the shooting device to the face of the person based on the reflection information and the second position information;

[0034] determine whether the standard light intensity is greater than the reflection light intensity, if yes, simulate a light image of the standard light intensity irradiating to the face of the person, and perform texture detection on the light image to obtain a texture detection result, if not, simulate a light image of the reflection light intensity irradiating to the face of the person, and perform texture detection on the light image to obtain a texture detection result.

[0035] In a third aspect, the present application provides a dynamic face recognition multi-modal fusion anti-cheating method, which adopts the following technical solution:

[0036] A dynamic face recognition multi-modal fusion anti-cheating method applied to a first server, comprising:

[0037] receive an anti-cheating verification scheme sent from a configuration platform and face video information sent from a mobile phone terminal, the anti-cheating verification scheme is obtained by acquiring personnel configuration requirements, and the anti-cheating verification scheme is generated according to the anti-cheating level, the function characteristic switch and the threshold value, the face video information is obtained by collecting a face recognition request of a user, and the face video information is obtained according to the face recognition request;

[0038] determine whether the anti-cheating verification scheme contains a second calculation scheme, if yes, determine a second threshold standard based on the second calculation scheme, the second calculation scheme includes frequency domain calculation and / or sparse optical flow calculation and / or face key point detection and / or time domain window sliding;

[0039] perform calculation processing on the face video information according to the second calculation scheme to obtain a second detection result;

[0040] determine whether the second detection result meets the second threshold standard, if yes, generate standard detection pass information, if not, generate verification exception information.

[0041] By adopting the technical scheme, the first server can receive the anti-cheating verification scheme from the configuration platform and the face video information from the mobile phone end, realizing centralized processing and collaborative work of information. The anti-cheating verification scheme is generated based on personnel configuration requirements, ensuring the pertinence and effectiveness of the strategy. By judging whether the anti-cheating verification scheme contains a second calculation scheme, such as frequency domain calculation, sparse optical flow calculation, face key point detection or time domain window sliding, the server can flexibly select suitable anti-cheating means. These calculation schemes can deeply analyze the frequency domain features, motion trajectories, key point changes and anomalies in time series in the video, so as to accurately identify cheating behaviors. The second detection result obtained based on the second calculation scheme is compared with the preset second threshold standard, which can efficiently judge the compliance of user behavior. If the detection is passed, standard detection pass information is generated to protect the rights and interests of legitimate users; if the detection is abnormal, verification exception information is generated to timely block cheating behaviors. This multi-modal fusion anti-cheating mechanism significantly improves the security and detection accuracy of the system.

[0042] In a fourth aspect, the present application provides a dynamic face recognition multi-modal fusion anti-cheating method, which adopts the following technical scheme:

[0043] A dynamic face recognition multi-modal fusion anti-cheating method applied to a second server, comprising:

[0044] Receiving an anti-cheating verification scheme sent from a configuration platform and face video information sent from a mobile phone end, the anti-cheating verification scheme being obtained by acquiring personnel configuration requirements and determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements, the anti-cheating verification scheme being generated according to the anti-cheating level, the function characteristic switch and the threshold value, the face video information being obtained by collecting a face recognition request of a user and determining the face video information according to the face recognition request;

[0045] Judging whether the anti-cheating verification scheme contains a third calculation scheme, if it contains, determining a third threshold standard based on the third calculation scheme, the third calculation scheme including high-precision optical flow calculation and / or adversarial sample defense;

[0046] According to the third calculation scheme, the face video information is calculated and processed to obtain a third detection result;

[0047] Judging whether the third detection result meets the third threshold standard, if it meets, generating high-precision detection pass information, if it does not meet, generating verification exception information.

[0048] By adopting the technical scheme, the second server can receive and integrate key data from the configuration platform and the mobile terminal, and provide comprehensive information support for subsequent anti-cheating detection. The anti-cheating verification scheme is dynamically generated based on personnel configuration requirements, ensuring the pertinence and effectiveness of the strategy. By judging whether the anti-cheating verification scheme contains a third calculation scheme such as high-precision optical flow calculation or adversarial sample defense, the server can select more advanced and accurate anti-cheating means. High-precision optical flow calculation can analyze the motion trajectory in the video in detail and capture subtle abnormal behavior; adversarial sample defense can effectively resist malicious attacks and improve the robustness of the system. The third detection result obtained based on the third calculation scheme is strictly compared with the preset third threshold standard, which can accurately judge the compliance of user behavior. If the detection is passed, high-precision detection pass information is generated to ensure user experience; if the detection is abnormal, verification exception information is generated to timely block cheating behavior. This high-precision and multi-level anti-cheating mechanism significantly improves the security and detection accuracy of the system, providing a more reliable service environment for users.

[0049] In a fifth aspect, the present application provides a dynamic face recognition multi-modal fusion anti-cheating system, which adopts the following technical scheme:

[0050] A dynamic face recognition multi-modal fusion anti-cheating system applied to a configuration platform, comprising:

[0051] A configuration acquisition module configured to acquire personnel configuration requirements and determine an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements;

[0052] A scheme generation module configured to generate an anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value.

[0053] In a sixth aspect, the present application provides a dynamic face recognition multi-modal fusion anti-cheating system, which adopts the following technical scheme:

[0054] A dynamic face recognition multi-modal fusion anti-cheating system applied to a mobile terminal, comprising:

[0055] A first receiving module configured to receive an anti-cheating verification scheme sent by the configuration platform, wherein the anti-cheating verification scheme is generated by acquiring personnel configuration requirements, determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements, and generating the anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value;

[0056] A request acquisition module configured to acquire a face recognition request of a user and determine face video information according to the face recognition request;

[0057] A calculation determination module is configured to determine a first calculation scheme and a first threshold standard according to the anti-cheating verification scheme, wherein the first calculation scheme comprises inter-frame difference calculation and / or Euler angle variance calculation.

[0058] A first calculation module is configured to perform calculation processing on the face video information according to the first calculation scheme to obtain a first detection result.

[0059] A first judgment module is configured to judge whether the first detection result meets the first threshold standard, and if yes, generate normal detection passing information, and if not, generate verification exception information.

[0060] In a seventh aspect, the present application provides a dynamic face recognition multi-modal fusion anti-cheating system, which adopts the following technical scheme:

[0061] A dynamic face recognition multi-modal fusion anti-cheating system is applied to a first service end and comprises:

[0062] A second receiving module is configured to receive an anti-cheating verification scheme sent from a configuration platform and face video information sent from a mobile phone end, wherein the anti-cheating verification scheme is obtained by acquiring personnel configuration requirements, determining an anti-cheating level, a function characteristic switch and a threshold according to the personnel configuration requirements, and generating the anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold, and the face video information is obtained by collecting a face recognition request of a user and determining the face video information according to the face recognition request.

[0063] A second judgment module is configured to judge whether the anti-cheating verification scheme contains a second calculation scheme, and if yes, determine a second threshold standard based on the second calculation scheme, wherein the second calculation scheme comprises frequency domain calculation and / or sparse optical flow calculation and / or face key point detection and / or time domain window sliding.

[0064] A second calculation module is configured to perform calculation processing on the face video information according to the second calculation scheme to obtain a second detection result.

[0065] A third judgment module is configured to judge whether the second detection result meets the second threshold standard, and if yes, generate standard detection passing information, and if not, generate verification exception information.

[0066] In an eighth aspect, the present application provides a dynamic face recognition multi-modal fusion anti-cheating system, which adopts the following technical scheme:

[0067] A dynamic face recognition multi-modal fusion anti-cheating system is applied to a second service end and comprises:

[0068] The third receiving module is configured to receive an anti-cheating verification scheme sent from a configuration platform and face video information sent from a mobile phone end, wherein the anti-cheating verification scheme is obtained by acquiring personnel configuration requirements, determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements, and generating the anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value, and the face video information is obtained by collecting a face recognition request of a user and determining the face video information according to the face recognition request;

[0069] The fourth judging module is configured to judge whether the anti-cheating verification scheme contains a third calculation scheme, and if so, determine a third threshold standard based on the third calculation scheme, wherein the third calculation scheme includes high-precision optical flow calculation and / or adversarial sample defense.

[0070] The third calculation module is configured to calculate and process the face video information according to the third calculation scheme to obtain a third detection result.

[0071] The fifth judging module is configured to judge whether the third detection result meets the third threshold standard, and if so, generate high-precision detection pass information, and if not, generate verification exception information.

[0072] In a ninth aspect, the present application provides a computer readable storage medium, which adopts the following technical scheme:

[0073] A computer readable storage medium, which stores a computer program, when the computer program is executed in a computer, the computer executes a dynamic face recognition multi-modal fusion anti-cheating method according to any one of the first aspect to the fourth aspect.

[0074] In summary, the present application includes at least one of the following beneficial technical effects:

[0075] 1. Acquiring personnel configuration requirements is the starting point of the entire anti-cheating process, which provides key information for subsequent strategy formulation. Different personnel configuration requirements reflect the differences in business scenarios and the personalized needs of users. Determining the anti-cheating level based on these requirements can preliminarily divide the intensity of dealing with cheating. The setting of the function characteristic switch gives the system flexibility, which can enable or disable some anti-cheating functions according to actual needs, avoiding resource waste. The determination of the threshold value further refines the rules. The generation of the anti-cheating verification scheme based on these factors makes the scheme more targeted, which can adjust the anti-cheating strategy according to different personnel configuration situations, effectively cover various potential cheating scenarios, reduce the possibility of cheating behavior, and improve the security and reliability of the system.

[0076] 2. The anti-cheating verification scheme is received from the configuration platform, enabling the mobile terminal to dynamically adapt to different anti-cheating strategies. This scheme is generated based on personnel configuration requirements, ensuring the relevance and effectiveness of anti-cheating measures. Face recognition requests from users are collected and face video information is determined, providing a necessary data foundation for subsequent anti-cheating detection. According to the anti-cheating verification scheme, a first calculation scheme is selected, such as inter-frame difference calculation or / and Euler angle variance calculation, which can effectively capture abnormal behaviors in the video, such as rapid blinking or unnatural head rotation. The first detection result is obtained through calculation processing and compared with the preset first threshold standard, which can quickly determine whether the user has cheated. If the detection is passed, normal detection pass information is generated, ensuring a smooth experience for normal users; if the detection is abnormal, verification exception information is generated, which can timely intercept potential cheating behavior, thereby improving the security and user experience of the system.

[0077] 3. The first server can receive anti-cheating verification schemes from the configuration platform and face video information from the mobile terminal, realizing centralized processing and collaborative work of information. The anti-cheating verification scheme is generated based on personnel configuration requirements, ensuring the relevance and effectiveness of the strategy. By judging whether the anti-cheating verification scheme contains a second calculation scheme, such as frequency domain calculation, sparse optical flow calculation, face key point detection or time domain window sliding, the server can flexibly select appropriate anti-cheating means. These calculation schemes can deeply analyze the frequency domain features, motion trajectories, key point changes and time series anomalies in the video, thereby accurately identifying cheating behavior. The second detection result based on the second calculation scheme is compared with the preset second threshold standard, which can efficiently judge the compliance of user behavior. If the detection is passed, standard detection pass information is generated, ensuring the rights and interests of legitimate users; if the detection is abnormal, verification exception information is generated, which can timely block cheating behavior. This multi-modal fusion anti-cheating mechanism significantly improves the security and detection accuracy of the system.

[0078] 4. The second service end can receive and integrate key data from the configuration platform and the mobile end, providing comprehensive information support for subsequent anti-cheating detection. The anti-cheating verification scheme is dynamically generated based on personnel configuration requirements, ensuring the pertinence and effectiveness of the strategy. By judging whether the anti-cheating verification scheme contains a third calculation scheme such as high-precision optical flow calculation or adversarial sample defense, the service end can choose more advanced and accurate anti-cheating means. High-precision optical flow calculation can analyze the motion trajectory in the video in detail and capture subtle abnormal behavior; adversarial sample defense can effectively resist malicious attacks and improve the robustness of the system. The third detection result obtained based on the third calculation scheme is strictly compared with the preset third threshold standard, which can accurately judge the compliance of user behavior. If the detection is passed, high-precision detection pass information is generated to ensure user experience; if the detection is abnormal, verification exception information is generated to block cheating behavior in a timely manner. This high-precision, multi-level anti-cheating mechanism significantly improves the security and detection accuracy of the system, providing a more reliable service environment for users. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 A first flowchart of a dynamic face recognition multi-modal fusion anti-cheating method provided by an embodiment of the present application.

[0080] Figure 2 A second flowchart of a dynamic face recognition multi-modal fusion anti-cheating method provided by an embodiment of the present application.

[0081] Figure 3 A third flowchart of a dynamic face recognition multi-modal fusion anti-cheating method provided by an embodiment of the present application.

[0082] Figure 4 A fourth flowchart of a dynamic face recognition multi-modal fusion anti-cheating method provided by an embodiment of the present application.

[0083] Figure 5 A first structural diagram of a dynamic face recognition multi-modal fusion anti-cheating system provided by an embodiment of the present application.

[0084] Figure 6 A second structural diagram of a dynamic face recognition multi-modal fusion anti-cheating system provided by an embodiment of the present application.

[0085] Figure 7 A third structural diagram of a dynamic face recognition multi-modal fusion anti-cheating system provided by an embodiment of the present application.

[0086] Figure 8 A fourth structural diagram of a dynamic face recognition multi-modal fusion anti-cheating system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0087] The accompanying drawings are incorporated into the present application to provide further description. Figures 1-8 The present application is further described in detail.

[0088] The present application is further described in detail.

[0089] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0090] In addition, the term "and / or" in this paper is only a description of the association between the associated objects, which means that there may be three relationships, for example, A and / or B, which means that there are three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper, unless otherwise specified, generally represents a "or" relationship between the associated objects before and after.

[0091] The embodiments of the present application will be further described below in combination with the drawings of the specification.

[0092] The embodiments of the present application disclose a dynamic face recognition multi-modal fusion anti-cheating method based on a configuration platform.

[0093] Referring to Figure 1 A dynamic face recognition multi-modal fusion anti-cheating method is applied to a configuration platform, comprising:

[0094] Step S10, obtaining personnel configuration requirements, and determining the anti-cheating level, the function characteristic switch and the threshold value according to the personnel configuration requirements.

[0095] Specifically, the anti-cheating level refers to the intensity of anti-cheating measures of different levels divided according to the degree of cheating risk, business importance, system security requirements and other factors. It is used to indicate the strictness of the anti-cheating means adopted by the system when facing different cheating threats. For example, for ordinary enterprise internal training examinations, the anti-cheating level may be low, and only basic face recognition verification is required; while for important national examinations, the anti-cheating level is high, and a combination of high-level anti-cheating technologies such as multi-biometric identification and behavior analysis will be used. In the embodiments of the present application, the anti-cheating level is divided into anti-cheating light level, anti-cheating standard level and anti-cheating high-precision level, wherein the anti-cheating light level includes inter-frame difference calculation and Euler angle variance calculation, the anti-cheating standard level includes frequency domain calculation, sparse optical flow calculation, face key point detection and time domain window sliding, and the anti-cheating high-precision level includes high-precision optical flow calculation and adversarial sample defense.

[0096] The function characteristic switch refers to the control of enabling or disabling various algorithm function modules in the dynamic face recognition multi-modal fusion anti-cheating system, that is, the control of enabling and disabling each algorithm in each anti-cheating level. It is used to flexibly enable or disable specific anti-cheating functions according to actual needs. For example, in some cases, it may not be necessary to enable a specific algorithm module, in which case the function characteristic switch can be turned off; while in high-risk scenarios, all related function characteristic switches need to be turned on to enhance anti-cheating ability. The threshold refers to the critical value used to determine whether the detection index of various algorithms in the dynamic face recognition multi-modal fusion anti-cheating process is abnormal. It is used to indicate a reasonable range set in the anti-cheating system for various detection data, and when the data exceeds this range, it is determined that there may be cheating behavior. For example, in face similarity detection, a similarity threshold is set, and when the detected face similarity is lower than the threshold, the system will issue a cheating warning.

[0097] Step S11, generating an anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold.

[0098] Specifically, the anti-cheating level determines the algorithm package applied in this anti-cheating verification, the function characteristic switch determines the algorithms needed to be turned on in the algorithm package, and the threshold determines the threshold standard of the algorithms needed to be turned on, thereby determining the verification scheme of this anti-cheating verification, that is, the algorithms needed to be turned on and the threshold standard of the algorithms.

[0099] The embodiment of the application provides a dynamic face recognition multi-modal fusion anti-cheating method, obtaining personnel configuration requirements is the starting point of the entire anti-cheating process, which provides key information for subsequent strategy making. Different personnel configuration requirements reflect the differences of business scenarios and the personalized requirements of users. Determining the anti-cheating level based on these requirements can preliminarily divide the intensity of coping with cheating. The setting of the function characteristic switch gives flexibility to the system, and some anti-cheating functions can be enabled or disabled according to actual requirements, avoiding resource waste. The determination of the threshold further refines the rules. The anti-cheating verification scheme is generated by comprehensively considering these factors, so that the scheme is more targeted, the anti-cheating strategy can be adjusted according to different personnel configuration, various potential cheating scenarios can be effectively covered, the possibility of cheating behavior is reduced, and the security and reliability of the system are improved.

[0100] Reference Figure 2 A dynamic face recognition multi-modal fusion anti-cheating method is applied to a mobile phone end and includes the following steps.

[0101] In step S20, an anti-cheating verification scheme sent by a configuration platform is received.

[0102] The anti-cheating verification scheme is obtained by obtaining personnel configuration requirements, determining an anti-cheating level, a function characteristic switch and a threshold according to the personnel configuration requirements, and generating the anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold.

[0103] Specifically, the mobile phone end and the configuration platform establish a communication connection. The mobile phone end initiates a connection request to the configuration platform through a network protocol (such as HTTP, WebSocket, etc.), and the configuration platform establishes a stable communication channel after verifying the identity legality of the mobile phone end. This step ensures that the mobile phone end can normally receive data from the configuration platform. Then, the mobile phone end sends a message to the configuration platform to request an anti-cheating verification scheme. When the mobile phone end needs to perform dynamic face recognition verification, it sends a specific request message to the configuration platform, which contains relevant information of the mobile phone end (such as device identifier, user identifier, etc.), so that the configuration platform can accurately return the corresponding anti-cheating verification scheme. The configuration platform searches for the corresponding anti-cheating verification scheme according to the information in the request message. The configuration platform searches for the anti-cheating verification scheme corresponding to the mobile phone end and the user in the database or storage system according to the device identifier, user identifier and other information sent by the mobile phone end. The configuration platform sends the found anti-cheating verification scheme to the mobile phone end. The configuration platform sends the anti-cheating verification scheme to the mobile phone end in a specific data format (such as JSON, XML, etc.) through the established communication channel. After receiving the scheme, the mobile phone end stores it in the local memory or a specific storage area, so as to perform anti-cheating verification according to the scheme in the subsequent dynamic face recognition process.

[0104] Step S21, collect the face recognition request of the user, and determine the face video information according to the face recognition request.

[0105] Specifically, an implementation based on a mobile application development framework is adopted. First, in mobile application development, the API interface provided by the corresponding development framework (such as Android SDK of Android and Xcode development tool of iOS) is used to listen to the face recognition button click event on the user interface. When the user clicks the button, a face recognition request is triggered. Then, through the camera access permission application function provided by the framework, the user is applied to use the camera permission. After the user's authorization, the camera management class (such as CameraManager of Android and AVCaptureDevice of iOS) provided by the framework is used to obtain the available camera device. Then, the parameters of the camera are configured, such as resolution, frame rate, focus mode, etc., to meet the requirements of face video information collection. After that, the video capture class (such as Camera2 API of Android and AVCaptureSession of iOS) provided by the framework is used to start the camera and begin collecting video data. The collected video data will be returned in the form of frames, and the system will combine these frames into face video information and perform necessary preprocessing.

[0106] Step S22, determine the first calculation scheme and the first threshold standard according to the anti-cheating verification scheme.

[0107] Among them, the first calculation scheme includes inter-frame difference calculation and / or Euler angle variance calculation.

[0108] Specifically, the inter-frame difference calculation refers to comparing and calculating the image data between consecutive frames in the face video information. By analyzing the changes in pixel values, feature point positions, and other factors between adjacent frames, it can detect whether there are abnormal actions or behaviors. It is used to determine whether the user has used fast switching, shielding, and other cheating methods during the face recognition process. Because normal facial movements are relatively smooth and continuous between adjacent frames, cheating behaviors may cause large and unreasonable differences between frames. For example, if someone quickly switches a photo during face recognition, the image difference between adjacent frames will be significantly larger than when a normal face rotates or changes expressions. The Euler angle in the Euler angle variance calculation is a parameter used to describe the pose of an object in three-dimensional space, including pitch, yaw, and roll. The Euler angle variance calculation refers to the statistical calculation of the Euler angles corresponding to each frame in the face video information, and the variance value is obtained. It is used to analyze the pose changes of the face during the video process to determine whether there are abnormal pose jitter or fixed pose cheating behaviors. Under normal circumstances, the user's head will have natural and slight movements during the face recognition process, and the Euler angle will have a certain fluctuation range, and its variance value will be within a reasonable range. If the user uses a photo or video to cheat, the face pose may remain relatively fixed, and the Euler angle variance will be significantly smaller than normal. The first threshold standard refers to the judgment threshold set for each calculation result in the first calculation scheme, which is used to distinguish between normal face recognition behaviors and possible cheating behaviors. It is used to indicate that when the calculation result exceeds or is lower than the threshold, the system determines that there is a risk of cheating. Different calculation schemes and business scenarios will correspond to different threshold standards, which are usually obtained through a large amount of experimental data and actual business demand analysis. For example, for inter-frame difference calculation, a frame difference threshold can be set, and when the difference between adjacent frames exceeds the threshold, the system considers that there may be cheating behavior; for Euler angle variance calculation, a variance threshold can also be set, and when the Euler angle variance is less than the threshold, the system issues a cheating warning.

[0109] Step S23, according to the first calculation scheme, the face video information is calculated and processed to obtain the first detection result.

[0110] Specifically, the mobile phone end pre-processes the face video information. The pre-processing steps include image graying, converting color images to grayscale images to reduce data volume and computational complexity; image normalization, adjusting the size, brightness and contrast of the image and other parameters, so that the images of different frames have consistent format and feature range, facilitating subsequent calculation and processing. For example, the size of all image frames is uniformly adjusted to a specific resolution, ensuring that the brightness and contrast of the image are within a suitable range, avoiding the influence of image quality differences on the accuracy of the calculation results. Then, if the first calculation scheme includes inter-frame difference calculation, the mobile phone end extracts consecutive image frames from the pre-processed face video information. For each pair of adjacent frames, the system calculates the pixel difference between them. Common calculation methods include mean square error (MSE) and structural similarity index (SSIM). The mean square error measures the difference between the two frames by calculating the average of the squares of the differences in the gray values of the corresponding pixels, with a larger value indicating a more obvious difference; the structural similarity index considers the similarity of the two frames from the aspects of brightness, contrast and structure, with a value closer to 1 indicating a higher similarity. The system will calculate the difference between adjacent frames in the entire video sequence to obtain the inter-frame difference distribution. If the first calculation scheme includes Euler angle variance calculation, the system will use a face key point detection algorithm (such as the MediaPipe 468-point detection algorithm) to detect the key points of the face in each frame of image. According to the detected face key point coordinates, the system can calculate the Euler angles (pitch angle, yaw angle and roll angle) corresponding to each frame of image. For example, by using trigonometric functions and geometric transformations, the values of the Euler angles are calculated based on the positional relationship of the key points. Then, the mobile phone end calculates the variance value of all frame Euler angles, which reflects the fluctuation degree of the face posture in the video process. The mobile phone end generates the first detection result based on the calculated data.

[0111] For the embodiments of the present application, the face video information is subjected to texture detection analysis to obtain a texture detection result, and it is determined whether the texture detection result is a non-screen texture. If so, the video grayscale frame and the inter-frame absolute difference are determined based on the face video information, it is determined whether the inter-frame absolute difference is not less than a preset absolute difference, if not, the periodic peak value of the face video information is calculated based on the video grayscale frame, the key points of the face video information are extracted, and the Euler angles of the key points corresponding to different time nodes in the face video information are obtained by calculating the Euler angles of the obtained face key points. The key point Euler angles are arranged in time sequence to obtain the Euler angle time sequence curve corresponding to different positions of the face of the person. The periodic peak value and the Euler angle time sequence curve are arranged to obtain the first detection result.

[0112] Specifically, the texture detection analysis is performed on the face video information to obtain a texture detection result, including: determining, based on the face video information, preceding video information and subsequent video information, the preceding video information being environment video information captured by the person, and the subsequent video information being face video information. The environment information and first position information of the face are determined based on the preceding video information, and the reflection information and second position information of the shooting device are determined based on the face video information. The light position group of different positions of the environment in which the person is located is determined based on the environment information. The first position information is matched with the light position group to obtain a standard light intensity. The reflection light intensity of the shooting device reflected to the face of the person is determined based on the reflection information and the second position information. It is judged whether the standard light intensity is greater than the reflection light intensity. If yes, a light image of the standard light intensity irradiated to the face of the person is simulated, and texture detection is performed on the light image to obtain the texture detection result. If no, a light image of the reflection light intensity irradiated to the face of the person is simulated, and texture detection is performed on the light image to obtain the texture detection result.

[0113] In the embodiments of the present application, the step is split into multiple sub-tasks, distributed to different computing nodes and executed in parallel. For example, in a large face recognition system, multiple cameras simultaneously capture face video information, and each camera corresponds to a computing node. Each computing node respectively pre-processes and extracts the face region from the video stream collected by itself, and then sends the extracted face region data to the center computing node for texture detection analysis. The center computing node then comprehensively processes the detection results of multiple nodes to obtain the final texture detection result.

[0114] In the embodiments of the present application, the step is split into multiple sub-tasks, distributed to different computing nodes and executed in parallel. For example, in a large face recognition system, multiple cameras simultaneously capture face video information, and each camera corresponds to a computing node. Each computing node respectively pre-processes and extracts the face region from the video stream collected by itself, and then sends the extracted face region data to the center computing node for texture detection analysis. The center computing node then comprehensively processes the detection results of multiple nodes to obtain the final texture detection result.

[0115] Specifically, a time sequence coordinate system is created, the X-axis of the time sequence coordinate system is different time nodes, the Y-axis of the time sequence coordinate system is different numerical degrees, the Euler angles of the key points at the same position are mapped into the time sequence coordinate system according to the time sequence, and the mapped dots are connected at the end to obtain an Euler angle time sequence curve corresponding to different positions of the face of the person.

[0116] Specifically, whether the Euler angle change trend of different positions of the face of the person has a mutation trend is determined based on the Euler angle time sequence curve. If there is, a starting point position of the mutation trend is determined. If not, the starting point of each floating period of the Euler angle time sequence curve is taken as the starting point position. The Euler angle time sequence curve is segmented based on the starting point position, a plurality of time sequence curves are obtained, the curve change score corresponding to each time sequence curve is calculated and accumulated, a total change score is obtained, whether the total change score meets a standard change score range is determined, if it meets, it is determined that the Euler angle curve time sequence change meets the preset time sequence change standard, if it does not meet, it is determined that the Euler angle curve time sequence change does not meet the preset time sequence change standard.

[0117] The total change score is obtained by calculating and accumulating the curve change score corresponding to each time sequence curve, including: determining the highest degree and the lowest degree of each time sequence curve, and calculating the mean value of the highest degree and the lowest degree to obtain the degree mean value corresponding to each time sequence curve. The degree mean value is taken as the numerator, and the time length of each time sequence curve is taken as the denominator to obtain the change score corresponding to each time sequence curve. The change scores are calculated and accumulated to obtain the total change score.

[0118] In the embodiment of the present application, the standard change score range is obtained by: collecting the standard Euler angle time sequence curves corresponding to the face video information passing the verification in the historical time period, determining whether the Euler angle change trend of different positions of the face of the person has a mutation trend based on each standard Euler angle time sequence curve, if there is, a target point position of the mutation trend is determined, if not, the starting point of each floating period of the standard Euler angle time sequence curve is taken as the target point position. Each standard Euler angle time sequence curve is segmented based on the target point position to obtain a plurality of target time sequence curves corresponding to each standard Euler angle time sequence curve. The highest degree and the lowest degree are determined according to the target time sequence curve, the mean value of the highest degree and the lowest degree is calculated, and the calculated degree mean value is taken as the numerator, and the time length of the target time sequence curve is taken as the denominator to obtain the standard change value. The standard change values are calculated and accumulated to obtain the standard total change score corresponding to each standard Euler angle time sequence curve, and the standard change score range is determined based on the standard total change score.

[0119] Step S24, determining whether the first detection result meets the first threshold standard, if it meets, generating normal detection passing information, if it does not meet, generating verification exception information.

[0120] Specifically, the first threshold criterion is a judgment threshold value preset for each calculation index (such as an inter-frame difference value, an Euler angle variance value, etc.) in the first calculation scheme. These threshold values are determined through a large amount of experimental data, actual business demand analysis, and security risk assessment, and are used to distinguish between normal face recognition behavior and possible cheating behavior. Different calculation indexes correspond to different threshold ranges. When the calculation result is within the threshold range, the behavior is considered normal; if it is outside or below the range, there may be a risk of cheating.

[0121] For the embodiments of the present application, a rule engine (such as Drools) is used to define the judgment rules. The first detection result and the first threshold criterion are input as fact data into the rule engine, and the rule engine performs matching and reasoning according to the pre-defined rules, thereby determining to generate normal detection pass information or verification exception information.

[0122] The embodiments of the present application provide a dynamic face recognition multi-modal fusion anti-cheating method, which receives an anti-cheating verification scheme from a configuration platform, so that a mobile terminal can dynamically adapt to different anti-cheating strategies. The scheme is generated based on personnel configuration requirements, ensuring the pertinence and effectiveness of the anti-cheating measures. Face recognition requests of a user are collected and face video information is determined, providing a necessary data basis for subsequent anti-cheating detection. According to the anti-cheating verification scheme, a first calculation scheme is selected, such as inter-frame difference calculation or / and Euler angle variance calculation. These calculation schemes can effectively capture abnormal behaviors in the video, such as rapid blinking or unnatural head rotation. The first detection result is obtained through calculation processing, and is compared with the pre-set first threshold criterion, which can quickly determine whether the user has cheating behavior. If the detection passes, normal detection pass information is generated, ensuring a smooth experience for normal users; if the detection is abnormal, verification exception information is generated, so as to timely intercept potential cheating behavior, thereby improving the security and user experience of the system.

[0123] Reference Figure 3 A dynamic face recognition multi-modal fusion anti-cheating method applied to a first service end, comprising:

[0124] Step S30, receiving an anti-cheating verification scheme sent from a configuration platform and face video information sent from a mobile terminal.

[0125] Among them, the anti-cheating verification scheme is obtained by acquiring personnel configuration requirements, and determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements, and generating the anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value. The face video information is obtained by collecting face recognition requests of a user, and determining the face video information according to the face recognition requests.

[0126] Step S31, determine whether the second calculation scheme is included in the anti-cheating verification scheme, if yes, determine the second threshold standard based on the second calculation scheme.

[0127] The second calculation scheme includes frequency domain calculation and / or sparse optical flow calculation and / or face key point detection and / or time domain window sliding.

[0128] Specifically, frequency domain calculation is a method of converting signals (in this scenario, face video information) from time domain to frequency domain for analysis. In the time domain, the signal is a sequence that changes over time, while the frequency domain shows the intensity distribution of different frequency components in the signal. Through mathematical tools such as Fourier transform, the time domain signal can be decomposed into a series of superimposed sine and cosine waves of different frequencies, so that the characteristics of the signal in the frequency domain can be analyzed.

[0129] In the application of anti-cheating verification, when using electronic screens (such as mobile phones, tablets, etc.) to display faces for cheating, the screen refresh will produce specific frequency characteristics. Frequency domain calculation can capture these frequency components and compare them with the frequency characteristics of normal face videos. If abnormal high-frequency components (possibly from screen refresh rate) are detected, it can be determined that there is a risk of cheating. In addition, videos of different qualities will have different performances in the frequency domain. Cheating videos may have compression distortion, noise, etc. These problems will reflect in the frequency domain as abnormal specific frequency components. By analyzing the frequency domain characteristics, it can be determined whether the quality of the video is normal, thereby assisting in detecting cheating behavior. In the embodiments of the present application, Fourier transform is used to realize frequency domain calculation, including discrete Fourier transform (DFT) and fast Fourier transform (FFT). FFT is a fast algorithm for DFT, which can efficiently calculate the frequency domain representation of signals. In face video processing, two-dimensional Fourier transform can be performed on each frame of image to obtain the frequency spectrum of the image in the frequency domain. Frequency domain feature extraction: extract features from the frequency spectrum, such as frequency energy, center frequency of frequency distribution, bandwidth, etc. These features can be used for subsequent classification and judgment.

[0130] The optical flow in sparse optical flow calculation refers to the motion vector of pixel points in the image between consecutive frames. Sparse optical flow calculation is a method of estimating the optical flow of some pixels with obvious features (such as corner points, edges, etc.) in the image. It describes the motion of objects by tracking the position changes of these feature points between different frames.

[0131] In the application of anti-cheating verification, detecting static photo cheating: if a static photo is used for face recognition cheating, the face in the photo will not produce natural movement. Sparse optical flow calculation can detect that the motion vector of feature points is very small or almost unchanged, so as to judge whether there is static photo cheating behavior. Analyzing the authenticity of facial movements: during normal facial movements such as speaking, blinking, and making expressions, facial feature points will have corresponding movements. Sparse optical flow calculation can analyze whether these movements conform to the natural facial movement pattern, detect whether there is abnormal movement (such as unnatural facial shaking, incoordination, etc.), and further judge whether there is cheating. In the embodiments of the present application, feature point detection: using a feature point detection algorithm (such as Harris corner detection, SIFT feature point detection, etc.) to extract feature points with obvious features in the image. Optical flow estimation: using an optical flow estimation algorithm (such as Lucas-Kanade algorithm, Horn-Schunck algorithm, etc.) to calculate the motion vector of feature points between consecutive frames. Motion analysis: analyzing the calculated optical flow vector, such as calculating the average size, direction change, etc. of the optical flow vector, to judge whether there is abnormal movement.

[0132] Face key point detection refers to automatically locating the positions of key points with specific semantics in a face image, such as the corners of the eyes, the tip of the nose, and the corners of the mouth. These key points can accurately describe the shape and structure information of the face. In the application of anti-cheating verification, detecting facial expression abnormalities: cheaters may try to deceive the system by using some unnatural facial expressions. Face key point detection can track the position changes of key points in real time, and analyze whether the facial expressions conform to the normal expression change rules. For example, if it is detected that the eyes do not blink for a long time, the shape of the mouth is abnormal, etc., there may be cheating suspicion. Verifying the authenticity of the face: by analyzing the geometric relationship and relative position between key points, it can be judged whether the face is a real three-dimensional face. Some cheating methods (such as 3D model masks) may cause differences in the geometric relationship between key points and real faces, so as to be detected.

[0133] In the embodiments of the present application, a large amount of image data labeled with face key points is used to train a machine learning model (such as a convolutional neural network CNN), so that the model learns to predict the positions of key points from face images. Regression method can also be used: the coordinates of the key points are directly regressed from the features of the face image by using a regression algorithm. Post-processing optimization: post-processing optimization is performed on the detected key points, such as filtering, smoothing, etc., to improve the accuracy of key point positioning.

[0134] Time window sliding is a method of analyzing signals in the time dimension. It divides continuous time series data into multiple overlapping or non-overlapping time windows, then processes and analyzes the data within each window, and observes the changes in the data within the window to obtain the characteristics and patterns of the time series.

[0135] In the application of anti-cheating verification, it detects abnormal behavior in a short time: cheating behavior may show abnormal characteristics in a short time. Time window sliding can set appropriate time window size to analyze the behavior of the face video in each window. For example, detect whether the rotation angle of the face, the speed of expression change, etc. in a short time exceeds the normal range, so as to find cheating behavior. Capture the change of behavior pattern: normal face recognition behavior usually has certain pattern and rule. Time window sliding can analyze the change of behavior pattern in different windows. If it is found that the behavior pattern changes dramatically in a short time, there may be a cheating risk.

[0136] In the embodiments of the present application, the size of the time window and the sliding step are determined according to the actual needs. The window size can be estimated according to the duration of the cheating behavior, and the sliding step can be set to a part of the window size to ensure a certain overlap between the windows and improve the continuity of the analysis. The features of the face video data in each window are extracted, such as inter-frame difference, key point motion trajectory, etc. Then statistical analysis, machine learning, etc. are used to analyze the features of each window to determine whether the behavior in the window is normal. At the same time, the feature changes between different windows are compared to find abnormal behavior patterns.

[0137] Step S32, according to the second calculation scheme, the face video information is calculated and processed to obtain the second detection result.

[0138] Specifically, if the second calculation scheme includes frequency domain calculation, the first server will perform Fourier transform on the face video frame to extract frequency domain features such as frequency energy, frequency distribution, etc. If it contains sparse optical flow calculation, the first server will detect feature points in the video frame and calculate the motion vector of these feature points between consecutive frames to analyze the change of the optical flow field. If it contains face key point detection, the first server will locate the key points in the face image, such as the positions of the eyes, nose, mouth, etc. and track the motion trajectory of the key points. If it contains time window sliding, the first server will divide the video information into multiple time windows and calculate and process the data in each window to analyze the change of the behavior pattern. The first server will comprehensively analyze the results obtained by various calculation methods to generate the second detection result.

[0139] Step S33, judge whether the second detection result meets the second threshold standard, if it meets, generate standard detection passing information, if it does not meet, generate verification exception information.

[0140] Specifically, if the second threshold criterion is a single numerical threshold value, the system compares the value of the second detection result with the threshold value. For example, if the second threshold criterion specifies a frequency domain energy threshold value of 80, and the second detection result obtains a frequency domain energy value of 70, since 70<80, it is determined that the second detection result meets the second threshold criterion. Interval threshold judgment: if the second threshold criterion is a numerical interval, the system checks whether the value of the second detection result is within the interval. For example, the second threshold criterion specifies a comprehensive score interval of [70, 100], and the second detection result obtains a score of 85, since 85 is within the interval [70, 100], it is determined that the second detection result meets the second threshold criterion. Multi-condition combination judgment: when the second threshold criterion contains multiple conditions, the system needs to satisfy all conditions at the same time to determine that the second detection result meets the criterion. For example, the second threshold criterion specifies that the face key point motion amplitude is greater than 0.5 and the optical flow field change rate is less than 0.2, and the second detection result has a face key point motion amplitude of 0.6 and an optical flow field change rate of 0.15, since both conditions are met, it is determined that the second detection result meets the second threshold criterion.

[0141] Referring to Figure 4 A dynamic face recognition multi-modal fusion anti-cheating method applied to a second server, comprising:

[0142] Step S40, receiving an anti-cheating verification scheme sent from a configuration platform and face video information sent from a mobile phone end.

[0143] Among them, the anti-cheating verification scheme is obtained by obtaining personnel configuration requirements, and the anti-cheating level, function characteristic switch and threshold value are determined according to the personnel configuration requirements, and the anti-cheating verification scheme is generated according to the anti-cheating level, function characteristic switch and threshold value, and the face video information is obtained by collecting the face recognition request of the user, and the face video information is determined according to the face recognition request.

[0144] Step S41, judging whether the third calculation scheme is contained in the anti-cheating verification scheme, if yes, determining the third threshold criterion based on the third calculation scheme.

[0145] Among them, the third calculation scheme includes high-precision optical flow calculation and / or adversarial sample defense.

[0146] Specifically, the optical flow is the instantaneous velocity of the pixel movement of the spatial moving object on the observation imaging plane, and is a method of finding the correspondence between the previous frame and the current frame by using the change of the pixel in the time domain and the correlation between the adjacent frames, so as to calculate the motion information of the object between the adjacent frames. High-precision optical flow calculation is crucial for tasks that require accurate motion estimation. In the embodiments of the present application, the high-precision optical flow calculation is realized based on the RAFT of PyTorch, and the high-precision optical flow estimation is realized through the recursive update of the Cost Volume and the iterative optimization of the GRU. The single fixed optical flow field can be maintained and updated at high resolution, which reduces the prediction error rate caused by low resolution, reduces the probability of missing small and fast moving targets, and also reduces the number of iterations usually required by training of more than 1M parameters.

[0147] Adversarial sample attack refers to making a slight modification to the original data to make the machine learning model produce an incorrect classification result. Adversarial sample defense refers to improving the machine learning model to better resist adversarial sample attacks.

[0148] Adversarial sample attack methods include fast gradient sign method (FGSM): based on the original data, a perturbation term is added iteratively multiple times, so that the perturbed data can deceive the machine learning model. Attack based on generative adversarial network (GAN): use GAN to generate adversarial samples, so that the generated samples can deceive the target model.

[0149] Adversarial sample defense methods: when training the machine learning model, add adversarial samples to make the model better adapt to adversarial sample attacks. For a given training data set, generate adversarial samples, mix the adversarial samples and the original data together to get a new training data set, and train the machine learning model on the new training data set. Use less information or more difficult to calculate gradients to generate adversarial samples, so that attackers are difficult to generate effective adversarial samples through gradient methods. Train the model at a higher temperature, and then use the model to train the final model at a lower temperature, so that the output of the model is smoother, thereby reducing the impact of adversarial samples. Integrate multiple different machine learning models to improve the robustness of the model. Train multiple different machine learning models, and when testing, integrate the prediction results of multiple models to get the final prediction result. Perform various conversion methods on the prediction sample to reduce the possible disturbance, and then input the converted sample into the original model for prediction, so that the adversarial sample is correctly classified again. For example, perform compression, noise reduction and other operations on the input image. Add an external model to detect whether the input sample is an adversarial sample without identifying the adversarial sample as a correct label.

[0150] Step S42, calculating and processing the face video information according to a third calculation scheme to obtain a third detection result.

[0151] In step S43, it is judged whether the third detection result meets the third threshold criterion. If yes, high-precision detection pass information is generated. If no, verification exception information is generated.

[0152] With reference to Figure 5 The embodiment of the present application also discloses a dynamic face recognition multi-modal fusion anti-cheating system 50 based on a configuration platform.

[0153] The dynamic face recognition multi-modal fusion anti-cheating system 50 is applied to the configuration platform and comprises:

[0154] The configuration acquisition module 51 is configured to acquire a personnel configuration requirement, and determine an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirement.

[0155] The scheme generation module 52 is configured to generate an anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value.

[0156] With reference to Figure 6 The embodiment of the present application also discloses a dynamic face recognition multi-modal fusion anti-cheating system 60 based on a mobile phone terminal.

[0157] The dynamic face recognition multi-modal fusion anti-cheating system 60 is applied to the mobile phone terminal and comprises:

[0158] The first receiving module 61 is configured to receive an anti-cheating verification scheme sent by the configuration platform. The anti-cheating verification scheme is generated by acquiring a personnel configuration requirement, determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirement, and generating the anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value.

[0159] The request acquisition module 62 is configured to acquire a face recognition request of a user, and determine face video information according to the face recognition request.

[0160] The calculation determination module 63 is configured to determine a first calculation scheme and a first threshold criterion according to the anti-cheating verification scheme. The first calculation scheme comprises inter-frame difference calculation and / or Euler angle variance calculation.

[0161] The first calculation module 64 is configured to perform calculation processing on the face video information according to the first calculation scheme, to obtain a first detection result.

[0162] The first judgment module 65 is configured to judge whether the first detection result meets the first threshold criterion. If yes, normal detection pass information is generated. If no, verification exception information is generated.

[0163] With reference to Figure 7The embodiment of the application also discloses a dynamic face recognition multi-modal fusion anti-cheating system 70 based on the first server side.

[0164] The dynamic face recognition multi-modal fusion anti-cheating system 70 is applied to the first server and comprises the following modules.

[0165] The second receiving module 71 is configured to receive the anti-cheating verification scheme sent by the configuration platform and the face video information sent by the mobile phone terminal, wherein the anti-cheating verification scheme is obtained by acquiring personnel configuration requirements and determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements, and the anti-cheating verification scheme is generated according to the anti-cheating level, the function characteristic switch and the threshold value; and the face video information is obtained by collecting a face recognition request of a user and determining the face video information according to the face recognition request.

[0166] The second judging module 72 is configured to judge whether the anti-cheating verification scheme contains a second calculation scheme, and if yes, determine a second threshold value standard based on the second calculation scheme, wherein the second calculation scheme comprises frequency domain calculation, sparse optical flow calculation, face key point detection and time domain window sliding.

[0167] The second calculation module 73 is configured to perform calculation and processing on the face video information according to the second calculation scheme, and obtain a second detection result.

[0168] The third judging module 74 is configured to judge whether the second detection result meets the second threshold value standard, and if yes, generate standard detection passing information, and if not, generate verification exception information.

[0169] Reference Figure 8 The embodiment of the application also discloses a dynamic face recognition multi-modal fusion anti-cheating system 80 based on the second server side.

[0170] The dynamic face recognition multi-modal fusion anti-cheating system 80 is applied to the second server and comprises the following modules.

[0171] The third receiving module 81 is configured to receive the anti-cheating verification scheme sent by the configuration platform and the face video information sent by the mobile phone terminal, wherein the anti-cheating verification scheme is obtained by acquiring personnel configuration requirements and determining an anti-cheating level, a function characteristic switch and a threshold value according to the personnel configuration requirements, and the anti-cheating verification scheme is generated according to the anti-cheating level, the function characteristic switch and the threshold value; and the face video information is obtained by collecting a face recognition request of a user and determining the face video information according to the face recognition request.

[0172] The fourth judgment module 82 is used to determine whether the anti-cheating verification scheme includes a third calculation scheme. If it does, the third threshold standard is determined based on the third calculation scheme. The third calculation scheme includes high-precision optical flow calculation and / or adversarial sample defense.

[0173] The third calculation module 83 is used to calculate and process the face video information according to the third calculation scheme to obtain the third detection result;

[0174] The fifth judgment module 84 is used to determine whether the third detection result meets the third threshold standard. If it does, high-precision detection pass information is generated; if it does not, verification exception information is generated.

[0175] The following describes a computer-readable storage medium provided by an embodiment of this application. The computer-readable storage medium described below can be referred to in correspondence with the method described above.

[0176] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the above-described dynamic face recognition multimodal fusion anti-cheating method.

[0177] Since the embodiments of the computer-readable storage medium portion correspond to the embodiments of the method portion, please refer to the description of the embodiments of the method portion for the embodiments of the computer-readable storage medium portion.

[0178] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0179] The above are only some embodiments of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A dynamic face recognition multi-modal fusion anti-cheating method, characterized in that, Be applied to configuration platform and mobile phone end, wherein, the method executed by the configuration platform comprises: Obtain personnel configuration requirements, and determine anti-cheating level, function characteristic switch and threshold value according to the personnel configuration requirements; Generate anti-cheating verification scheme according to anti-cheating level, function characteristic switch and threshold value; The method executed by the mobile phone end comprises: Receive the anti-cheating verification scheme sent by the configuration platform, which is obtained by obtaining personnel configuration requirements, determining anti-cheating level, function characteristic switch and threshold value according to the personnel configuration requirements, and generating anti-cheating verification scheme according to anti-cheating level, function characteristic switch and threshold value; Collect user's face recognition request, and determine face video information according to the face recognition request; Determine first calculation scheme and first threshold standard according to the anti-cheating verification scheme, wherein the first calculation scheme comprises interframe difference calculation and Euler angle variance calculation; According to the first calculation scheme, the face video information is calculated and processed to obtain the first detection result; The method for calculating and processing the face video information according to the first calculation scheme to obtain the first detection result comprises: Texture detection analysis is performed on the face video information to obtain texture detection result; Determine whether the texture detection result is non-screen texture, if yes, determine video gray frame and interframe absolute difference based on the face video information; Determine whether the interframe absolute difference is not less than the preset absolute difference, if not, calculate the periodic peak value of the face video information based on the video gray frame; Key points are extracted from the face video information, and Euler angle calculation is performed on the obtained face key points to obtain key point Euler angle corresponding to different time nodes in the face video information; The key point Euler angle is arranged in time sequence to obtain Euler angle time sequence curve corresponding to different positions of personnel face; The periodic peak value and the Euler angle time sequence curve are arranged to obtain the first detection result; Determine whether the first detection result meets the first threshold standard, if yes, generate normal detection pass information, if not, generate verification exception information; The method further comprises: Determine whether the Euler angle change trend of different positions of personnel face exists sudden change trend based on the Euler angle time sequence curve, if yes, determine the starting point of the sudden change trend, if not, take the starting point of each floating period of the Euler angle time sequence curve as the starting point, segment the Euler angle time sequence curve based on the starting point to obtain multiple time sequence curves, calculate and accumulate the curve change score corresponding to each time sequence curve to obtain total change score, and determine whether the total change score meets the standard change score range.

2. The dynamic face recognition multi-modal fusion anti-cheating method according to claim 1, characterized in that, The method for performing texture detection analysis on the face video information to obtain texture detection result comprises: Determine presequence video information and subsequent video information based on the face video information, wherein the presequence video information is environmental video information shot by personnel, and the subsequent video information is face video information; Determine environmental information and first position information where the face is located according to the presequence video information; Determine the reflection information and the second position information of the shooting device according to the face video information; Determine the light position group of different positions in the environment where the personnel is located according to the environment information; Match the first position information with the light position group to obtain the standard light intensity; Determine the reflected light intensity of the shooting device reflected to the face of the personnel based on the reflection information and the second position information; Determine the reflected light intensity of the shooting device reflected to the face of the personnel based on the reflection information and the second position information; 3. A dynamic face recognition multi-modal fusion anti-cheating system, characterized in that, If the standard light intensity is greater than the reflected light intensity, simulate the light image irradiated to the face of the personnel by the standard light intensity, and perform texture detection on the light image to obtain the texture detection result; if the standard light intensity is less than the reflected light intensity, simulate the light image irradiated to the face of the personnel by the reflected light intensity, and perform texture detection on the light image to obtain the texture detection result. Be applied to configuration platform and cell -phone end, wherein, the module that configuration platform carries out, include: The configuration acquisition module is used for obtaining personnel configuration requirements, and determining the anti-cheating level, the function characteristic switch and the threshold value according to the personnel configuration requirements; The scheme generation module is used for generating an anti-cheating verification scheme according to the anti-cheating level, the function characteristic switch and the threshold value; The module that the cell -phone end carries out, include: The first receiving module is used for receiving the anti-cheating verification scheme sent from the configuration platform, and the anti-cheating verification scheme is obtained by obtaining personnel configuration requirements, and determining the anti-cheating level, the function characteristic switch and the threshold value according to the personnel configuration requirements, and the anti-cheating verification scheme is generated according to the anti-cheating level, the function characteristic switch and the threshold value; The request acquisition module is used for acquiring the face recognition request of the user, and determining the face video information according to the face recognition request; The calculation determination module is used for determining a first calculation scheme and a first threshold standard according to the anti-cheating verification scheme, and the first calculation scheme includes interframe difference calculation and Euler angle variance calculation; The first calculation module is used for calculating and processing the face video information according to the first calculation scheme to obtain a first detection result; When the first calculation module calculates and processes the face video information according to the first calculation scheme to obtain a first detection result, it is specifically used for: Texture detection analysis is performed on the face video information to obtain a texture detection result; Determine whether the texture detection result is a non-screen texture, if yes, determine the video gray frame and the interframe absolute difference based on the face video information; Determine whether the interframe absolute difference is not less than a preset absolute difference, if not, calculate the periodic peak value of the face video information based on the video gray frame; Key points are extracted from the face video information, and Euler angles are calculated based on the obtained face key points to obtain the key point Euler angles corresponding to different time nodes in the face video information; The key point Euler angles are arranged in time sequence to obtain the Euler angle time sequence curve corresponding to different positions of the face of the personnel; The periodic peak value and the Euler angle time sequence curve are arranged to obtain a first detection result; The first judging module is configured to judge whether the first detection result meets the first threshold standard, and if yes, generate normal detection passing information, and if not, generate verification exception information; The system further comprises: The system further comprises: The system further comprises: determining a starting point of the mutation trend if the mutation trend exists, and if not, taking a starting point of each floating period of the Euler angle time series curve as a starting point, segmenting the Euler angle time series curve based on the starting point, obtaining a plurality of time series curves, calculating and accumulating a curve change score corresponding to each time series curve, and obtaining a total change score, and judging whether the total change score meets a standard change score range.

Images