Data detection method and related device

Abstract

The embodiment of the invention discloses a data detection method and a related device, and the method comprises the steps: detecting a first to-be-detected image through a plurality of multi-type universal classification models, and determining a first undetermined result of the first to-be-detected image together according to a plurality of obtained first probabilities, thereby improving the accuracy of the first undetermined result. In order to expand the range of an application scene, multiple scene supplementary models are trained in advance, and different scene supplementary models are suitable for different detection scene types, so that a target scene supplementary model corresponding to a scene identifier is obtained from the multiple scene supplementary models based on the scene identifier, and the scene is detected. And the first to-be-detected image is detected based on the target scene supplement model to obtain a second undetermined result, so that the accuracy of the second undetermined result is improved. Through the first undetermined result with universality and the second undetermined result with pertinence, the first target detection result of the first to-be-detected image is comprehensively determined, and the detection accuracy is further improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a data detection method and related device. BACKGROUND

[0002] With the rapid development of electronic, computer, communication and other technologies, terminal devices (such as mobile phones, tablet computers, etc.) have been widely used in people's daily life. In order to meet people's demand for the functional diversity of terminal devices, various applications (Application) have emerged.

[0003] When a user uses an application on a terminal device, some abnormal behaviors may occur, for example, when playing a game through a game application, the behavior of using a plug-in may occur. In related technologies, the log data of the application may be collected, and the application log data is detected in combination with a model, so as to determine whether the application has abnormal behavior.

[0004] However, the accuracy of detection based on this method is low, and the application range is small. SUMMARY

[0005] In order to solve the above technical problems, the present application provides a data detection method and related device for improving the accuracy of detection and expanding the application range.

[0006] The embodiments of the present application disclose the following technical solutions:

[0007] In one aspect, the present application provides a data detection method, which comprises:

[0008] Obtaining a first to-be-detected image and a scene identifier, the first to-be-detected image being an image obtained by taking a screenshot during the running of a target application, and the scene identifier being used to identify the detection scene type corresponding to the target application;

[0009] Detecting the first to-be-detected image by a plurality of multi-type general classification models respectively to obtain a first probability that the first to-be-detected image is of a target type output by each of the multi-type general classification models, and determining a first pending result of the first to-be-detected image according to a plurality of the first probabilities, wherein the plurality of multi-type general classification models are different models;

[0010] According to the scene identifier, a target scene supplement model suitable for the target application is determined from a plurality of scene supplement models, and the first to-be-detected image is detected by the target scene supplement model to obtain a second pending result, wherein different scene supplement models are suitable for different detection scene types;

[0011] According to the first undetermined result and the second undetermined result, a first target detection result of the first to-be-detected image is determined, and the first target detection result is used to identify whether a type of the first to-be-detected image is the target type.

[0012] In another aspect, an embodiment of the present application provides a data detection device, and the device comprises an acquisition unit, a first detection unit, a second detection unit and a determination unit.

[0013] The acquisition unit is configured to acquire a first to-be-detected image and a scene identifier, the first to-be-detected image is an image obtained by taking a screenshot during running of a target application, and the scene identifier is used to identify a detection scene type corresponding to the target application.

[0014] The first detection unit is configured to detect the first to-be-detected image by using a plurality of multi-type general classification models respectively, to obtain a first probability that the first to-be-detected image is of a target type output by each of the plurality of multi-type general classification models, and to determine a first undetermined result of the first to-be-detected image according to a plurality of the first probabilities, wherein the plurality of multi-type general classification models are different models.

[0015] The second detection unit is configured to determine a target scene supplement model suitable for the target application from a plurality of scene supplement models according to the scene identifier, to detect the first to-be-detected image by using the target scene supplement model, and to obtain a second undetermined result, wherein different scene supplement models are suitable for different detection scene types.

[0016] The determination unit is configured to determine a first target detection result of the first to-be-detected image according to the first undetermined result and the second undetermined result, and the first target detection result is used to identify whether a type of the first to-be-detected image is the target type.

[0017] In another aspect, an embodiment of the present application provides a computer device, and the computer device comprises a processor and a memory.

[0018] The memory is configured to store a computer program and transmit the computer program to the processor.

[0019] The processor is configured to execute the method according to the instructions in the computer program.

[0020] In another aspect, an embodiment of the present application provides a computer readable storage medium, and the computer readable storage medium is configured to store a computer program, and the computer program is configured to execute the method according to the above aspect.

[0021] In another aspect, an embodiment of the present application provides a computer program product or computer program, which comprises computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method of the above aspect.

[0022] From the above technical solution, in the first aspect, instead of using application log data for identification, a screenshot is taken during the running of the target application, and the obtained image is used as the first detection image. Since the first detection image is a real process, it is less likely to be tampered with and is not easy to be attacked, thereby improving the accuracy of subsequent detection. In the second aspect, instead of using a binary classification model for a certain abnormal behavior for detection, a multi-classification model, i.e., a multi-type general classification model, is used for detection, thereby realizing detection of multiple abnormal behaviors by one model. In the third aspect, in order to improve the detection accuracy of the multi-type general classification model, instead of using one model, multiple multi-type general classification models are used to detect the first detection image respectively, to obtain the first probability that the first detection image output by the multi-type general classification model is of a target type, thereby determining the first pending result of whether the type of the first detection image is the target type according to multiple first probabilities, to improve the accuracy of the first pending result. In the fourth aspect, in order to expand the range of application scenarios, multiple scene supplement models are pre-trained, different scene supplement models are suitable for different detection scene types, thereby obtaining a target scene supplement model suitable for the detection scene type corresponding to the scene identifier from the multiple scene supplement models based on the obtained scene identifier, and then detecting the first detection image based on the target scene supplement model to obtain a second pending result, thereby improving the accuracy of the second pending result. In the fifth aspect, the first pending result with universality obtained based on the multiple multi-type general classification models and the second pending result with pertinence for the corresponding scene obtained based on the target scene supplement model are comprehensively determined to obtain the first target detection result of the first detection image, i.e., whether the first detection image is of the target type, such as whether the first detection image includes abnormal behavior, thereby further improving the accuracy of detection. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0024] Figure 1An application scenario of a data detection method provided by an embodiment of the present application is shown in the figure;

[0025] Figure 2 A specific application scenario of a data detection method provided by an embodiment of the present application is shown in the figure;

[0026] Figure 3 A specific application scenario of a data detection method provided by an embodiment of the present application is shown in the figure;

[0027] Figure 4 A specific application scenario of a data detection method provided by an embodiment of the present application is shown in the figure;

[0028] Figure 5 A flowchart of a data detection method provided by an embodiment of the present application is shown in the figure;

[0029] Figure 6 A model used by a data detection method provided by an embodiment of the present application is shown in the figure;

[0030] Figure 7 A detection based on a local graph general classification model provided by an embodiment of the present application is shown in the figure;

[0031] Figure 8 A method for marking an opposite game character by a box provided by an embodiment of the present application is shown in the figure;

[0032] Figure 9 A method for marking an opposite game character by a bone provided by an embodiment of the present application is shown in the figure;

[0033] Figure 10 A blood volume of a game character provided by an embodiment of the present application is shown in the figure;

[0034] Figure 11 A detection based on a multi-type special classification model provided by an embodiment of the present application is shown in the figure;

[0035] Figure 12 A detection based on a multi-type special classification model provided by an embodiment of the present application is shown in the figure;

[0036] Figure 13 A game plug-in including a function directory provided by an embodiment of the present application is shown in the figure;

[0037] Figure 14 A detection based on an easily recognizable general classification model and a character recognition model provided by an embodiment of the present application is shown in the figure;

[0038] Figure 15 A detection based on a multi-type special classification model and an easily recognizable special classification model provided by an embodiment of the present application is shown in the figure;

[0039] Figure 16 A schematic diagram of detection based on a whole image classification model provided for an embodiment of the present application;

[0040] Figure 17 A schematic diagram of detection based on a first whole image classification model and a second whole image classification model provided for an embodiment of the present application;

[0041] Figure 18 A schematic diagram of detection based on a local image special classification model provided for an embodiment of the present application;

[0042] Figure 19 A schematic diagram of detection based on a first local image special classification model and a second local image special classification model provided for an embodiment of the present application;

[0043] Figure 20 A schematic diagram of a data detection system provided for an embodiment of the present application;

[0044] Figure 21 A schematic diagram of data detection for a real-time processing scenario provided for an embodiment of the present application;

[0045] Figure 22 A structural schematic diagram of a data detection apparatus provided for an embodiment of the present application;

[0046] Figure 23 A structural schematic diagram of a server provided for an embodiment of the present application;

[0047] Figure 24 A structural schematic diagram of a terminal device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0048] Embodiments of the present application are described below with reference to the accompanying drawings.

[0049] In the related art, application log data is generally used for detection, but the application log data is vertical data, and numerical data is easy to modify or resist, thereby resulting in low accuracy of the detection result. Moreover, the application log data is generally detected based on one model, thereby obtaining whether the application is a certain abnormal behavior, i.e., is the abnormal behavior or is not the abnormal behavior, i.e., the model is generally a binary classification model, and can only recognize one kind of abnormal behavior, thereby resulting in a small application range of detection. However, there are many kinds of abnormal behaviors in the application, and only one kind of abnormal behavior is detected, which results in low accuracy of detection. Moreover, only one model is used for detection, and the detection result depends on the detection accuracy of the model, and if the detection accuracy of the model is low, the detection accuracy may also be low.

[0050] Based on this, the embodiment of the present application provides a data detection method and related device, which detects the first to-be-detected image through a plurality of multi-type general classification models and a target scene supplement model suitable for a target application, determines the first target detection result based on the obtained first and second to-be-determined results. Thus, the accuracy of detection is improved, and the application range is expanded.

[0051] The data detection method provided by the present application can be applied to a computer device with data detection capability, such as a terminal device and a server.

[0052] The terminal device can be a desktop computer, a notebook computer, a smart phone, a tablet computer, an Internet of Things device, and a portable wearable device. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle-mounted device, etc. The smart vehicle-mounted device can be a vehicle-mounted navigation terminal and a vehicle-mounted computer, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc., but is not limited thereto.

[0053] The server can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server or a server cluster providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN), and big data and artificial intelligence platforms, etc. The terminal device and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application.

[0054] In order to facilitate understanding of the data detection method provided by the embodiment of the present application, the application scenario of the data detection method is exemplarily introduced below taking the server as an example.

[0055] Referring to Figure 1 The figure is a schematic diagram of the application scenario of the data detection method provided by the embodiment of the present application. As shown in Figure 1As shown, the application scenario includes a terminal device 110 and a server 120, and the terminal device 110 and the server 120 can communicate through a communication network. The communication network uses standard communication technology and / or protocol, and is usually the Internet, but can also be any network, including but not limited to Bluetooth, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), mobile, a private network, or any combination of a virtual private network. In some embodiments, custom or dedicated data communication technology can be used instead of or in addition to the above data communication technology.

[0056] The terminal device 110 is installed with a client for providing a type detection service for an application installed in the terminal device. The server 120 corresponds to the server of the client, and is used to provide the client with a type detection service for the application.

[0057] After starting the client, the application to be detected, i.e., the target application, can be selected, such as Figure 1 As shown, application 2 is selected as the target application, so that a screenshot is taken during the running of the target application to obtain a first to-be-detected image, and a scene identifier corresponding to the target application, which is used to identify the detection scene type corresponding to the target application. The terminal device 110 sends the first to-be-detected image and the scene identifier to the server 120. The first to-be-detected image is obtained by taking a screenshot during the running of the target application, and is a real process, which is less likely to be tampered with and is not easy to be attacked, so that subsequent detection based on the screenshot has high accuracy.

[0058] After the server 120 obtains the first to-be-detected image and the scene identifier, the first to-be-detected image is detected by a plurality of multi-type general classification models respectively, to obtain a first probability that the first to-be-detected image is of a target type output by each of the plurality of multi-type general classification models, so as to determine a first tentative result that whether the type of the first to-be-detected image is the target type according to the plurality of first probabilities. As shown, Figure 1 As shown, two multi-type general classification models are installed on the server 120, and two first probabilities are obtained respectively, so as to obtain the first tentative result based on the two first probabilities.

[0059] That is, instead of using a binary classification model for a certain abnormal behavior for detection, a multi-classification model, that is, a multi-type general classification model, is used for detection, so as to realize detection of multiple abnormal behaviors by one model. Moreover, in order to improve the detection accuracy of the multi-type general classification model, instead of using one model, multiple multi-type general classification models are used to detect the first to-be-detected image respectively, so as to improve the detection accuracy.

[0060] In addition, the server 120 will also determine the target scene supplement model applicable to the target application according to the scene identifier, from multiple scene supplement models. As shown in Figure 1 The server 120 selects the target scene supplement model corresponding to the scene identifier, that is, the target scene supplement model applicable to the target application, that is, the scene supplement model 2, from the three scene supplement models. Further, the first to-be-detected image is detected based on the target scene supplement model to obtain a second to-be-determined result. Different scene supplement models are applicable to different detection scene types, so that multiple scene supplement models are applicable to different detection scene types, thereby expanding the range of application scenarios.

[0061] Finally, the server 120 comprehensively determines the first target detection result of the first to-be-detected image, that is, whether the first to-be-detected image is of the target type, such as whether the first to-be-detected image includes abnormal behavior, by combining the first to-be-determined result with universality obtained based on multiple multi-type general classification models and the second to-be-determined result with pertinence to the corresponding scene obtained by the target scene supplement model, thereby further improving the detection accuracy.

[0062] The data detection method provided in the embodiments of the present application can be executed by a server. However, in other embodiments of the present application, a terminal device can also have similar functions as the server, so as to execute the data detection method provided in the embodiments of the present application, or the data detection method provided in the embodiments of the present application is executed by the terminal device and the server together, which is not limited in the present embodiment.

[0063] The data detection method provided in the present application can be applied to various scenes, including but not limited to game plug-ins, ticket grabbing software, social platform data fraud, office software batch processing documents, security testing software simulating hacker attacks, etc. The following takes three scenes as examples.

[0064] Scene one, detection scene for using game plug-ins to play games.

[0065] Game hacking is a way of cheating in games, which refers to a program or plug-in that modifies or enhances the game through unofficial and improper means. These hacks can often provide players with abilities or advantages that they would not normally have in the game, such as automatic aiming, unlimited lives, accelerated movement, and wall penetration, thereby disrupting the balance and fairness of the game.

[0066] As shown in Figure 2 To identify whether a player is using game hacks, a screenshot of the game interface during gameplay can be taken as the first image to be detected, and the data detection method provided by the present application can be used to identify the first image to be detected, thereby obtaining the first target detection result of the first image to be detected. If the first target detection result identifies the type of the first image to be detected as the target type, i.e., the abnormal type of using game hacks, the player's account can be processed, such as being banned, to ensure the fairness of the game and the legitimate rights and interests of the player.

[0067] Scenario two, detection scenario for using ticket snatching software to snatch tickets.

[0068] Ticket snatching software hacks can help users automatically monitor and obtain some platform-issued tickets, such as concert tickets and train tickets, through improper means. These software are usually considered as hacks or cheating tools because they obtain benefits through improper means and may infringe the fair competition rights of other users.

[0069] As shown in Figure 3 To help the platform detect whether a user has used a ticket snatching software hack, a screenshot of the ticket purchase process interface can be taken as the first image to be detected during the user's ticket purchase process through the platform, and the data detection method provided by the present application can be used to identify the first image to be detected, thereby obtaining the first target detection result of the first image to be detected. If the first target detection result identifies the type of the first image to be detected as the target type, i.e., the abnormal type of using ticket snatching software hacks, the user's account can be processed, such as invalidating the ticket purchase through the account, to ensure the fairness of the platform ticket purchase and the legitimate rights and interests of other users.

[0070] Scenario three, detection scenario for data falsification on social platforms.

[0071] On social platforms, some users may use data falsification hacks to automatically like, comment, or forward content, etc., to increase exposure or influence. However, this behavior can also be considered as an improper promotion method, which negatively affects the platform ecosystem.

[0072] As shown in Figure 4As shown, in order to help the social media platform to detect whether the user uses the data cheating plug-in, the behavior interface for generating data can be screenshot during the user generates data using the social media platform, and the screenshot is taken as the first to be detected image, and the first to be detected image is identified through the data detection method provided in the application, so as to obtain the first target detection result of the first to be detected image. If the first target detection result identifies that the type of the first to be detected image is the target type, that is, the abnormal type of using the data cheating plug-in, the account of the user can be processed, such as the ban processing, etc., to ensure the fairness of the social media platform data and the legitimate rights and interests of other users.

[0073] It should be noted that the above application scenarios are only examples, and the data detection method provided in the embodiment can also be applied to other scenarios, which are not limited here.

[0074] The data detection method provided in the application will be described in detail below by the method embodiment.

[0075] Referring to Figure 5 , the figure is a flowchart of the data detection method provided in the embodiment of the application. In order to facilitate description, the following embodiments are still described by taking the server as an example of the execution subject of the data detection method. As Figure 5 shown, the data detection method includes S501-S504.

[0076] S501: obtaining a first to be detected image and a scene identifier.

[0077] The first to be detected image is an image obtained by screenshot during the running of the target application. The first to be detected image is truly generated during the running of the target application. Based on the real first to be detected image, it can be accurately judged from the visual whether there is abnormal behavior in the game image, such as whether the game plug-in is used during the game running. Compared with the data in the target running process, the credibility of the to be detected image obtained by screenshot is high, so that the accuracy of subsequent detection based on the first to be detected image is improved, that is, the accuracy of the first target detection result is improved.

[0078] As a possible implementation, during the use of the target application configured on the client, if the server detects that there may be abnormal behavior based on the client's such as application log data, the server sends an image acquisition instruction to the client; the client responds to the image acquisition instruction, screenshots the application display screen of the target application to obtain the first to be detected image, and sends the first to be detected image to the server. Correspondingly, the server receives the first to be detected image sent by the client, and performs subsequent detection based on the first to be detected image.

[0079] As another possible implementation, during the use process of the target application configured on the client, the client initiates to take screenshots of the application display screen in the use process of the target application according to the set image interception rule, obtains a first to-be-detected image, and sends the first to-be-detected image to the server. Correspondingly, the server receives the first to-be-detected image sent by the client, and performs subsequent detection based on the first to-be-detected image.

[0080] The image interception rule is not specifically limited in the embodiments of the present application. For example, the interception is periodically performed according to a set time interval to obtain the first to-be-detected image. Taking the detection scenario of game plug-ins as an example, the target application can be a game application program, and the target application running process can refer to the period from when the game player starts the game to when the game ends. During the game process of the game player through the game application program configured on the client, the server obtains the game log data of the client, and detects whether there is suspected plug-in behavior based on the obtained game log data. For example, for a gun battle game, if it is found according to the game log data that even if the distance between the game character of the player and the game character of the opponent is very far, the distance is usually almost impossible to discover the position of the game character of the opponent by the visual of the game player, or even if the game character of the opponent is in a hidden place, it is almost impossible to discover the position of the game character of the opponent by the visual of the game player, but the game player can still accurately hit the game character of the opponent, at this time, it can be determined that the game player has suspected plug-in behavior, that is, there is a high possibility of plug-in behavior. If it is detected that the game player has suspected plug-in behavior, an image acquisition instruction is sent to the client, the client responds to the image acquisition instruction, takes a screenshot of the game screen during the game process to obtain a first to-be-detected image, and sends the first to-be-detected image to the server. Correspondingly, the server receives the first to-be-detected image sent by the client, and performs subsequent detection based on the first to-be-detected image.

[0081] The scene identifier is used to identify the detection scenario type corresponding to the target application. Different scene identifiers correspond to different detection scenario types, so that a scene supplement model suitable for the detection scenario type corresponding to the scene can be selected based on the scene identifier in the subsequent detection, so that the detection of the target application is more comprehensive and more accurate.

[0082] As a possible implementation, the number of types of abnormal behaviors in different target applications is different, the detection scenario types can be divided based on the number of types of abnormal behaviors, and therefore different scenes can be set based on different target applications. For example, different detection scenario types correspond to a smaller number of types of abnormal behaviors and a larger number of types of abnormal behaviors.

[0083] The embodiments of the present application do not specifically limit the acquisition manner of the scene identifier, which can be that the target application pre-acquires a corresponding detection scene type (equivalent to the detection permission acquired from the server), and sends the scene identifier corresponding to the detection scene type together each time the first to-be-detected image is sent. Alternatively, the target application can send the first to-be-detected image and the application identifier of the target application to the server, the server determines the detection service purchased by the target application based on the application identifier, and thus determines the scene identifier of the detection scene type corresponding to the detection service of the target application.

[0084] It can be understood that in the specific embodiments of the present application, all the data (such as the first to-be-detected image, the second to-be-detected image, the scene identifier, etc.) collected by the present application needs to obtain individual permission or individual consent of the object (such as a user, an institution or an enterprise) to which the data belongs when the embodiments of the present application are applied to specific products or technologies, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0085] S502: detecting the first to-be-detected image by using the plurality of multi-type general classification models respectively to obtain a plurality of first probabilities that the first to-be-detected image is of a target type output by each of the multi-type general classification models, and determining a first undetermined result of the first to-be-detected image according to the plurality of first probabilities.

[0086] In the embodiments of the present application, the multi-type general classification model can be a multi-classification model, which can recognize a plurality of abnormal behaviors, such as a plurality of abnormal behaviors in the target application, a plurality of abnormal behaviors in a plurality of applications, etc. Thus, compared with the related art which uses a binary classification model for detecting a certain abnormal behavior, the multi-classification model, i.e., the multi-type general classification model, is used for detection, which can realize detection of a plurality of abnormal behaviors by using one model and improve the detection range.

[0087] In the embodiments of the present application, the multi-type general classification model is trained based on images of various types in a plurality of applications. Taking the detection of whether there is a game plug-in in a game application as an example, the multi-type general classification model can be trained based on black images (images using various game plug-ins) and white images (images not using game plug-ins) in a plurality of games. The embodiments of the present application do not specifically limit the structure of the multi-type general classification model, which can be a MobileNetV3 model, or a shallow classification network designed based on a residual structure, etc.

[0088] As a possible implementation manner, if the target type is the type corresponding to the abnormal behavior, the network layer number of the multi-type general classification model can not be particularly deep, so as to reduce the calculation cost while filtering a large number of first to-be-detected images that do not include abnormal behaviors. In other words, the first to-be-detected image is identified by the plurality of multi-type general classification models, so as to filter out images that do not include abnormal behaviors from a large number of first to-be-detected images, because a large number of users will not use game cheats and the like to disturb the normal operation of the target application during the running process, so as to realize the function of coverage.

[0089] Moreover, the multi-type general classification model includes a plurality of multi-type general classification models, and the plurality of multi-type general classification models are different models, that is, the model structures or model parameters of different multi-type general classification models are different. For example, the model structures of the plurality of multi-type general classification models are the same, but the sample images used in the training process are different, so that the model parameters of the plurality of multi-type general classification models are different, and the detection effects on the same to-be-detected image are different. For example, some multi-type general classification models have good detection effects on A game, and some multi-type general classification models have good detection effects on B game. For another example, some multi-type general classification models have good detection effects on A type of abnormal behavior, and some multi-type general classification models have good detection effects on B type of abnormal behavior. That is, by ensuring the difference between the plurality of multi-type general classification models, it is beneficial to cover the detection of a plurality of abnormal behaviors, so as to improve the accuracy of the first target detection result.

[0090] After the server obtains the first to-be-detected image, the server calls the plurality of multi-type general classification models to identify the first to-be-detected image respectively, each multi-type general classification model outputs a probability that the first to-be-detected image is of the target type, that is, a first probability, so as to obtain a plurality of first probabilities based on the plurality of multi-type general classification models. The target type is a type to be identified. For example, if the identification is an abnormal behavior, the target type can be an abnormal behavior category, if the identification is a normal behavior, the target type can be a normal behavior category, and the like.

[0091] According to the plurality of first probabilities, a first pending result of the first to-be-detected image is determined. The first pending result is obtained based on the multi-type general classification model and is used to describe whether the first to-be-detected image is of the target type. The embodiments of the present application do not specifically limit the manner of obtaining the first pending result based on the plurality of first probabilities. For example, the average value of the plurality of first probabilities is determined, and if the average value of the plurality of first probabilities is greater than a first preset probability threshold, the first pending result indicates that the first to-be-detected image is of the target type. For another example, if the number of the plurality of first probabilities greater than a second preset probability threshold is greater than a first preset number threshold, the first pending result indicates that the first to-be-detected image is of the target type.

[0092] S503: determining a target scene supplement model suitable for the target application from the multiple scene supplement models according to the scene identifier, and detecting the first to-be-detected image by using the target scene supplement model to obtain a second to-be-detected result.

[0093] Although the multi-type general classification model can recognize multiple types, such as multiple abnormal behaviors, that is, the multi-type general classification model is more universal, due to the data amount of training samples and model parameters, the detection accuracy of a model suitable for a certain detection scene type is generally higher than that of a model suitable for multiple detection scene types.

[0094] Therefore, in order to improve the detection accuracy of the detection scene type suitable for the application model, in the embodiments of the present application, multiple scene supplement models are provided, and different scene supplement models are suitable for different detection scene types, so as to cover more detection scene types by using multiple scene supplement models, thereby expanding the range of application scenarios.

[0095] The detection scene type is related to the target application, so the scene supplement model suitable for the target application is determined from the multiple scene supplement models based on the scene identifier suitable for the target application, that is, the target scene supplement model, so as to detect the first to-be-detected image based on the target scene supplement model to obtain a second to-be-detected result. The second to-be-detected result is obtained by using the target scene supplement model and is used to describe whether the first to-be-detected image is of the target type.

[0096] The embodiments of the present application do not specifically limit the scene supplement model, which can be set according to actual application by those skilled in the art, and will be described later, and will not be repeated here.

[0097] It should be noted that the embodiments of the present application do not specifically limit the order of using the target scene supplement model and the multi-type general classification model, such as the order of S502 and S503, which can be executed simultaneously or sequentially, and can be set according to actual needs by those skilled in the art. The same applies later, and will not be repeated here.

[0098] S504: determining a first target detection result of the first to-be-detected image according to the first to-be-detected result and the second to-be-detected result.

[0099] Referring to Figure 6 The figure is a schematic diagram of a model used in a data detection method provided by the embodiments of the present application. In the Figure 6In the specific implementation, the first target detection result is determined based on the first tentative result and the second tentative result. The first target detection result is used to identify whether the type of the first detection image is a target type. Compared with the first tentative result or the second tentative result, the accuracy of the first target detection result determined based on the first tentative result and the second tentative result is higher.

[0100] The first target detection result is determined based on the first tentative result and the second tentative result, and is used to identify whether the type of the first detection image is a target type. Compared with the first tentative result or the second tentative result, the accuracy of the first target detection result determined based on the first tentative result and the second tentative result is higher.

[0101] The first tentative result obtained based on the plurality of multi-type general classification models has universality, and the second tentative result obtained based on the target scene supplement model is specific to the corresponding scene. The first target detection result of the first detection image, i.e., whether the first detection image is of a target type, is determined based on the first tentative result and the second tentative result, thereby further improving the accuracy of detection.

[0102] The embodiments of the present application do not specifically limit the manner in which the first target detection result is determined based on the first tentative result and the second tentative result. For example, if the first tentative result identifies that the first detection image is of a target type, or the second tentative result identifies that the first detection image is of a target type, the first target detection result identifies that the first detection image is of a target type. That is, the first target detection result is obtained based on the combination of the first tentative result and the second tentative result, so as to ensure that as many images of a target type as possible are not missed, thereby ensuring coverage.

[0103] As a possible implementation, the first target detection result identifies that the first detection image is of a target type, indicating that the probability that the first detection image is of a target type is high, so that the first detection image determined to be of a target type can be sent to artificial review, so as to further improve the accuracy of detection. It can be understood that, although the first detection image is identified in combination with artificial, since many images that are not of a target type can be filtered out by the multi-type general classification model and the target scene supplement model, the artificial review is only a small part, such as one hundred or two hundred, so that the artificial review amount is low while the accuracy of detection is improved.

[0104] The above technical solution can be seen from the following aspects. First, instead of using application log data for identification, a screenshot is taken during the running of the target application, and the obtained image is used as the first detection image. Since the first detection image is a real process, it is less likely to be tampered with and is not easy to be attacked, thereby improving the accuracy of subsequent detection. Second, instead of using a binary classification model for a certain abnormal behavior for detection, a multi-classification model, i.e., a multi-type general classification model, is used for detection, thereby realizing detection of multiple abnormal behaviors by one model. Third, to improve the detection accuracy of the multi-type general classification model, instead of using one model, multiple multi-type general classification models are used to detect the first detection image, and the first probability that the first detection image is of a target type output by the multiple multi-type general classification models is obtained, thereby determining whether the type of the first detection image is the target type based on the multiple first probabilities, improving the accuracy of the first tentative result. Fourth, to expand the range of application scenarios, multiple scene supplement models are pre-trained, different scene supplement models are suitable for different detection scene types, and a target scene supplement model suitable for a detection scene type corresponding to a scene identifier is obtained from the multiple scene supplement models based on the obtained scene identifier, and then the first detection image is detected based on the target scene supplement model to obtain a second tentative result. Fifth, the first tentative result with universality obtained based on the multiple multi-type general classification models and the second tentative result with pertinence for the corresponding scene obtained based on the target scene supplement model are comprehensively determined to obtain the first target detection result of the first detection image, i.e., whether the first detection image is of the target type, such as whether the first detection image includes abnormal behavior, thereby further improving the accuracy of detection.

[0105] In the embodiments of the present application, multiple scene supplement models are included, such as a local map general classification model for a small amount of cheating, an easily recognizable general classification model for a large amount of cheating, and the like. Different scene supplement models are used to detect different detection scene types.

[0106] Scene one: cold start scene.

[0107] It can be understood that the target type can be a large category, which can include multiple behavior categories. Taking the detection of game plug-ins as an example, the target type can be a type of using game plug-ins (abnormal behavior), and the abnormal behavior category can include multiple behavior categories, such as seeing an opponent game character that should not be seen, increasing the blood volume of the game character, and the like.

[0108] If the scene identifier is currently a cold start scene, such as an application is just online, there are few external plug-ins for the application, and thus the amount of cheating is small. That is, the cold start scene has a small number of types of cheating behaviors implemented by game external plug-ins (referred to as a small amount of cheating), such as the detection scene corresponding to the long tail game. At this time, the target scene supplement model selected in S503 is the local graph general classification model.

[0109] Referring to Figure 7 , this figure is a schematic diagram of detection based on a local graph general classification model provided by an embodiment of the present application. In Figure 7 , two (i.e., n is 2) multi-type general classification models and one local graph general classification model are included. The first to-be-detected image is detected based on the two multi-type general classification models, two first probabilities are obtained, a first tentative result is obtained based on the two first probabilities, the first to-be-detected image is recognized by the local graph general classification model, a second tentative result is obtained, and a first target detection result is obtained based on the first tentative result and the second tentative result.

[0110] In the embodiment of the present application, the local graph general classification model is trained based on local image samples. In order to facilitate understanding of the local image samples, the following takes a gun battle game as an example to explain why local image samples are needed.

[0111] In a gun battle game, a game external plug-in generally helps a game player to quickly find an opponent game character. Referring to Figure 8 , this figure is a schematic diagram of marking an opponent game character by a square in an embodiment of the present application. In Figure 8 , the opponent game character is blocked by grass, and theoretically the game player cannot see the opponent game character, but through the game external plug-in, it marks the position of the opponent game character by a square 801, so that the home game character can shoot the opponent game character. Referring to Figure 9 , this figure is a schematic diagram of marking an opponent game character by a skeleton in an embodiment of the present application. In Figure 9 , the opponent game character is blocked by grass, and theoretically the home game character cannot see the opponent game character, but through the game external plug-in, it marks the position of the opponent game character by a skeleton 901, so that the home game character can shoot the opponent game character. In addition, the game external plug-in can also help the game character to modify attribute parameters, such as the home game character will not lose blood after being hit, that is, the blood volume remains unchanged, as shown by 1001 in Figure 10 .

[0112] The box, the skeleton and the blood amount are all regions in the first to-be-detected image that are relatively easy to identify whether a game plug-in is used, and in order to avoid the influence of other regions in the first to-be-detected image on the regions where the box, the skeleton and the blood amount are located, the regions where the box, the skeleton and the blood amount are located can be extracted from the first to-be-detected image as local image samples, so as to improve the learning ability of the subsequent local general classification model and improve the detection accuracy of the local general classification model.

[0113] Based on this, the training process of the local general classification model is described below, referring to A1-A4.

[0114] A1: Obtain a plurality of image samples.

[0115] The image samples have labels, which are used to identify the types of the image samples, such as belonging to a target type or not belonging to a target type, or can be further divided into belonging to a target type (an abnormal behavior category with a box), belonging to a target type (an abnormal behavior category with a skeleton), etc.

[0116] Moreover, the types of the plurality of obtained image samples are various, so that a local general classification model capable of identifying various types is trained through various image samples, such as image samples with a box, image samples without a box, image samples with a skeleton, image samples without a skeleton, image samples with blood amount and image samples without blood amount, and the local general classification model is obtained based on the above-mentioned six kinds of image samples.

[0117] A2: Each image sample is cropped according to the label of the image sample to obtain a plurality of local image samples.

[0118] Different labels can correspond to different cropping regions. For example, if the label identification is based on the appearance of a box in the image sample to be identified as a target type, the cropping region is the region including the box. For example, if the label identification is based on the appearance of a skeleton in the image sample to be identified as a target type, the cropping region is the region including the skeleton, etc. For example, the size of the image sample is 1024x768, and the size of the local image sample can be 400x300.

[0119] A3: Each local image sample is detected by the initial local general classification model to obtain a first detection result of each image sample.

[0120] For example, the local image sample is input into the initial local general classification model, and the initial local general classification model is detected to obtain the first detection result of the local image sample. Therefore, the plurality of local image samples are respectively input into the initial local general classification model, and the initial local general classification model is respectively detected to obtain each local image sample, i.e., the first detection result corresponding to each image sample.

[0121] Compared with the local graph general classification model, the initial local graph general classification model is a model that has not been trained, and after being trained, a local graph general classification model is obtained. The initial local graph general classification model is not specifically limited in the embodiments of the present application, and can be a middle-layer classification network designed based on a residual structure, etc., wherein the network depth of the middle-layer classification network is greater than that of a shallow-layer classification network.

[0122] A4: Adjusting the model parameters of the initial local graph general classification model according to the difference between each first detection result and the label of the corresponding image sample, to obtain a local graph general classification model.

[0123] For example, for the difference between a first detection result and the label of the image sample corresponding to the first detection result, the model parameters of the initial local graph general classification model are adjusted so that the difference between the first detection result and the label of the image sample corresponding to the first detection result becomes smaller and smaller, and when a preset condition is met (such as the number of iterations or model convergence), a local graph general classification model is obtained. The local graph general classification model can detect the types of multiple image samples.

[0124] In addition, it should be noted that although the local graph general classification model is trained based on local image samples, the local graph general classification model can also be trained to recognize images larger than the local image samples. Alternatively, the first detection image can be cropped to obtain multiple local detection images, and the multiple local detection images are input into the local graph general classification model. If one of the local detection images is recognized as a target type, the first detection image is of the target type.

[0125] Therefore, by cropping the image samples based on the labels, local image samples with pertinence can be obtained, and the detection accuracy of the local graph general classification model trained based on the local image samples with less redundant information is higher, so that in the subsequent detection process for the cold start scene, the first tentative result is obtained by global detection based on the multi-type general classification model, the second tentative result is obtained by local detection based on the local graph general classification model, and the first target detection result is obtained based on the first tentative result and the second tentative result. That is, for a scene with less cheating, the accuracy of the first target detection result is improved by combining global detection and local detection (local detection can improve the coverage rate of detection by about 10%).

[0126] As a possible implementation manner, for a cold start scene, the cheating amount is small, that is, there are a large number of to-be-detected images that do not include the target type, in order to ensure the accuracy of detection, as much as possible, to-be-detected images that do not include the target type can be filtered out, but for to-be-detected images that may be the target type, all need to be identified, so the first target detection result can be obtained based on the union of the first pending result and the second pending result.

[0127] Specifically, if the first pending result identifies the first to-be-detected image as the target type, or the second pending result identifies the first to-be-detected image as the target type, it is determined that the first target detection result of the first to-be-detected image is used to identify that the first to-be-detected image is the target type. Thus, as much as possible, to-be-detected images belonging to the target type are not missed to ensure coverage. Similarly, if the first pending result identifies that the first to-be-detected image is not the target type, and the second pending result identifies that the first to-be-detected image is not the target type, it is determined that the first target detection result of the first to-be-detected image is used to identify that the first to-be-detected image is not the target type.

[0128] As a possible implementation manner, taking a long tail game as an example, the first to-be-detected image is a cheating level relatively low volume version scene, that is, there are a large number of to-be-detected images that are not the target type (referred to as normal images) in many applications, so a large number of normal images can be filtered out by the server, and the to-be-detected images belonging to the target type (referred to as abnormal images) detected by the server have high accuracy. In order to further improve the accuracy of detection, the abnormal images can be further manually audited. It can be understood that, since the server has filtered out a large number of normal images, the number of manual audits (the audit level is about a few hundred) is low, thereby ensuring the accuracy of detection while reducing the amount of manual audit.

[0129] As a possible implementation manner, the multi-type general classification model is a model suitable for multiple applications, and the multi-type general classification model can be fine-tuned again based on the review result, so that the fine-tuned multi-type special classification model is more suitable for the target application. The following will be specifically explained, see B1-B5.

[0130] B1: Obtain a review result for the first to-be-detected image.

[0131] The accuracy of the review result is greater than the accuracy threshold, that is, the accuracy of the review result is high, so that the multi-type general classification model is fine-tuned based on the review result with high accuracy, and the accuracy of the model obtained by fine-tuning, that is, the multi-type special classification model, is higher.

[0132] As a possible implementation manner, the review result can be obtained based on manual auditing. For example, after a first to-be-detected image belonging to the target type is manually audited for one or two weeks, a certain number of review results can be accumulated, the first to-be-detected image belonging to the target type is detected based on the server, the manual auditing result of the first to-be-detected image belonging to the target type is determined based on the manual auditing, the manual auditing result is taken as the review result, and the multi-type general classification model is fine-tuned based on the review result, so as to realize a closed loop of data, ensure accuracy, reduce the amount of manual auditing, and improve the accuracy of the model.

[0133] B2: If the review result is different from the first target detection result, the first to-be-detected image is detected by the target multi-type general classification model to obtain a second probability that the first to-be-detected image is of the target type.

[0134] If the review result is different from the first target detection result, it indicates that the first target detection result is incorrect, that is, the detection capability of the multi-type general classification model is insufficient, and therefore the multi-type general classification model can be fine-tuned based on the review result.

[0135] As a possible implementation manner, all multi-type general classification models can be fine-tuned, or part of the multi-type general classification models can be fine-tuned, which is not limited in the present application, and one of the multi-type general classification models, that is, the target multi-type general classification model, is taken as an example for description below.

[0136] For example, if the first target detection result of the first to-be-detected image is different from the review result of the first to-be-detected image, the first to-be-detected image can be input into the target multi-type general classification model, and the target multi-type general classification model is used for detection to obtain a probability that the first to-be-detected image is of the target type, that is, a second probability.

[0137] B3: A second detection result of the first to-be-detected image is determined according to the second probability.

[0138] For example, if the second probability is greater than a third preset probability threshold, it can be determined that the second detection result indicates that the first to-be-detected image belongs to the target type. For another example, if the second probability is less than or equal to the third preset probability threshold, it can be determined that the second detection result indicates that the first to-be-detected image does not belong to the target type.

[0139] B4: A multi-type special classification model is obtained by adjusting model parameters of the target multi-type general classification model according to a difference between the second detection result and the review result.

[0140] According to the difference between the second detection result and the review result, the model parameters of the target multi-type general classification model are adjusted, so that the difference between the second detection result and the review result becomes smaller and smaller, so as to obtain the trained target multi-type general classification model, that is, the multi-type special classification model.

[0141] As a possible implementation, the plurality of multi-type general classification models can be respectively taken as the target multi-type general classification model, and the multi-type special classification model corresponding to each multi-type general classification model is obtained through B2-B4.

[0142] B5: replacing the multi-type general classification model with the multi-type special classification model.

[0143] The multi-type general classification model is replaced with the multi-type special classification model, so that the detection of the to-be-detected image based on the multi-type general classification model is performed in subsequent identification, and the detection accuracy is improved.

[0144] Therefore, the multi-type special classification model is obtained by fine-tuning the multi-type general classification model based on the review result, that is, the multi-type special classification model is fine-tuned on the multi-type general classification model based on the to-be-detected image of the target application. Compared with the multi-type general classification model suitable for multiple applications, the multi-type special classification model is more suitable for the target application, so that the detection of the to-be-detected image from the target application is more accurate.

[0145] It should be noted that in the embodiments of the present application, the multi-type general classification model is not fine-tuned based on the review result, because the review result only identifies whether the to-be-detected image is of the target type. However, if the multi-type general classification model is to be fine-tuned, not only does the review result need to identify whether the to-be-detected image is of the target type, but also needs to identify which abnormal behavior is hit in the to-be-detected image, that is, it is difficult to obtain data for fine-tuning the multi-type general classification model.

[0146] Based on this, the multi-type general classification model is not fine-tuned in the embodiments of the present application, but only the multi-type general classification model is fine-tuned based on the easily obtained review result. How to use the multi-type special classification model is described below, see C1-C4.

[0147] C1: taking a plurality of multi-type general classification models as target multi-type general classification models respectively to obtain multi-type special classification models corresponding to each multi-type general classification model.

[0148] All multi-type general classification models are fine-tuned based on the review result, so that each multi-type general classification model

[0149] C2: obtaining a second to-be-detected image.

[0150] The second to-be-detected image is an image obtained by taking a screenshot during running of the target application. The second to-be-detected image can be the same as or different from the first to-be-detected image, and the present application does not make a specific limitation thereon. However, it can be understood that in actual application, the second to-be-detected image is more likely to be different from the first to-be-detected image.

[0151] It can be understood that not only the second to-be-detected image can be obtained, but also a scene identifier can be obtained, which is used to indicate that the target scene supplement model is a local map general classification model. However, in the process of using the multi-type special classification model, since the multi-type special classification model has high accuracy, the local map general classification model can not be used, so whether the scene identifier is obtained does not need to be specially concerned.

[0152] C3: detecting the second to-be-detected image by using the plurality of multi-type special classification models respectively to obtain a third probability that the second to-be-detected image output by each multi-type special classification model is of the target type.

[0153] The second to-be-detected image is detected by using the multi-type special classification model to obtain a probability that the second to-be-detected image output by the multi-type special classification model is of the target type, that is, the third probability. Therefore, the second to-be-detected image is detected by using the plurality of multi-type special classification models respectively to obtain a third probability that the second to-be-detected image output by each multi-type special classification model is of the target type.

[0154] C4: determining a second target detection result of the second to-be-detected image according to the plurality of third probabilities.

[0155] The second target detection result is used to identify whether the second to-be-detected image is of the target type. As a possible implementation manner, if the second target detection result is used to identify that the second to-be-detected image is of the target type, the second to-be-detected image can be detected based on a manual review manner to improve the accuracy of detection.

[0156] The present application does not specifically limit the manner of determining the second target detection result. For example, an average value of the plurality of third probabilities is determined, and if the average value of the plurality of third probabilities is greater than a fourth preset probability threshold, the second target detection result identifies that the second to-be-detected image is of the target type. For another example, if the number of third probabilities greater than a fifth preset probability threshold is greater than a second preset number threshold, the second target detection result identifies that the second to-be-detected image is of the target type, and the like.

[0157] Referring to Figure 11 , the figure is a schematic diagram of detection based on a multi-type special classification model provided by an embodiment of the present application. In Figure 11In the second embodiment, the second to-be-detected image is respectively input into two multi-type special classification models to obtain two third probabilities, and the second target detection result is obtained based on the two third probabilities.

[0158] Therefore, the multi-type special classification model is obtained by fine-tuning the multi-type general classification model, and the multi-type general classification models that have been used are replaced respectively, and the multi-type special classification model is used for detection. The multi-type special classification model has high accuracy for the to-be-detected image from the target application, so it is not necessary to fine-tune the local map general classification model, the accuracy is improved, and the fine-tuning difficulty is reduced.

[0159] As a possible implementation manner, although the local map general classification model is not fine-tuned, the detection can be combined with the local map general classification model to improve the coverage rate of the detection. Specifically, the second to-be-detected image is detected by the plurality of multi-type special classification models to obtain a plurality of third probabilities, and the third to-be-determined result of the second to-be-detected image is determined based on the plurality of third probabilities. The second to-be-detected image is detected according to the local map general classification model to obtain the fourth to-be-determined result of the second to-be-detected image. Finally, the second target detection result of the second to-be-detected image is determined according to the third to-be-determined result and the fourth to-be-determined result.

[0160] The manner of obtaining the second target detection result based on the third to-be-determined result and the fourth to-be-determined result can refer to the manner of obtaining the first intermediate to-be-determined result based on the first to-be-determined result and the second to-be-determined result, which will not be described herein again.

[0161] Referring to Figure 12 , the figure is a schematic diagram of detection based on a multi-type special classification model provided by an embodiment of the present application. In Figure 12 , the second to-be-detected image is respectively input into two multi-type special classification models to obtain two third probabilities, and the third to-be-determined result is obtained based on the two third probabilities. The second to-be-detected image is input into the local map general classification model to obtain the fourth to-be-determined result, and the second target detection result is obtained based on the third to-be-determined result and the fourth to-be-determined result.

[0162] In addition, when the model of the embodiment shown in Figure 11 or Figure 12 is used for detection, if the second to-be-detected image is detected as a target type, manual review can also be performed to further improve the detection accuracy. Since the accuracy of the fine-tuned multi-type special classification model is high, the amount of manual review will be reduced, for example, the amount of review of more than one hundred will be reduced to within one hundred, the amount of review of more than two hundred will be reduced to within two hundred, and the number of detected cheating people will be more.

[0163] Therefore, in the subsequent confrontation operation process, only the review results obtained through manual review need to be accumulated, and the multi-type general classification model or the multi-type special classification model can be continuously fine-tuned. After the fine-tuned model is evaluated to be qualified, the model can be updated and put online, so as to continuously improve the accuracy of the fine-tuned model and improve the effect of the confrontation operation.

[0164] Scenario two: real-time processing scenario.

[0165] If the scene identifier represents that the detection scene type corresponding to the target application is a real-time processing scenario, in this scenario, it is necessary to quickly confront the plug-in and process the account using the plug-in in the application in real time, so it is necessary to detect more accurately, so that the processed account will not be misjudged and will not affect the user's experience of using the target application. At this time, the target scene supplement model selected in S503 is the easily recognizable general classification model and the text recognition model.

[0166] The following describes the detection of the first to-be-detected image based on the easily recognizable general classification model and the text recognition model to obtain a second to-be-determined result. For details, see D1-D4.

[0167] D1: Detect the first to-be-detected image through the easily recognizable general classification model to obtain a first intermediate to-be-determined result.

[0168] The easily recognizable general classification model is used to identify a first behavior category in the plurality of behavior categories included in the target type, and the identification difficulty of the first behavior category is less than a difficulty threshold. That is, the easily recognizable general classification model is used to detect a behavior category with relatively small identification difficulty, so as to improve the accuracy of detection.

[0169] The target type can be a normal type corresponding to a normal behavior and an abnormal type corresponding to an abnormal behavior. As a possible implementation manner, the target type is the abnormal type, and the abnormal type can include behavior categories respectively corresponding to a plurality of abnormal behaviors, that is, one abnormal behavior corresponds to one behavior category.

[0170] Taking a gun battle game as the target application as an example, some game plug-ins can be displayed in the form of a function directory on the game interface, and the function directory is easy to recognize.

[0171] Referring to Figure 13 , the figure is a schematic diagram of a game plug-in including a function directory provided by an embodiment of the present application. As Figure 13 shown, a plug-in function directory box 1301 is displayed in the upper left corner of the game screen, including three plug-in mode options of a square box, a skeleton, and a blood volume. When the game player selects the square box plug-in mode, the position of the opponent game character will be automatically marked with a square box, as Figure 8 shown, to prompt the home game player. Figure 13The text information "external function list, box, skeleton, blood volume" shown is text associated with abnormal behavior that appears extra in the game interface due to the existence of a game external plug-in. That is, if the function list is identified, the account has a high probability of using a game external plug-in, so the first behavior category can be determined as the behavior of using the function list, and the easy-to-identify general classification model is used to identify the function list.

[0172] The training method of the easy-to-identify general classification model is not limited in the embodiments of the present application. Taking the foregoing as an example, the easy-to-identify general classification model can be obtained based on image samples including the function list and image samples not including the function list. As a possible implementation manner, the easy-to-identify general classification model can be a deep classification network designed based on a residual structure, and the network depth of the deep classification network is greater than that of the middle-level classification network. As a possible implementation manner, a plurality of easy-to-identify general classification models can also be designed, such as an easy-to-identify general classification model with an input image size of 800x600 and an easy-to-identify general classification model with an input image size of 1024x768.

[0173] D2: If the first intermediate tentative result identifies the first detected image as the first behavior category in the target type, performing text recognition on the first detected image through a text recognition model to obtain a first text recognition result.

[0174] Since the identification difficulty of the first behavior category is low, the probability of misidentifying the first behavior category is low, so if the first intermediate tentative result identifies the first detected image as the first behavior category in the target type, it means that the first detected image is probably the first behavior category and belongs to the target type.

[0175] However, in order to avoid detection errors and misjudgments, when the first detected image is identified as the first behavior category based on the easy-to-identify general classification model, the text recognition model can be used for detection again. The text recognition model can perform text recognition on the first detected image to obtain a first text recognition result, and the first text recognition result can identify the text included in the first detected image.

[0176] The embodiments of the present application do not limit the text recognition model, and a person skilled in the art can set it according to actual needs. For example, the network structure of the text recognition model can be PaddleOCR (Paddle Optical Character Recognition), where OCR is Optical Character Recognition. For another example, based on the text screenshot label in the game (such as data volume in thousands to tens of thousands), combined with the OCR detection algorithm, the text recognition model is obtained by using the open source pre-training weight to fine-tune the text detection frame and the text recognition model.

[0177] D3: If the first text recognition result is the same as the keyword in the keyword set, it is determined that the second pending result is used to identify that the first to-be-detected image is the target type.

[0178] The keyword set includes keywords related to the first behavior category, or in other words, a plurality of keywords are determined in advance based on the text related to the first behavior category, and the plurality of keywords constitute the keyword set.

[0179] Therefore, if the first text recognition result is the same as the keyword in the keyword set, it indicates that the text recognition model also detects that the first to-be-detected image has text related to the first behavior category, and the first to-be-detected image belongs to the first behavior category and belongs to the target type. At this time, both the text recognition model and the easy-to-recognize general classification model determine that the first to-be-detected image is the target type, and it is determined that the second pending result is used to identify that the first to-be-detected image is the target type.

[0180] D4: If the first text recognition result is not the same as the keyword in the keyword set, it is determined that the second pending result is used to identify that the first to-be-detected image is not the target type.

[0181] If the first text recognition result is not the same as the keyword in the keyword set, at this time, the easy-to-recognize general classification model determines that the first to-be-detected image is the target type, and the text recognition model determines that the first to-be-detected image is not the target type. It may be that the text recognition model or the easy-to-recognize general classification model fails to recognize, at this time, it is not possible to accurately determine whether the first to-be-detected image is the target type, in order to avoid subsequent misjudgment, the second pending result can be used to identify that the first to-be-detected image is not the target type.

[0182] Referring to Figure 14 , the figure is a schematic diagram of detection based on an easy-to-recognize general classification model and a text recognition model provided by an embodiment of the present application. In Figure 14 , the target scene supplement model is an easy-to-recognize general classification model and a text recognition model, and includes two multi-type general classification models.

[0183] The first to-be-detected image is detected by the two multi-type general classification models to obtain two first probabilities, and the first to-be-detected image is obtained based on the two first probabilities. The first pending result identifies whether the first to-be-detected image is the target type.

[0184] The first intermediate undetermined result is obtained by identifying the first to-be-detected image through the easily recognizable general classification model. If the first intermediate undetermined result identifies the first to-be-detected image as the first behavior category, i.e., belonging to the target type, then the first to-be-detected image is subjected to character recognition through a character recognition model to obtain a first character recognition result. If the first character recognition result is the same as a keyword in the keyword set, then it is determined that the second undetermined result is used to identify the first to-be-detected image as the target type. In this case, the account corresponding to the first to-be-detected image can be directly processed, such as being banned. If the first character recognition result is not the same as the keyword in the keyword set, then it is determined that the second undetermined result is used to identify the first to-be-detected image as not being of the target type. Then, the first target detection result can be obtained based on a comprehensive determination of the first undetermined result.

[0185] For example, the first undetermined result and the second undetermined result are subjected to deduplication processing to ensure that the first to-be-detected image is not manually reviewed twice, improve the efficiency of manual review, and reduce the workload of manual review.

[0186] Therefore, in the process of detecting in real-time processing scenarios, since detection through the model can directly process the corresponding account, to ensure user experience, the accuracy of detection needs to be improved to prevent misprocessing. Based on this, the easily recognizable general classification model for detecting and identifying the first behavior category with low difficulty can be pre-trained. When the first to-be-detected image is of the target type based on the combined detection of the easily recognizable general classification model and the character recognition model, the corresponding account is directly processed. Through twice detection and both detection results being that the first to-be-detected image is of the target type, the accuracy is high, and the amount of misjudgment is reduced. Moreover, the magnitude of the first behavior category is low, which can reduce the computational load of the character recognition model. The rest of the detection process can refer to the process shown in Figure 7 That is, for real-time processing scenarios, the easily recognizable general classification model and the character recognition model can accurately identify to-be-detected images belonging to the target type, thereby quickly processing and achieving fast anti-hacking.

[0187] As a possible implementation manner, the multi-type general classification model and the easily recognizable general classification model can also be fine-tuned to improve the detection accuracy of the model. For details, see E1-E4.

[0188] E1: Obtain a review result for the first to-be-detected image.

[0189] The accuracy of the review result is greater than an accuracy threshold. For details, see the foregoing B1.

[0190] E2: If the review result is different from the first target detection result, the first to-be-detected image is detected by the target multi-type general classification model to obtain a fourth probability that the first to-be-detected image is of the target type, and the first to-be-detected image is detected by the easily recognizable general classification model to obtain a fifth probability that the first to-be-detected image is of the target type.

[0191] If the review result is different from the first target detection result, it indicates that the first target detection result is incorrect, that is, the detection capability of the multi-type general classification model is insufficient, and therefore the multi-type general classification model and the easily recognizable general classification model can be fine-tuned based on the review result.

[0192] First, the multi-type general classification model is fine-tuned. As a possible implementation manner, all multi-type general classification models can be fine-tuned, or part of the multi-type general classification models can be fine-tuned, which is not limited in the present application. Hereinafter, one of the multi-type general classification models, that is, the target multi-type general classification model, is taken as an example for description.

[0193] For example, if the first target detection result of the first to-be-detected image is different from the review result for the first to-be-detected image, the first to-be-detected image can be input into the target multi-type general classification model, and the target multi-type general classification model is used for detection to obtain a probability that the first to-be-detected image is of the target type, that is, the fourth probability.

[0194] Then, the easily recognizable general classification model is fine-tuned. The first to-be-detected image is detected by the easily recognizable general classification model to obtain a probability that the first to-be-detected image is of the target type, that is, the fifth probability.

[0195] E3: According to the fourth probability, a third detection result, that is, a second target detection result, of the first to-be-detected image is determined, and according to the difference between the third detection result, that is, the second target detection result, and the review result, the model parameters of the target multi-type general classification model are adjusted to obtain a multi-type special classification model, and the multi-type special classification model is used to replace the multi-type general classification model.

[0196] The relevant parts are described in the foregoing B3-B5, which will not be described here.

[0197] E4: According to the fifth probability, a fourth detection result of the first to-be-detected image is determined, and according to the difference between the fourth detection result and the review result, the model parameters of the easily recognizable general classification model are adjusted to obtain an easily recognizable special classification model, and the easily recognizable special classification model is used to replace the easily recognizable general classification model.

[0198] For example, if the fifth probability is greater than a sixth preset probability threshold, it can be determined that the fourth detection result indicates that the first to-be-detected image belongs to the target type. For another example, if the fifth probability is less than or equal to the sixth preset probability threshold, it can be determined that the fourth detection result indicates that the first to-be-detected image does not belong to the target type.

[0199] According to the difference between the fourth detection result and the review result, the model parameters of the target multi-type general classification model are adjusted, so that the difference between the fourth detection result and the review result becomes smaller and smaller, so as to obtain a trained easily recognizable general classification model, that is, an easily recognizable special classification model.

[0200] The easily recognizable special classification model is used to replace the easily recognizable general classification model, so as to detect the to-be-detected image based on the easily recognizable general classification model in subsequent recognition, and improve the detection accuracy. It should be noted that the fine-tuning can be based on the multi-type general classification model at the beginning or the easily recognizable general classification model, or based on the multi-type general classification model or the easily recognizable general classification model obtained by the last fine-tuning. The embodiments of the present application do not make specific limitations.

[0201] Therefore, by fine-tuning the multi-type general classification model and the easily recognizable general classification model based on the review result, the multi-type special classification model and the easily recognizable special classification model are obtained, that is, the multi-type special classification model is fine-tuned based on the to-be-detected image of the target application on the basis of the multi-type general classification model, and the easily recognizable special classification model is fine-tuned based on the to-be-detected image of the target application on the basis of the easily recognizable general classification model. Compared with the multi-type general classification model and the easily recognizable general classification model suitable for multiple applications, the multi-type special classification model and the easily recognizable special classification model are more suitable for the target application, so that the detection of the to-be-detected image from the target application is more accurate.

[0202] As a possible implementation manner, after the multi-type special classification model and the easily recognizable special classification model are obtained by fine-tuning, the multi-type special classification model is used to replace the multi-type general classification model, and the easily recognizable special classification model is used to replace the easily recognizable general classification model, so that the multi-type special classification model and the easily recognizable special classification model more suitable for the target application are used to detect the to-be-detected image, thereby further improving the detection accuracy. For details, see F1-F7.

[0203] F1: Obtain a third to-be-detected image.

[0204] The third to-be-detected image is an image obtained by taking a screenshot during the running of the target application, which can be the same as the second to-be-detected image or different from the second to-be-detected image. The present application does not make specific limitations.

[0205] F2: detecting the third to-be-detected image by the plurality of multi-type special classification models respectively to obtain a sixth probability that the third to-be-detected image is of the target type output by each multi-type special classification model.

[0206] For example, the third to-be-detected image is input into the multi-type special classification model, and detection is performed by the multi-type special classification model, so as to obtain a probability output by the multi-type special classification model, i.e., the sixth probability for describing that the third to-be-detected image is of the target type.

[0207] F3: determining a second intermediate undetermined result of the third to-be-detected image according to the plurality of sixth probabilities.

[0208] For example, an average value of the plurality of sixth probabilities is determined, and if the average value of the plurality of sixth probabilities is greater than a seventh preset probability threshold, the second intermediate undetermined result identifies that the third to-be-detected image is of the target type. For another example, if the number of the plurality of sixth probabilities greater than the seventh preset probability threshold is greater than a third preset number threshold, the second intermediate undetermined result identifies that the third to-be-detected image is of the target type.

[0209] As a possible implementation manner, if the average value of the plurality of sixth probabilities is greater than or equal to a first probability threshold, it is determined that the second intermediate undetermined result is used to identify that the third to-be-detected image is of the target type, and F4-F7 can be executed subsequently.

[0210] If the average value of the plurality of sixth probabilities is less than the first probability threshold and the average value of the plurality of sixth probabilities is greater than or equal to a second probability threshold, it is determined that the second intermediate undetermined result is used to identify that the third to-be-detected image is of a to-be-verified type, wherein the second probability threshold is less than the first probability threshold. For example, the to-be-detected image of the to-be-verified type can continue to be detected based on a manual review manner, so as to improve detection accuracy.

[0211] If the average value of the plurality of sixth probabilities is less than the second probability threshold, it is determined that the second intermediate undetermined result is used to identify that the third to-be-detected image is not of the target type.

[0212] As a possible implementation manner, if the second intermediate undetermined result identifies that the third to-be-detected image is of the to-be-verified type or is not of the target type, the fifth undetermined result can be determined based on the second intermediate undetermined result to identify that the third to-be-detected result is not of the target type, so that the third target detection result can be determined based on the fifth undetermined result subsequently.

[0213] Thus, based on the relationship between the average of the plurality of sixth probabilities and the probability threshold (i.e., the first probability threshold and the second probability threshold), the second intermediate tentative result can be determined, and the second intermediate tentative result is refined by the first probability threshold and the second probability threshold, so that subsequent detection can be continued based on the refined second intermediate tentative result based on the character recognition model, and then the account corresponding to the to-be-detected image is directly processed through model combination, improving processing efficiency and ensuring processing real-time performance. Or, combined with the artificial review method for detection, further improve the detection accuracy and improve the user experience.

[0214] F4: If the second intermediate tentative result identifies the third to-be-detected image as the target type, the third to-be-detected image is subjected to character recognition through the character recognition model to obtain a second character recognition result. If the second character recognition result is the same as the keyword in the keyword set, it is determined that the fifth tentative result is used to identify the third to-be-detected image as the target type. If the second character recognition result is not the same as the keyword in the keyword set, it is determined that the fifth tentative result is used to identify the third to-be-detected image as the target type.

[0215] If the intermediate tentative result identifies the third to-be-detected image as the target type, in order to improve the accuracy of detection, the third to-be-detected image is subjected to character recognition using the character recognition model to obtain a second character recognition result. The second character recognition result is used to identify which characters are included in the third to-be-detected image.

[0216] If the second character recognition result is the same as the keyword in the keyword set, it means that the third to-be-detected image includes characters related to the first behavior category or the target type, thereby determining that the fifth tentative result is used to identify the third to-be-detected image as the target type. Conversely, if the second character recognition result is not the same as the keyword in the keyword set, it is determined that the fifth tentative result is used to identify the third to-be-detected image as the target type.

[0217] Thus, by combining a plurality of multi-type special classification models and a character recognition model, the third to-be-detected image is subjected to secondary detection, so that both detections determine that the third to-be-detected image is the target type, and only when both detections determine that the third to-be-detected image is the target type, the sixth tentative recognition result is used to identify the third to-be-detected image as the target type, thereby improving the accuracy of the sixth tentative recognition result.

[0218] F5: The third to-be-detected image is subjected to detection through the easy-to-recognize special classification model to obtain a third intermediate tentative result.

[0219] For example, the third to-be-detected image is input into the easy-to-identify special classification model, detection is performed through the easy-to-identify special classification model, and a third intermediate to-be-determined result output by the easy-to-identify special classification model is obtained. The third intermediate to-be-determined result is used to identify whether the third to-be-detected image is of the target type, and the easy-to-identify special classification model can further refine which sub-class of the target type the third to-be-detected image is, that is, whether the third to-be-detected image is of the first behavior class.

[0220] F6: If the third intermediate to-be-determined result identifies that the third to-be-detected image is of the first behavior class, text recognition is performed on the third to-be-detected image through the text recognition model, and a third text recognition result is obtained. If the third text recognition result is the same as a keyword in the keyword set, it is determined that the sixth to-be-determined result is used to identify that the third to-be-detected image is of the target type. If the third text recognition result is different from the keyword in the keyword set, it is determined that the sixth to-be-determined result is used to identify that the third to-be-detected image is not of the target type.

[0221] If the third intermediate to-be-determined result identifies that the third to-be-detected image is of the first behavior class, that is, the third to-be-detected image is of the target type, in order to improve the detection accuracy, the third to-be-detected image can be further detected through the text recognition model, that is, detection is performed through the text recognition model, and a third text recognition result is obtained. The third text recognition result is used to identify which text is included in the third to-be-detected image.

[0222] If the third text recognition result is the same as the keyword in the keyword set, it indicates that the third to-be-detected image includes text related to the first behavior class or the target type, and thus it is determined that the sixth to-be-determined result is used to identify that the third to-be-detected image is of the target type. Conversely, if the third text recognition result is different from the keyword in the keyword set, it is determined that the sixth to-be-determined result is used to identify that the third to-be-detected image is not of the target type.

[0223] Therefore, the third to-be-detected image is detected twice through the combination of the multiple multi-type special classification models and the text recognition model, and it is determined that the seventh to-be-determined result is used to identify that the third to-be-detected image is of the target type only when both detections determine that the third to-be-detected image is of the target type, thereby improving the accuracy of the seventh to-be-determined result.

[0224] F7: If the fifth to-be-determined result identifies that the third to-be-detected image is of the target type and the sixth to-be-determined result identifies that the third to-be-detected image is of the target type, it is determined that a third target detection result of the third to-be-detected image is used to identify that the third to-be-detected image is of the target type, and the account corresponding to the third to-be-detected image is processed.

[0225] If the fifth pending result and the sixth pending result both identify the third to-be-detected image as the target type, it indicates that the detection results of the third to-be-detected image are all the target type, and the third target detection result of the third to-be-detected image is determined to identify the third to-be-detected image as the target type. At this time, the third target detection result has a higher credibility, and the account corresponding to the third to-be-detected image can be directly processed, such as a ban processing, a downgrade processing, and the like, which is not limited in the present application.

[0226] Similarly, if the fifth pending result identifies the third to-be-detected image as not the target type, and the sixth pending result identifies the third to-be-detected image as the target type, it indicates that the third to-be-detected image is not determined as the target type. In addition, if the fifth pending result identifies the third to-be-detected image as the target type, and the sixth pending result identifies the third to-be-detected image as not the target type, it indicates that the third to-be-detected image is not determined as the target type. In order to ensure the accuracy of detection, the third target detection result of the third to-be-detected image is determined to identify the third to-be-detected image as the to-be-verified type, so as to prevent misprocessing and cause poor user experience. As a possible implementation manner, if the third target detection result identifies the third to-be-detected image as the to-be-verified type, the accuracy of detection can be further improved through artificial auditing.

[0227] If the fifth pending result identifies the third to-be-detected image as not the target type, and the sixth pending result identifies the third to-be-detected image as not the target type, the third target detection result of the third to-be-detected image is determined to identify the third to-be-detected image as not the target type, and the artificial auditing manner is not required, thereby reducing the number of artificial auditing.

[0228] In addition, it should be noted that the text recognition model used after the multi-type special classification model and the text recognition model used after the easy-to-identify special classification model can be the same text recognition model, or can not be the same text recognition model, which is not limited in the present application.

[0229] As a possible implementation manner, if the text recognition model used after the multi-type special classification model and the text recognition model used after the easy-to-identify special classification model are the same text recognition model, a de-duplication operation can be performed after using the text recognition model, so that the text recognition is performed only once using the text recognition model, thereby improving the detection efficiency.

[0230] Referring to Figure 15 The figure is a schematic diagram of detection based on a multi-type special classification model and an easy-to-identify special classification model according to an embodiment of the present application.

[0231] In Figure 15In the specific implementation, the two multi-type special classification models, the easy-to-recognize special classification model, and the character recognition model are included. The two multi-type special classification models are used to detect the third to-be-detected image to obtain two sixth probabilities. If it is determined based on the two sixth probabilities that the second intermediate to-be-determined result indicates that the third to-be-detected image is of the target type, the character recognition model is further used to recognize the third to-be-detected image to obtain a second character recognition result. If the second character recognition result is the same as a keyword in the keyword set, the fifth to-be-determined result indicates that the third to-be-detected image is of the target type, otherwise, the third to-be-detected image is not of the target type.

[0232] The easy-to-recognize special classification model is used to detect the third to-be-detected image to obtain a third intermediate to-be-determined result. If the third intermediate to-be-determined result indicates that the third to-be-detected image is of the target type, the character recognition model is further used to recognize the third to-be-detected image to obtain a third character recognition result. If the third character recognition result is the same as a keyword in the keyword set, the sixth to-be-determined result indicates that the third to-be-detected image is of the target type, otherwise, the third to-be-detected image is not of the target type.

[0233] If the fifth to-be-determined result indicates that the third to-be-detected image is of the target type and the sixth to-be-determined result indicates that the third to-be-detected image is of the target type, it is determined that the third target detection result indicates that the third to-be-detected image is of the target type, otherwise, the third to-be-detected image is not of the target type. When the third target detection result indicates that the third to-be-detected image is of the target type, the account corresponding to the third target detection image is directly processed to ensure the real-time performance of the processing.

[0234] Therefore, the multi-type special classification model and the character recognition model are combined to obtain the fifth to-be-determined result with higher accuracy, and the easy-to-recognize special classification model and the character recognition model are combined to obtain the sixth to-be-determined result with higher accuracy. Thus, based on the fifth to-be-determined result with higher accuracy and the sixth to-be-determined result with higher accuracy, the third target detection result of the third to-be-detected image is comprehensively determined to improve the accuracy of the third target detection result. Moreover, to ensure that subsequent processing does not cause misprocessing and affect the user experience, when the fifth to-be-determined result and the sixth to-be-determined result both indicate that the third to-be-detected image is of the target type, it is determined that the third to-be-detected image is of the target type, and the account is directly processed to ensure the accuracy of the processing.

[0235] As a possible implementation, if the amount of cheating for the target application is large, that is, the number of behavior categories included in the target type is large, the corresponding embodiments of Figure 14 or Figure 15 may be used for detection. As the amount of cheating for the target application increases, that is, the number of behavior categories included in the target type also increases, in the real-time processing scenario, not only the first recognition category with low recognition difficulty needs to be recognized, but also other behavior categories with some recognition difficulty need to be recognized, so as to improve the accuracy of detection.

[0236] Based on this, after the fine-tuning of the multiple types of special classification models and the easily recognizable special classification models, the embodiments of the present application can not only detect in combination with the character recognition model, but also detect in combination with the whole image classification model, and F8 and F9 are added on the basis of F1-F7, which are specifically described as follows.

[0237] F8: If the second intermediate undetermined result identifies the third to-be-detected image as the target type, the whole image classification model is used to detect the third to-be-detected image, and a seventh undetermined result of the third to-be-detected image is obtained.

[0238] The whole image classification model is used to identify other behavior categories except the first behavior category in the target type, such as the second behavior category or the third behavior category in the target type, and the first behavior category, the second behavior category and the third behavior category are different behavior types. Taking the gun battle game as an example, the first behavior category can be embodied as a function directory on the game interface, the second behavior category can be embodied as a skeleton on the game interface, and the third behavior category can be embodied as a box on the game interface.

[0239] The whole image classification model can be a deep classification network based on a residual structure, which is trained based on image samples including the corresponding behavior category and image samples not including the corresponding behavior category. Moreover, the whole image classification model is trained based on image samples with a resolution greater than a resolution threshold, that is, the resolution of the image samples is relatively large.

[0240] Therefore, if the second intermediate undetermined result identifies the third to-be-detected image as the target type, it means that the probability of the third to-be-detected image being the target type is relatively large, and the third to-be-detected image can be further identified by the whole image classification model to obtain the seventh undetermined result of the third to-be-detected image based on the whole image classification model, and the seventh undetermined result is used to identify whether the third to-be-detected image is the target type.

[0241] F9: If the seventh undetermined result identifies the third to-be-detected image as the target type, and the third intermediate undetermined result identifies the third to-be-detected image as the target type, it is determined that the third target detection result of the third to-be-detected image is the target type, and the account corresponding to the third to-be-detected image is processed.

[0242] If the seventh pending result identifies the third to-be-detected image as the target type and the third intermediate pending result identifies the third to-be-detected image as the target type, it indicates that the third to-be-detected image not only includes the first behavior category but also includes other behavior categories, that is, the same to-be-detected image appears two behavior categories included in the target type, and the to-be-detected image has a very high possibility of being the target type. Based on this, the third target detection result of the third to-be-detected image can be determined as the target type. Since the result is reliable, the account corresponding to the third to-be-detected image can be directly processed to improve the accuracy of detection and improve the user experience.

[0243] Referring to Figure 16 The figure is a schematic diagram of detection based on an integral image classification model provided by an embodiment of the application. Figure 16 The embodiment of the application is based on the embodiment of the application Figure 15 The embodiment of the application is based on the embodiment of the application

[0244] If the seventh pending result identifies the third to-be-detected image as the target type and the third intermediate pending result identifies the third to-be-detected image as the target type, it indicates that the third to-be-detected image not only includes the first behavior category but also includes other behavior categories, that is, the same to-be-detected image appears two behavior categories included in the target type, and the to-be-detected image has a very high possibility of being the target type. Based on this, the third target detection result of the third to-be-detected image can be determined as the target type. Since the result is reliable, the account corresponding to the third to-be-detected image can be directly processed to improve the accuracy of detection and improve the user experience.

[0245] It should be noted that for the same third to-be-detected image, there are multiple ways to determine whether the third to-be-detected image is the target type through the combination of multiple models, and some ways can directly process the account corresponding to the third to-be-detected image belonging to the target type to ensure the requirements of real-time scenarios. If the account corresponding to the third to-be-detected image has been directly processed by one of the ways, other ways can not need to continue or can not process the account again after being completed to ensure that the same account is processed only once within a period of time, improving the trust and experience of users.

[0246] Thus, based on the combination of the multi-type special classification model and the character recognition model, the combination of the easy-to-recognize special classification model and the character recognition model, the comprehensive determination of the third target detection result of the third to-be-detected image based on the fifth to-be-determined result with higher accuracy and the sixth to-be-determined result with higher accuracy, the newly-added whole-image classification model is used to recognize more behavior categories in the target type, and the detection accuracy is improved. Moreover, through the combination of the multi-type special classification model, the easy-to-recognize special classification model and the whole-image classification model, if the same to-be-detected image hits two behavior categories in the target type, the account corresponding to the to-be-detected image can be directly processed to ensure the accuracy and real-time performance of the processing and improve the user experience.

[0247] As a possible implementation manner, the whole-image classification model can include a plurality of whole-image classification models, each of which can recognize different behavior categories, so as to increase the number of whole-image classification models to increase the recognition of the types of behavior categories in the target type and improve the detection accuracy. The following takes two whole-image classification models as an example for description.

[0248] That is, the whole-image classification model includes a first whole-image classification model and a second whole-image classification model, wherein the first whole-image classification model is used to recognize a second behavior category in a plurality of behavior categories included in the target type, and the second whole-image classification model is used to recognize a third behavior category in the plurality of behavior categories included in the target type. The following describes the specific implementation manner of detecting the third to-be-detected image by the whole-image classification model in F8 to obtain the seventh to-be-determined result of the third to-be-detected image, which is described with reference to F81-F83.

[0249] F81: detecting the third to-be-detected image by the first whole-image classification model to obtain a fourth intermediate to-be-determined result of the third to-be-detected image.

[0250] For example, the third to-be-detected image is input into the first whole-image classification model, and the first whole-image classification model is used for detection to obtain a probability that the third to-be-detected image is the target type. If the probability is greater than a corresponding judgment threshold, it is determined that the fourth intermediate to-be-determined result is used to identify that the third to-be-detected image is the target type, otherwise, the fourth intermediate to-be-determined result is used to identify that the third to-be-detected image is not the target type, that is, the fourth intermediate to-be-determined result is used to identify whether the third to-be-detected image is the target type.

[0251] F82: detecting the third to-be-detected image by the second whole-image classification model to obtain a fifth intermediate to-be-determined result of the third to-be-detected image.

[0252] For example, the third to-be-detected image is input into the second whole-image classification model, detection is performed by the second whole-image classification model, a probability that the third to-be-detected image is of the target type is obtained, if the probability is greater than a corresponding judgment black threshold, it is determined that the fifth intermediate to-be-determined result is used to identify that the third to-be-detected image is of the target type, otherwise, the fifth intermediate to-be-determined result is used to identify that the third to-be-detected image is not of the target type, that is, the fifth intermediate to-be-determined result is used to identify whether the third to-be-detected image is of the target type.

[0253] F83: According to the fourth intermediate to-be-determined result and the fifth intermediate to-be-determined result, a seventh to-be-determined result of the third to-be-detected image is obtained.

[0254] Referring to Figure 17 , the figure is a schematic diagram of detection based on the first whole-image classification model and the second whole-image classification model provided by the embodiment of the application. Compared with the embodiment of Figure 16 , the embodiment of Figure 17 includes the first whole-image classification model and the second whole-image classification model. And detection is performed on the third to-be-detected image based on the first whole-image classification model, a fourth intermediate to-be-determined result of the third to-be-detected image is obtained, detection is performed on the third to-be-detected image based on the second whole-image classification model, a fifth intermediate to-be-determined result of the third to-be-detected image is obtained. Finally, according to the fourth intermediate to-be-determined result and the fifth intermediate to-be-determined result, a seventh to-be-determined result of the third to-be-detected image is obtained.

[0255] If the third intermediate to-be-determined result identifies that the third to-be-detected image is of a first behavior category in the target type, and any one of the fourth intermediate to-be-determined result and the fifth intermediate to-be-determined result identifies that the third to-be-detected image is of the target type, that is, the fourth intermediate to-be-determined result identifies that the third to-be-detected image is of a second behavior category in the target type, or the fifth intermediate to-be-determined result identifies that the third to-be-detected image is of a third behavior category in the target type, the seventh to-be-determined result identifies that the third to-be-detected image is of the target type.

[0256] If the third intermediate to-be-determined result identifies that the third to-be-detected image is not of the first behavior category in the target type, and the fourth intermediate to-be-determined result and the fifth intermediate to-be-determined result both identify that the third to-be-detected image is of the target type, that is, the fourth intermediate to-be-determined result identifies that the third to-be-detected image is of the second behavior category in the target type, and the fifth intermediate to-be-determined result identifies that the third to-be-detected image is of the third behavior category in the target type, the seventh to-be-determined result identifies that the third to-be-detected image is of the target type.

[0257] Therefore, by combining the multiple-type special classification model, the easily-recognized special classification model, the first whole-image model and the second whole-image model, the number of recognized behavior categories is further improved, and the accuracy of the third target detection result is improved. If any two of the third intermediate pending result, the fourth intermediate pending result and the fifth intermediate pending result indicate that the third to-be-detected image belongs to the corresponding behavior category in the target type, it means that the third to-be-detected image hits at least two behavior categories in the three behavior categories, and the possibility of the third to-be-detected image being the target type is greater. Therefore, the third target detection result of the third to-be-detected image is determined to identify that the third to-be-detected image is the target type, and the account corresponding to the third to-be-detected image is directly processed to ensure the accuracy and real-time performance of the processing and improve the user experience. In addition, the first whole-image classification model and the second whole-image classification model can be used as a supplementary model of the character recognition model. When there are traditional Chinese characters that cannot be recognized by the character recognition model, the same image can hit two behavior categories and be processed in real time to improve the detection accuracy.

[0258] As a possible implementation manner, since the whole-image recognition model needs to detect the complete third to-be-detected image and hit at least two behavior categories, in some applications, the traces (referred to as cheating traces) of some behavior categories are relatively small, and the judgment threshold is generally high, so the third to-be-detected image may not be correctly detected as the target type, and the manual review mode may also be combined.

[0259] In order to reduce the amount of manual review and improve the detection accuracy, it is found through analysis that the use of external plug-ins by the target application is generally reflected in the center of the application interface, such as the perspective cheating (such as displaying the opponent game character that is blocked, etc.) around the center position of the game interface in the gun battle game, and the proportion of such behavior categories is relatively large, that is, the range of the non-behavior category region in the to-be-detected image accounts for a large proportion, and the scene changes are complex.

[0260] Based on this, the embodiment of the present application adds a local image special classification model, which is used to identify behavior categories other than the first behavior category in the target type, such as the second behavior category or the third behavior category, etc. Unlike the whole-image classification model, the local image special classification model is obtained based on image samples with a resolution less than or equal to a resolution threshold.

[0261] In other words, the size of the image sample used to train the local map dedicated classification model is smaller than the image sample used to train the whole map classification model. For example, the image sample used to train the whole map classification model can be cropped to obtain the image sample used to train the local map dedicated classification model, with the center of the application interface as the center. Thus, the range of the non-behavior category area included in the image sample used to train the local map dedicated classification model is smaller, and the complexity of the scene change is smaller, so that the local map dedicated classification model obtained by training can be targeted for detection, thereby improving the detection accuracy.

[0262] The process of using the local map dedicated classification model is described below, and specific reference is made to F10-F12.

[0263] F10: If the second intermediate tentative result identifies the third to-be-detected image as the target type, the third to-be-detected image is cropped based on the center of the application interface of the target application to obtain a third to-be-detected local image.

[0264] If the second intermediate tentative result identifies the third to-be-detected image as the target type, it means that the probability of the third to-be-detected image being the target type is relatively large. In order to remove the proportion of other areas, the third to-be-detected image can be cropped, such as cropping the third to-be-detected image based on the center of the application interface to obtain a local image with a size of 400x300, i.e., the third to-be-detected local image. Thus, compared with the third to-be-detected image, the third to-be-detected local image filters out many irrelevant backgrounds, which can increase the proportion of including a certain behavior category and reduce the recognition difficulty.

[0265] F11: The third to-be-detected local image is detected by the local map dedicated classification model to obtain an eighth tentative result of the third to-be-detected image.

[0266] The third to-be-detected local image is continuously recognized by the local map dedicated classification model to obtain an eighth tentative result of the third to-be-detected image based on the local map dedicated classification model, and the eighth tentative result is used to identify whether the third to-be-detected image is the target type.

[0267] For example, the third to-be-detected local image is input into the local map dedicated classification model, and the third to-be-detected local image is detected by the local map dedicated classification model to obtain an eighth tentative result of the third to-be-detected image.

[0268] The local map dedicated classification model is not specifically limited in the embodiments of the present application, and can be a middle-level classification network designed based on a residual structure, etc.

[0269] F12: If the eighth pending result indicates that the third to-be-detected image is of the target type, it is determined that the third target detection result of the third to-be-detected image is of the target type, and the account corresponding to the third to-be-detected image is processed.

[0270] If the eighth pending result indicates that the third to-be-detected image is of the target type, it means that there is a greater probability of a small and difficult-to-identify behavior category in the third to-be-detected image. Therefore, it can be determined that the third target detection result of the third to-be-detected image is of the target type. That is, not only is the third to-be-detected image determined to be of the target type by the local image special recognition model, but also is determined to be of the target type based on the multi-type special classification model or the easy-to-identify special classification model. The detection result has a higher credibility, and the account corresponding to the third to-be-detected image can be directly processed to improve the accuracy of detection and improve the user experience while processing in real time.

[0271] Referring to Figure 18 , the figure is a schematic diagram of detection based on a local image special classification model provided by an embodiment of the present application. Figure 18 The embodiment of the present application is based on the embodiment of the present application Figure 15 The embodiment of the present application is based on the embodiment of the present application

[0272] If the eighth pending result indicates that the third to-be-detected image is of the target type, it is determined that the third target detection result of the third to-be-detected image is of the target type, and the account corresponding to the third to-be-detected image is processed. If the eighth pending result indicates that the third to-be-detected image is not of the target type, it is determined that the third target detection result of the third to-be-detected image is not of the target type, and the account corresponding to the third to-be-detected image is not directly processed. It can be processed indirectly, such as determining whether to process the account corresponding to the third to-be-detected image in combination with an artificial auditing method.

[0273] It should be noted that for the same third to-be-detected image, there are multiple ways to determine whether the third to-be-detected image is of the target type through the combination of multiple models, and some of the ways can directly process the account corresponding to the third to-be-detected image of the target type to ensure the requirements of real-time scenarios. If one of the ways has already directly processed the account corresponding to the third to-be-detected image, the other ways can not need to continue or can not need to process the account again after being completed to ensure that the same account is processed only once within a period of time, thereby improving the trust and experience of the user.

[0274] Thus, on the basis of combining the multi-type special classification model and the character recognition model, combining the easily recognizable special classification model and the character recognition model, and comprehensively determining the third target detection result of the third to-be-detected image based on the fifth to-be-determined result with higher accuracy and the sixth to-be-determined result with higher accuracy, a local image special classification model is added to recognize more behavior categories in the target type, and the third to-be-detected image is removed by cropping, so that more fine behavior categories can be recognized, the detection accuracy is improved, and if the eighth to-be-determined result obtained by the local image special classification model identifies the third to-be-detected image as the target type, the account corresponding to the to-be-detected image can be directly processed to ensure the accuracy and real-time performance of the processing and improve the user experience.

[0275] As a possible implementation manner, the local image special classification model can include a plurality of local image special classification models, each of which can recognize different behavior categories, so as to increase the number of local image special classification models to increase the recognition of the types of behavior categories in the target type and improve the detection accuracy. The following takes two local image special classification models as an example for description.

[0276] That is, the local image special classification model includes a first local image special classification model and a second local image special classification model, wherein the first local image special classification model is used to recognize a second behavior category in a plurality of behavior categories included in the target type, and the second local image special classification model is used to recognize a third behavior category in the plurality of behavior categories included in the target type. For example, the first local image special classification model is used to recognize whether the third to-be-detected image includes the aforementioned box, and the second local image special classification model is used to recognize whether the third to-be-detected image includes the aforementioned bone, etc.

[0277] The following describes a specific implementation manner of detecting the third to-be-detected local image by the local image special classification model in F11 to obtain the eighth to-be-determined result of the third to-be-detected image, which is described with reference to F111-F113.

[0278] F111: detecting the third to-be-detected local image by the first local image special classification model to obtain a sixth intermediate to-be-determined result of the third to-be-detected image.

[0279] For example, the third local image to be detected is input into the first local image dedicated classification model. The first local image dedicated classification model is used to detect the third local image to be detected, or the probability that the third local image to be detected is the target type. If the probability is greater than the corresponding black threshold, the sixth intermediate pending result is determined to identify the third local image to be detected as the target type. Otherwise, the sixth intermediate pending result is used to identify the third local image to be detected as not being the target type. That is, the sixth intermediate pending result is used to identify whether the third local image to be detected is the target type.

[0280] F112: The third local image to be detected is detected by the second local image-specific classification model, and the seventh intermediate undetermined result of the third local image to be detected is obtained.

[0281] For example, the third local image to be detected is input into the second local image-specific classification model. The second local image-specific classification model is used for detection to obtain the third local image to be detected, or in other words, the probability that the third local image to be detected is the target type. If the probability is greater than the corresponding black threshold, the seventh intermediate pending result is used to identify the third local image to be detected as the target type. Otherwise, the seventh intermediate pending result is used to identify the third local image to be detected as not being the target type. That is, the seventh intermediate pending result is used to identify whether the third local image to be detected is the target type.

[0282] F113: If the sixth intermediate pending result identifies the third image to be detected as the second row category, or the seventh intermediate pending result identifies the third image to be detected as the third row category, then the eighth pending result is determined to identify the third image to be detected as the target type.

[0283] See Figure 19 This figure is a schematic diagram illustrating a detection method based on a first local graph-specific classification model and a second local graph-specific classification model, provided in an embodiment of this application. Compared to Figure 18 Implementation examples, Figure 19 The embodiment includes a first local image-specific classification model and a second local image-specific classification model. The third local image to be detected is detected based on the first local image-specific classification model to obtain a sixth intermediate pending result for the third local image to be detected. The third local image to be detected is then detected based on the second local image-specific classification model to obtain a seventh intermediate pending result for the third local image to be detected. Finally, based on the sixth and seventh intermediate pending results, an eighth pending result for the third local image to be detected is obtained.

[0284] If the sixth intermediate pending result identifies the third image to be detected as the second row category, that is, if the sixth intermediate pending result identifies the third image to be detected as the target type, then the eighth pending result is determined to identify the third image to be detected as the target type.

[0285] If the seventh intermediate undetermined result identifies the third to-be-detected image as the second behavior category, i.e., the seventh intermediate undetermined result identifies the third to-be-detected image as the target type, it is determined that the eighth undetermined result is used to identify the third to-be-detected image as the target type.

[0286] Therefore, by combining the multi-type special classification model, the easy-to-identify special classification model, the first local image special model, and the second local image special classification model, the number of recognized behavior categories is further improved, and the accuracy of the third target detection result is improved. Moreover, after the second intermediate undetermined result identifies the third to-be-detected image as the target type, the third to-be-detected image is cropped to remove irrelevant backgrounds, and then the cropped third to-be-detected image is detected based on the first local image special model and the second local image special classification model, respectively. If the third target detection result of the third to-be-detected image is used to identify the third to-be-detected image as the target type, the corresponding account of the third to-be-detected image is directly processed to ensure the accuracy and real-time performance of the processing and improve the user experience. In addition, the first local image special classification model and the second local image special classification model can be used as a supplementary model of the character recognition model. When there are traditional Chinese characters that cannot be recognized by the character recognition model, the third to-be-detected local image can be combined with the corresponding behavior category to make real-time processing and improve the detection accuracy.

[0287] In addition, it should be noted that the number of images obtained by the user by taking screenshots of the target application installed in the terminal device during use of the target application is large, which may cause a sharp increase in subsequent detection and unable to quickly detect.

[0288] Therefore, the pre-filtering classification model is installed in the terminal device used by the user, so as to preferentially filter out images that are not of the target type based on the pre-filtering classification model, reduce the number of to-be-detected images (such as the first to-be-detected image, the second to-be-detected image, and the third to-be-detected image), and improve the detection speed.

[0289] Specifically, during the running of the target application, the running process of the target application is taken to obtain an undetermined to-be-detected image. The pre-filtering classification model is used to detect the undetermined to-be-detected image. If the probability that the undetermined to-be-detected image is of the target type is greater than a third probability threshold, the undetermined to-be-detected image is determined as a to-be-detected image. If the probability that the undetermined to-be-detected image is of the target type is less than or equal to the third probability threshold, the undetermined to-be-detected image is deleted, i.e., the subsequent process does not need to continue to detect the undetermined to-be-detected image, thereby reducing the detection quantity and improving the detection speed.

[0290] The pre-filter classification model can be trained based on images belonging to the target type and images not belonging to the target type as training data. For example, the pre-filter classification model can be based on a GhostNet model.

[0291] As a possible implementation, the pre-filter classification model has a magnitude lower than a preset magnitude threshold, thereby reducing the performance requirements of the pre-filter classification model on the terminal device and improving user experience. Further, if the pre-filter classification model is small and has weak detection capability, the proportion of images belonging to the target type in the training data can be increased, or the weight of images belonging to the target type in the loss function can be increased, to enhance the identification capability of the pre-filter classification model for images belonging to the target type, so as to avoid screening out images belonging to the target type, thereby improving the detection accuracy while improving the subsequent detection speed.

[0292] To facilitate further understanding of the technical solutions provided in the embodiments of the present application, the following takes a server as the execution subject of the data detection method provided in the embodiments of the present application, takes a game as the target application, and takes using a game plug-in as the target type as an example to introduce the data detection method in the whole.

[0293] The following first describes the data detection system.

[0294] Referring to Figure 20 , the figure is a schematic diagram of a data detection system provided in an embodiment of the present application. In Figure 20 , the data detection system includes a data layer 2010, an algorithm layer 2020, a base model layer 2030, a system layer 2040, and a scenario layer 2050. The following are described respectively.

[0295] (1) The data layer 2010 is configured to collect images for training or fine-tuning a model. The data layer 2010 includes general game images 2011, special game images 2012, local game images 2013, and text game images 2014.

[0296] The general game images 2011 are configured to collect screenshots from multiple games, including screenshots using game plug-ins (i.e., images belonging to the target type) and screenshots not using game plug-ins (i.e., images not belonging to the target type). The data volume is generally in the order of hundreds of thousands to millions.

[0297] The special game image 2012 is used to collect complete screenshots of each behavior category corresponding to the game plug-in in multiple games, i.e., complete images belonging to each behavior category in the target type, such as complete screenshots including function directories (generally in the order of magnitude of ten thousand), complete screenshots including boxes (generally in the order of magnitude of several thousand to tens of thousands), complete screenshots including skeletons (generally in the order of magnitude of several thousand to tens of thousands), etc. Thus, the multiple-type special classification model and the easily-recognized special classification model are obtained by fine-tuning the images collected based on the special game image 2012.

[0298] The local game image 2013 is used to collect partial screenshots of each behavior category corresponding to the game plug-in in multiple games, i.e., partial images belonging to each behavior category in the target type, such as partial screenshots including boxes, partial screenshots including skeletons, etc. The data volume is generally in the order of magnitude of several thousand to tens of thousands.

[0299] The text game image 2014 is used to collect screenshots of text boxes and text annotations using the game plug-in in multiple games. The data volume is generally in the order of magnitude of several thousand to tens of thousands.

[0300] (2) The algorithm layer 2020 is used to provide algorithms used by the model. The algorithm layer 2020 includes a pre-filtering algorithm 2021, a multiple-type classification algorithm 2022, an entire image classification algorithm 2023, a local image classification algorithm 2024, and a text recognition algorithm 2025.

[0301] The pre-filtering algorithm 2021 is used to provide an algorithm for training a pre-filtering classification model, such as GhostNet. The screenshots for the target application collected using the general game image 2011 can not be distinguished as belonging to the target type.

[0302] The multiple-type classification algorithm 2022 is used to provide an algorithm for training a multiple-type general model and a multiple-type special model, such as MobileNetV3, a shallow classification network based on a residual structure design, etc., and uses the screenshots of multiple applications and the screenshots for the target application collected using the general game image 2011.

[0303] The entire image classification algorithm 2023 is used to provide an algorithm for training an entire image classification model, such as a deep classification network based on a residual structure design, etc., and uses the screenshots for the target application collected using the special game image 2012.

[0304] The local image classification algorithm 2024 is used to provide an algorithm for training an easily-recognized general classification model, an easily-recognized special classification model, a local image general classification model, and a local image special classification model, such as a middle-level classification network based on a residual structure design, etc., and uses the screenshots for the target application collected using the local game image 2013.

[0305] The text recognition algorithm 2025 is used to improve the algorithm for training the text recognition model, such as PaddleOCR, and uses the screenshots collected by the text game image 2014 for the target application.

[0306] (3) The base model layer 2030 is used to provide a base model, i.e., a model structure used for training the model. It is used for cold start on new game screenshots without accumulated labeled game screenshots, or fine-tuning on game screenshots with a moderate amount of labeled screenshot data. The base model layer 2030 includes a pre-filter base model 2031, a multi-type classification base model 2032, an entire image classification base model 2033, a local image classification base model 2034, and a text base model 2035.

[0307] The pre-filter base model 2031 is used to provide a base model for training a pre-filter classification model, and the pre-filter classification model is trained by combining the algorithm provided by the pre-filter layer algorithm 2021. The pre-filter classification model is trained to filter a large number of non-target type screenshots on the terminal device side, reducing the computational cost.

[0308] The multi-type classification base model 2032 is used to provide a base model for training a multi-type general classification model, and the multi-type general classification model and the multi-type special classification model are trained by combining the algorithm provided by the multi-type classification algorithm 2022. This is done to filter a large number of non-target type screenshots on the server side, reducing the computational cost.

[0309] The entire image classification base model 2033 is used to provide a base model for training an entire image classification model, and the entire image classification model is trained by combining the algorithm provided by the entire image classification algorithm 2033. This is done to improve the detection types of behavior categories and improve detection accuracy.

[0310] The local image classification base model 2034 is used to provide a base model for training a local image general classification model and a base model for training an easily recognizable general classification model, and the easily recognizable general classification model, the easily recognizable special classification model, the local image general classification model, and the local image special classification model are trained by combining the algorithm provided by the local image classification algorithm 2024. This is done to improve the detection types of behavior categories and improve detection accuracy.

[0311] The text base model 2035 is used to provide a base model for training a text recognition model, and the text recognition model is trained by combining the algorithm provided by the text recognition algorithm 2025. This is done to improve the detection types of behavior categories and improve detection accuracy.

[0312] (4) The system layer 2040 is used for assisting data collection, data labeling, various training sample arrangement and model version management, assisting model training and effect evaluation, multi-model online and overall process strategy configuration, etc. The system layer includes data preprocessing 2041, model training and evaluation 2042, model prediction 2043, strategy configuration 2044 and keyword set configuration 2045.

[0313] The data preprocessing 2041 is used for assisting data collection, data labeling, data generation, data cleaning, etc. to improve the effectiveness of the data.

[0314] The model training and evaluation 2042 is used for various training sample arrangement and model version management, such as setting various base models in the base model layer 2030, setting training data, initial learning rate, maximum iteration number, batch number, learning strategy, fine-tuning strategy, etc. as shown in Table 1, thereby assisting model training and effect evaluation.

[0315] Table 1

[0316]

[0317] The model prediction 2043 is used for providing multi-model online service.

[0318] The strategy configuration 2044 is used for improving the overall process strategy configuration, such as determining which model combination can directly process the target detection result of the account, etc.

[0319] The keyword set configuration 2045 is used for configuring the keywords included in the keyword set of various applications.

[0320] (5) The scene layer 2050 is used for configuring the relationship between the scene identification and the model. The scene layer 2050 includes a cold start scene 2051 and a real-time processing scene 2052.

[0321] The cold start scene 2051 can use a combination of multi-type general classification models and local graph general classification models to obtain a target detection result for a to-be-detected image, etc. as shown in the embodiment of Figure 7 , the embodiment of Figure 11 or the embodiment of Figure 12 .

[0322] The real-time processing scene 2052 can use a combination of multi-type general classification models, easily recognizable general classification models and text recognition models to obtain a target detection result for a to-be-detected image, etc. as shown in the embodiment of Figures 14-19 .

[0323] As can be seen from the above data detection system, the embodiments of this application provide a flexible and rapid adaptation solution to meet different screenshot countermeasure requirements. Different models can be combined and quickly fine-tuned during online and operational phases to meet different screenshot countermeasure requirements.

[0324] For example, in a newly launched game, i.e., a cold start scenario, a pre-filtering model is trained based on screenshots collected using the general game screenshot 2011 model, and a pre-filtering basis model 2031 and a pre-filtering algorithm 2021. This model is then embedded into the terminal device. This lightweight pre-filtering model filters a large number of normal screenshots (i.e., non-target type screenshots) locally within the game application, reducing suspicious reports (screenshots that may belong to the target type). When the images to be detected are reported to the cloud server for detection, a multi-type general classification model is deployed to ensure coverage of classic cheating types (i.e., frequently occurring target types). Simultaneously, a general classification model is deployed to identify local images such as bounding boxes, skeletons, and health points to ensure coverage of subtle cheating traces, thus addressing the cold start scenario. For details, please refer to [link to relevant documentation]. Figure 7 The illustrated embodiments Figure 11 The illustrated embodiments or Figure 12 The example shown.

[0325] For example, in real-time processing scenarios, where it is necessary to quickly detect and process relevant accounts, an easy-to-identify general classification model and a text recognition model for the function directory can be launched to quickly and accurately identify screenshots belonging to the target type.

[0326] Furthermore, after the model has been online for a period of time and a certain number of review screenshots have been accumulated, i.e., review results, the multi-type general classification model and the easily identifiable general classification model can be fine-tuned based on the review results, thereby obtaining multi-type special classification models and easily identifiable special classification models, improving accuracy and coverage.

[0327] Furthermore, by annotating data such as function directories, bounding boxes, and skeletons, specialized classification models for local images and whole-image classification models are trained to refine the identification of target types. This allows for real-time automatic punishment of most cheating screenshots through machine review, with only a small number of applications requiring manual review to cover the remaining issues, thus reducing the workload of manual review. This enables rapid integration with game screenshot detection, flexible adaptation to different scenario requirements for screenshot cheating detection, and efficient operation against cheating. For details, please refer to [link / reference]. Figures 14-19 The illustrated embodiment, or Figure 21 The combined scheme shown is to obtain the target detection result by detecting in parallel through multiple methods.

[0328] Moreover, in Figure 21In this system, for the same third image to be detected, multiple models can be combined to determine whether the third image belongs to the target type. Some models can directly process the account corresponding to the third image that belongs to the target type to ensure the requirements of real-time scenarios. If one model has already processed the account corresponding to the third image, other models do not need to continue, or after processing, they will not process the account again, to ensure that the same account is processed only once within a certain period of time, thereby improving user trust and experience.

[0329] As can be seen from the above technical solution, when a new game is launched, screenshots can be quickly detected, anti-cheating measures can be implemented, a large number of normal screenshots can be filtered out, and suspicious cheating images can be selected for manual review, reducing the workload of manual review. After one or two weeks of operation, based on the review results obtained by the reviewers or the newly marked data, [further details can be added]. Figure 20 The system rapidly trains new models across its base model layer, algorithm layer, and system layer, improving the accuracy and coverage of cheat detection and enabling efficient game operation. Alternatively, when model updates are needed during adversarial gameplay, only training data is required for rapid model fine-tuning; this can involve fine-tuning only the last layer or performing full-scale training.

[0330] Furthermore, the embodiments of this application are flexibly adaptable to games with varying degrees of urgency and intensity of confrontation. They can quickly adapt to the needs of real-time penalties, providing rapid response methods for different types of cheating. Moreover, only training data is required to train a refined model for penalty requirements. This allows for flexible handling of diverse confrontational demands.

[0331] For games that only need to ensure cheat coverage and rely on manual review as a fallback, a system can be built. Figure 11 The model combination ensures a review volume of only a few hundred or even a hundred, with lightweight operation. For games requiring real-time security monitoring, deployment is possible immediately upon game integration. Figure 14 The model combination allows for one-click screenshotting and use. Furthermore, it can be customized based on the game's importance and required combat effects. Figures 15-19 The model combination is designed to meet different proportions of automatic direct penalties, satisfying the game's required combat tasks while minimizing the provider's manpower costs.

[0332] In response to the data detection method described above, this application also provides a corresponding data detection device so that the above data detection method can be applied and implemented in practice.

[0333] See Figure 22 The figure is a schematic diagram of the structure of a data detection device provided in an embodiment of this application. Figure 22As shown, the data detection apparatus 2200 comprises an acquisition unit 2201, a first detection unit 2202, a second detection unit 2203, and a determination unit 2204.

[0334] The acquisition unit 2201 is configured to acquire a first to-be-detected image and a scene identifier, the first to-be-detected image being an image obtained by taking a screenshot during running of a target application, and the scene identifier being used to identify a detection scene type corresponding to the target application.

[0335] The first detection unit 2202 is configured to detect the first to-be-detected image by using a plurality of multi-type general classification models respectively, to obtain a first probability that the first to-be-detected image is of a target type output by each of the multi-type general classification models, and to determine a first pending result of the first to-be-detected image according to a plurality of the first probabilities, the plurality of multi-type general classification models being different models.

[0336] The second detection unit 2203 is configured to determine a target scene supplement model suitable for the target application from a plurality of scene supplement models according to the scene identifier, to detect the first to-be-detected image by using the target scene supplement model, and to obtain a second pending result, different scene supplement models being suitable for different detection scene types.

[0337] The determination unit 2204 is configured to determine a first target detection result of the first to-be-detected image according to the first pending result and the second pending result, the first target detection result being used to identify whether a type of the first to-be-detected image is the target type.

[0338] It can be seen from the technical solution that the data detection device provided by the embodiment of the application includes an acquisition unit, a first detection unit, a second detection unit and a determination unit. In the first aspect, the application log data is no longer used for identification, but a screenshot is taken during the running of the target application, and the obtained image is used as a first detection image. Since the first detection image is a real process, the possibility of being tampered with is small, and it is not easy to be counteracted, thereby improving the accuracy of subsequent detection. In the second aspect, instead of using a binary classification model for a certain abnormal behavior for detection, a multi-classification model, i.e., a multi-type general classification model, is used for detection, thereby realizing detection of multiple abnormal behaviors by one model. In the third aspect, in order to improve the detection accuracy of the multi-type general classification model, instead of using one model, multiple multi-type general classification models are used to detect the first detection image respectively, and a first probability that the first detection image is of a target type output by the multi-type general classification model is obtained, thereby improving the accuracy of the first detection image. In the fourth aspect, in order to expand the range of application scenarios, multiple scene supplement models are pre-trained, different scene supplement models are suitable for different detection scene types, and then based on the obtained scene identifier, a target scene supplement model suitable for the detection scene type corresponding to the scene identifier is obtained from the multiple scene supplement models, and the first detection image is detected based on the target scene supplement model to obtain a second tentative result. In the fifth aspect, by using the first tentative result with universality obtained based on the multiple multi-type general classification models and the second tentative result obtained by the target scene supplement model, the first target detection result of the first detection image is determined, i.e., whether the first detection image is of the target type, such as whether the first detection image includes abnormal behavior, thereby further improving the accuracy of detection.

[0339] As a possible implementation manner, if the scene identifier represents that the detection scene type corresponding to the target application is a cold start scene, the target scene supplement model is a local image general classification model, and the device includes a training unit configured to:

[0340] Obtain multiple image samples, the image samples have labels for identifying the types of the image samples, and the types of the multiple image samples have multiple types;

[0341] Crop each of the image samples according to the label of each of the image samples to obtain multiple local image samples;

[0342] Detect each of the local image samples by using an initial local image general classification model to obtain a first detection result of each of the image samples;

[0343] According to a difference between each first detection result and a label of a corresponding image sample, a model parameter of the initial local graph general classification model is adjusted to obtain the local graph general classification model.

[0344] As a possible implementation manner, the determination unit 2204 is specifically configured to:

[0345] If the first pending result indicates that the first to-be-detected image is of the target type or the second pending result indicates that the first to-be-detected image is of the target type, a first target detection result of the first to-be-detected image is determined to indicate that the first to-be-detected image is of the target type.

[0346] As a possible implementation manner, the apparatus comprises a fine-tuning unit configured to:

[0347] A review result of the first to-be-detected image is obtained, and accuracy of the review result is greater than an accuracy threshold;

[0348] If the review result is different from the first target detection result, the first to-be-detected image is detected by a target multi-type general classification model to obtain a second probability that the first to-be-detected image is of the target type, and the target multi-type general classification model is one of the plurality of multi-type general classification models;

[0349] According to the second probability, a second detection result of the first to-be-detected image is determined;

[0350] According to a difference between the second detection result and the review result, a model parameter of the target multi-type general classification model is adjusted to obtain a multi-type special classification model;

[0351] The multi-type special classification model is used to replace the multi-type general classification model.

[0352] As a possible implementation manner, the apparatus further comprises an application unit configured to:

[0353] After the multi-type special classification model is used to replace the multi-type general classification model, the plurality of multi-type general classification models are respectively used as the target multi-type general classification model to obtain a multi-type special classification model corresponding to each multi-type general classification model;

[0354] A second to-be-detected image is obtained, and the second to-be-detected image is an image obtained by taking a screenshot during running of the target application;

[0355] determining, according to the plurality of third probabilities, a third pending result of the second to-be-detected image;

[0356] determining, according to the plurality of third probabilities, a third pending result of the second to-be-detected image;

[0357] As a possible implementation manner, the application unit is specifically configured to:

[0358] determining, according to the plurality of third probabilities, a third pending result of the second to-be-detected image;

[0359] detecting, by the local graph general classification model, the second to-be-detected image to obtain a fourth pending result of the second to-be-detected image;

[0360] determining, according to the third pending result and the fourth pending result, a second target detection result of the second to-be-detected image.

[0361] As a possible implementation manner, if the scene identifier represents that the detection scene type corresponding to the target application is a real-time processing scene, the target scene supplement model includes an easily recognizable general classification model and a character recognition model, the easily recognizable general classification model is used to identify a first behavior category in a plurality of behavior categories included in the target type, and the identification difficulty of the first behavior category is less than a difficulty threshold; the second detection unit 2203 is specifically configured to:

[0362] detecting, by the easily recognizable general classification model, the first to-be-detected image to obtain a first intermediate pending result;

[0363] if the first intermediate pending result identifies the first to-be-detected image as the first behavior category, performing character recognition on the first to-be-detected image by the character recognition model to obtain a first character recognition result;

[0364] if the first character recognition result is the same as a keyword in a keyword set, determining that the second pending result is used to identify that the first to-be-detected image is the target type, and the keyword in the keyword set is related to the first behavior category;

[0365] if the first character recognition result is not the same as the keyword in the keyword set, determining that the second pending result is used to identify that the first to-be-detected image is not the target type.

[0366] As a possible implementation manner, the apparatus further includes a fine-tuning unit configured to:

[0367] obtaining a review result for the first image to be detected, an accuracy of the review result being greater than an accuracy threshold;

[0368] if the review result is different from the first target detection result, detecting the first image to be detected by a target multi-type general classification model to obtain a fourth probability that the first image to be detected is of the target type, and detecting the first image to be detected by the easily recognizable general classification model to obtain a fifth probability that the first image to be detected is of the target type, the target multi-type general classification model being one of the plurality of multi-type general classification models;

[0369] determining a third detection result of the first image to be detected according to the fourth probability, adjusting model parameters of the target multi-type general classification model according to a difference between the third detection result and the review result, obtaining a multi-type special classification model, and replacing the multi-type general classification model with the multi-type special classification model;

[0370] determining a fourth detection result of the first image to be detected according to the fifth probability, adjusting model parameters of the easily recognizable general classification model according to a difference between the fourth detection result and the review result, obtaining an easily recognizable special classification model, and replacing the easily recognizable general classification model with the easily recognizable special classification model.

[0371] As a possible implementation manner, the apparatus further includes an application unit configured to:

[0372] obtaining a third image to be detected, the third image to be detected being an image obtained by taking a screenshot during running of the target application;

[0373] detecting the third image to be detected by the plurality of multi-type special classification models respectively to obtain a sixth probability that the third image to be detected is of the target type output by each of the plurality of multi-type special classification models;

[0374] determining a second intermediate tentative result of the third image to be detected according to the plurality of sixth probabilities;

[0375] if the second intermediate tentative result indicates that the third image to be detected is of the target type, performing character recognition on the third image to be detected by the character recognition model to obtain a second character recognition result, if the second character recognition result is the same as a keyword in the keyword set, determining a fifth tentative result for indicating that the third image to be detected is of the target type, and if the second character recognition result is not the same as the keyword in the keyword set, determining the fifth tentative result for indicating that the third image to be detected is not of the target type;

[0376] detecting the third to-be-detected image by using the easily recognizable special classification model to obtain a third intermediate to-be-determined result;

[0377] If the third intermediate to-be-determined result indicates that the third to-be-detected image belongs to the first behavior category, performing text recognition on the third to-be-detected image by using the text recognition model to obtain a third text recognition result; if the third text recognition result is the same as a keyword in the keyword set, determining a sixth to-be-determined result for indicating that the third to-be-detected image belongs to the target type; if the third text recognition result is not the same as the keyword in the keyword set, determining the sixth to-be-determined result for indicating that the third to-be-detected image does not belong to the target type.

[0378] If the fifth to-be-determined result indicates that the third to-be-detected image belongs to the target type and the sixth to-be-determined result indicates that the third to-be-detected image belongs to the target type, determining a third target detection result of the third to-be-detected image for indicating that the third to-be-detected image belongs to the target type, and processing an account corresponding to the third to-be-detected image.

[0379] As a possible implementation manner, the apparatus further includes an application unit configured to:

[0380] If the fifth to-be-determined result indicates that the third to-be-detected image does not belong to the target type and the sixth to-be-determined result indicates that the third to-be-detected image does not belong to the target type, determining a third target detection result of the third to-be-detected image for indicating that the third to-be-detected image does not belong to the target type.

[0381] If the fifth to-be-determined result indicates that the third to-be-detected image belongs to the target type and the sixth to-be-determined result indicates that the third to-be-detected image does not belong to the target type, determining a third target detection result of the third to-be-detected image for indicating that the third to-be-detected image belongs to a to-be-verified type.

[0382] If the fifth to-be-determined result indicates that the third to-be-detected image does not belong to the target type and the sixth to-be-determined result indicates that the third to-be-detected image belongs to the target type, determining a third target detection result of the third to-be-detected image for indicating that the third to-be-detected image belongs to the to-be-verified type.

[0383] As a possible implementation manner, the apparatus further includes an application unit configured to:

[0384] If the average value of the plurality of sixth probabilities is greater than or equal to a first probability threshold, determining that the second intermediate to-be-determined result is used for indicating that the third to-be-detected image belongs to the target type.

[0385] If the average of the plurality of sixth probabilities is less than the first probability threshold and the average of the plurality of sixth probabilities is greater than or equal to a second probability threshold, it is determined that the second intermediate pending result is used to identify that the third to-be-detected image is of a to-be-verified type, and the second probability threshold is less than the first probability threshold.

[0386] If the average of the plurality of sixth probabilities is less than the second probability threshold, it is determined that the second intermediate pending result is used to identify that the third to-be-detected image is not of the target type.

[0387] As a possible implementation manner, the apparatus further includes an application unit configured to:

[0388] After the second intermediate pending result of the third to-be-detected image is determined according to the plurality of sixth probabilities, if the second intermediate pending result identifies that the third to-be-detected image is of the target type, the third to-be-detected image is detected by using a whole-image classification model to obtain a seventh pending result of the third to-be-detected image, the whole-image classification model is used to identify other behavior categories than the first behavior category in the plurality of behavior categories included in the target type, and the whole-image classification model is trained based on image samples with a resolution greater than a resolution threshold.

[0389] If the seventh pending result identifies that the third to-be-detected image is of the target type and the third intermediate pending result identifies that the third to-be-detected image is of the target type, it is determined that a third target detection result of the third to-be-detected image is of the target type, and an account corresponding to the third to-be-detected image is processed.

[0390] As a possible implementation manner, the whole-image classification model includes a first whole-image classification model and a second whole-image classification model, the first whole-image classification model is used to identify a second behavior category in the plurality of behavior categories included in the target type, the second whole-image classification model is used to identify a third behavior category in the plurality of behavior categories included in the target type, and the apparatus further includes an application unit configured to:

[0391] The third to-be-detected image is detected by using the first whole-image classification model to obtain a fourth intermediate pending result of the third to-be-detected image.

[0392] The third to-be-detected image is detected by using the second whole-image classification model to obtain a fifth intermediate pending result of the third to-be-detected image.

[0393] The seventh pending result of the third to-be-detected image is obtained according to the fourth intermediate pending result and the fifth intermediate pending result.

[0394] If the third intermediate tentative result indicates that the third to-be-detected image is not of the target type, and the fourth intermediate tentative result and the fifth intermediate tentative result both indicate that the third to-be-detected image is of the target type, the seventh tentative result indicates that the third to-be-detected image is of the target type.

[0395] If the third intermediate tentative result indicates that the third to-be-detected image is not of the target type, and the fourth intermediate tentative result and the fifth intermediate tentative result both indicate that the third to-be-detected image is of the target type, the seventh tentative result indicates that the third to-be-detected image is of the target type.

[0396] As a possible implementation, the apparatus further includes an application unit configured to:

[0397] After determining the second intermediate tentative result of the third to-be-detected image according to the plurality of sixth probabilities, if the second intermediate tentative result indicates that the third to-be-detected image is of the target type, the third to-be-detected image is cropped based on a center of an application interface of the target application to obtain a third to-be-detected local image.

[0398] The third to-be-detected local image is detected by a local image special classification model to obtain an eighth tentative result of the third to-be-detected image, the local image special classification model is used to identify other behavior categories in the plurality of behavior categories included in the target type except the first behavior category, and the local image special classification model is trained based on image samples with a resolution less than or equal to a resolution threshold.

[0399] If the eighth tentative result indicates that the third to-be-detected image is of the target type, a third target detection result of the third to-be-detected image is determined to be of the target type, and an account corresponding to the third to-be-detected image is processed.

[0400] As a possible implementation, the local image special classification model includes a first local image special classification model and a second local image special classification model, the first local image special classification model is used to identify a second behavior category in the plurality of behavior categories included in the target type, the second local image special classification model is used to identify a third behavior category in the plurality of behavior categories included in the target type, and the apparatus further includes an application unit configured to:

[0401] The third to-be-detected local image is detected by the first local image special classification model to obtain a sixth intermediate tentative result of the third to-be-detected image.

[0402] The second local image-specific classification model is used to detect the third to-be-detected local image, to obtain a seventh intermediate to-be-determined result of the third to-be-detected image.

[0403] If the sixth intermediate to-be-determined result indicates that the third to-be-detected image belongs to the second behavior category, or the seventh intermediate to-be-determined result indicates that the third to-be-detected image belongs to the third behavior category, it is determined that the eighth to-be-determined result is used to indicate that the third to-be-detected image belongs to the target type.

[0404] As a possible implementation manner, the to-be-detected image is an image obtained based on a pre-filtering classification model, and the pre-filtering classification model is used to filter a to-be-determined to-be-detected image.

[0405] If the probability that the to-be-determined to-be-detected image belongs to the target type is greater than a third probability threshold, the to-be-determined to-be-detected image is taken as the to-be-detected image; if the probability that the to-be-determined to-be-detected image belongs to the target type is less than or equal to the third probability threshold, the to-be-determined to-be-detected image is deleted, and the to-be-detected image is the first to-be-detected image, the second to-be-detected image, or the third to-be-detected image.

[0406] The embodiments of the present application also provide a computer device, which can be a server or a terminal device. In the following, the computer device provided by the embodiments of the present application is introduced from the perspective of hardware materialization. The computer device can be a server or a terminal device. Figure 23 FIG. 1 shows a structural schematic diagram of a server, Figure 24 FIG. 2 shows a structural schematic diagram of a terminal device.

[0407] Referring to FIG. 1, Figure 23 FIG. 1 is a structural schematic diagram of a server provided by the embodiments of the present application. The server 1400 can have great differences due to different configurations or performances, and can include one or more processors 1422, such as a central processing unit (CPU), a memory 1432, one or more storage media 1430 (for example, one or more mass storage devices) of an application program 1442 or data 1444. The memory 1432 and the storage medium 1430 can be temporary storage or persistent storage. The program stored in the storage medium 1430 can include one or more modules (not shown in the figure), and each module can include a series of instruction operations in the server. Further, the processor 1422 can be configured to communicate with the storage medium 1430 and execute a series of instruction operations in the storage medium 1430 on the server 1400.

[0408] The server 1400 can also include one or more power supplies 1426, one or more wired or wireless network interfaces 1450, one or more input / output interfaces 1458, and / or one or more operating systems 1441, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM , and the like.

[0409] The steps performed by the server in the above embodiments can be based on the server structure shown in FIG. 14. Figure 23

[0410] The processor 1422 is configured to perform the following steps:

[0411] obtain a first to-be-detected image and a scene identifier, the first to-be-detected image being an image obtained by taking a screenshot during running of a target application, and the scene identifier being used to identify a detection scene type corresponding to the target application;

[0412] detect the first to-be-detected image by using a plurality of multi-type general classification models respectively, to obtain a first probability that the first to-be-detected image is of a target type output by each of the multi-type general classification models, and determine a first tentative result of the first to-be-detected image according to a plurality of the first probabilities, the plurality of multi-type general classification models being different models;

[0413] determine a target scene supplement model suitable for the target application from a plurality of scene supplement models according to the scene identifier, and detect the first to-be-detected image by using the target scene supplement model to obtain a second tentative result, different scene supplement models being suitable for different detection scene types;

[0414] determine a first target detection result of the first to-be-detected image according to the first tentative result and the second tentative result, the first target detection result being used to identify whether a type of the first to-be-detected image is the target type.

[0415] Optionally, the processor 1422 can also perform the method steps of any of the specific implementation manners of the data detection method in the embodiments of the present application.

[0416] Referring to Figure 24 , FIG. 14 is a structural schematic diagram of a terminal device provided by an embodiment of the present application. The terminal device is taken as an example of a smart phone for description, Figure 24 ​A block diagram of a partial structure of the smartphone is shown, which includes a Radio Frequency (RF) circuit 1510, a memory 1520, an input unit 1530, a display unit 1540, a sensor 1550, an audio circuit 1560, a Wireless Fidelity (WiFi) module 1570, a processor 1580, and a power supply 1590, etc. Those skilled in the art can understand that Figure 24 The structure of the smartphone shown in the figure is not intended to limit the smartphone, and the smartphone can include more or less components than the figure, or combine some components, or different arrangement of components.

[0417] The following will be described in detail Figure 24 The various components of the smartphone will be described in detail:

[0418] The RF circuit 1510 can be used for receiving and sending signals in the process of receiving or calling information, and in particular, receiving the downlink information of the base station and processing it by the processor 1580; in addition, sending the uplink data to the base station.

[0419] The memory 1520 can be used to store software programs and modules, and the processor 1580 can realize various functions of the smartphone and data processing by running the software programs and modules stored in the memory 1520.

[0420] The input unit 1530 can be used to receive input digital or character information, and generate key signal input related to user settings and function control of the smartphone. Specifically, the input unit 1530 can include a touch panel 1531 and other input devices 1532. The touch panel 1531, also known as a touch screen, can collect user touch operations on or near it and drive the corresponding connection device according to the pre-set program. In addition to the touch panel 1531, the input unit 1530 can also include other input devices 1532. Specifically, the other input devices 1532 can include one or more of a physical keyboard, function keys (such as volume control buttons, on-off buttons, etc.), trackballs, mice, joysticks, etc.

[0421] The display unit 1540 can be used to display information input by the user or information provided to the user and various menus of the smartphone. The display unit 1540 can include a display panel 1541, which can be configured in the form of a Liquid Crystal Display (LCD), an Organic Light-Emitting Diode (OLED), etc.

[0422] The smart phone can further include at least one sensor 1550, such as a light sensor, a motion sensor, and other sensors. As to other sensors that can be configured to the smart phone, such as a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, and the like, no further description is provided herein.

[0423] The audio circuit 1560, the speaker 1561, and the microphone 1562 can provide an audio interface between the user and the smart phone. The audio circuit 1560 can convert received audio data into an electrical signal, transmit the electrical signal to the speaker 1561, and convert the electrical signal into a sound signal output by the speaker 1561; on the other hand, the microphone 1562 can convert a sound signal collected into an electrical signal, and the audio circuit 1560 can convert the electrical signal into audio data, output the audio data to the processor 1580 for processing, and transmit the audio data to another smart phone via the RF circuit 1510, or output the audio data to the memory 1520 for further processing.

[0424] The processor 1580 is a control center of the smart phone, and connects various parts of the smart phone via various interfaces and lines, executes various functions of the smart phone and processes data by running or executing software programs and / or modules stored in the memory 1520 and calling data stored in the memory 1520. Optionally, the processor 1580 can include one or more processing units.

[0425] The smart phone further includes a power supply 1590 (such as a battery) for supplying power to various components. Preferably, the power supply can be logically connected to the processor 1580 via a power management system, so as to realize functions such as charge management, discharge management, and power consumption management via the power management system.

[0426] Although not shown, the smart phone can further include a camera, a Bluetooth module, and the like, and no further description is provided herein.

[0427] In the embodiments of the present application, the memory 1520 included in the smart phone can store a computer program and transmit the computer program to the processor.

[0428] The processor 1580 included in the smart phone can execute the data detection method provided in the above embodiments according to instructions in the computer program.

[0429] The embodiments of the present application further provide a computer readable storage medium for storing a computer program, and the computer program is used to execute the data detection method provided in the above embodiments.

[0430] The embodiment of the present application further provides a computer program product or computer program, which comprises computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the data detection method provided in various optional implementation manners of the above aspect.

[0431] The terms "first", "second", "third", "fourth", "fifth", "sixth", "seventh", "eighth" and the like, if any, used in the specification and claims, and above description, are used to distinguish similar objects, and are not necessarily used to describe a particular order or sequence. It should be understood that the data thus used can be interchanged, where appropriate, so that the embodiments of the present application described herein can be carried out in other than the order shown or described herein. Furthermore, the terms "comprise" and "comprising" and their any variations, are intended to cover non-exclusive inclusion, for example, processes, methods, systems, products, or devices that comprise a list of steps or units not necessarily limited to those clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products, or devices.

[0432] It can be understood by those skilled in the art that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, and the foregoing program can be stored in a computer readable storage medium, and the program executes the steps of the above-mentioned method embodiments when executed; and the foregoing storage medium can be at least one of the following media: read-only memory (English: Read-Only Memory, abbreviation: ROM), RAM, magnetic disc or optical disc and various computer program storage media.

[0433] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as processing circuit or memory) or combination thereof. Similarly, one processor (or multiple processors or memory) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an integral module or unit that includes the function of the module or unit.

[0434] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0435] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data detection method, characterized in that, The method includes: A first image to be detected and a scene identifier are obtained. The first image to be detected is an image captured during the operation of the target application. The scene identifier is used to identify the detection scene type corresponding to the target application. The first image to be detected is detected by multiple multi-type general classification models respectively, and the first probability of the first image to be detected being the target type is obtained by the output of each of the multi-type general classification models respectively. Based on the multiple first probabilities, the first undetermined result of the first image to be detected is determined, and the multiple multi-type general classification models are different models. Based on the scene identifier, a target scene supplementary model suitable for the target application is determined from multiple scene supplementary models. The first image to be detected is detected by the target scene supplementary model to obtain a second pending result. Different scene supplementary models are applicable to different detection scene types. Based on the first pending result and the second pending result, a first target detection result of the first image to be detected is determined, wherein the first target detection result is used to identify whether the type of the first image to be detected is the target type.

2. The method according to claim 1, characterized in that, If the scene identifier indicates that the detection scene type corresponding to the target application is a cold start scene, then the target scene supplementary model is a local graph general classification model, and the method further includes: Multiple image samples are acquired, each image sample having a label to identify its type; the multiple image samples may have various types. Each image sample is cropped according to its label to obtain multiple local image samples; Each of the local image samples is detected using an initial local image general classification model to obtain the first detection result for each of the image samples; Based on the differences between each of the first detection results and the labels of the corresponding image samples, the model parameters of the initial local graph general classification model are adjusted to obtain the local graph general classification model.

3. The method according to claim 2, characterized in that, The step of determining the first target detection result of the first image to be detected based on the first pending result and the second pending result includes: If the first pending result identifies the first image to be detected as the target type, or the second pending result identifies the first image to be detected as the target type, then the first target detection result of the first image to be detected is determined to identify the first image to be detected as the target type.

4. The method according to claim 2, characterized in that, The method further includes: Obtain the verification result for the first image to be detected, wherein the accuracy of the verification result is greater than the accuracy threshold; If the verification result is different from the first target detection result, the first image to be detected is detected by a target multi-type general classification model to obtain a second probability that the first image to be detected is the target type, wherein the target multi-type general classification model is one of a plurality of multi-type general classification models; Based on the second probability, determine the second detection result of the first image to be detected; Based on the difference between the second detection result and the verification result, the model parameters of the target multi-type general classification model are adjusted to obtain a multi-type specific classification model; Replace the general classification type with the specific classification model.

5. The method according to claim 4, characterized in that, After replacing the multi-type general classification with the multi-type specialized classification model, the method further includes: Each of the multiple multi-type general classification models is used as the target multi-type general classification model to obtain the multi-type specific classification model corresponding to each of the multi-type general classification models. Acquire a second image to be detected, which is an image captured during the operation of the target application; The second image to be detected is detected by multiple specialized classification models of the various types, and the third probability that the second image to be detected is the target type is obtained by the output of each specialized classification model of the various types. The second target detection result of the second image to be detected is determined based on multiple third probabilities.

6. The method according to claim 5, characterized in that, Determining the second target detection result of the second image to be detected based on a plurality of the third probabilities includes: Based on multiple third probabilities, a third pending result is determined for the second image to be detected; The second image to be detected is detected by the local image general classification model to obtain the fourth undetermined result of the second image to be detected; Based on the third and fourth pending results, the second target detection result of the second image to be detected is determined.

7. The method according to claim 1, characterized in that, If the scene identifier indicates that the detection scene type corresponding to the target application is a real-time processing scene, then the target scene supplementary model includes an easily identifiable general classification model and a text recognition model. The easily identifiable general classification model is used to identify the first behavior category among the multiple behavior categories included in the target type, and the identification difficulty of the first behavior category is less than the difficulty threshold. The step of detecting the first image to be detected using the target scene supplementary model to obtain a second undetermined result includes: The first image to be detected is detected using the easily identifiable general classification model to obtain a first intermediate pending result; If the first intermediate pending result identifies the first image to be detected as the first behavior category, then the first image to be detected is subjected to text recognition by the text recognition model to obtain the first text recognition result; If the first character recognition result is the same as the keyword in the keyword set, then the second pending result is determined to identify the first image to be detected as the target type, and the keywords included in the keyword set are related to the first behavior category; If the first character recognition result is different from the keywords in the keyword set, then the second pending result is determined to identify that the first image to be detected is not the target type.

8. The method according to claim 7, characterized in that, The method further includes: Obtain the verification result for the first image to be detected, wherein the accuracy of the verification result is greater than the accuracy threshold; If the verification result is different from the first target detection result, the first image to be detected is detected by the target multi-type general classification model to obtain the fourth probability that the first image to be detected is the target type, and the first image to be detected is detected by the easily identifiable general classification model to obtain the fifth probability that the first image to be detected is the target type. The target multi-type general classification model is one of the multiple multi-type general classification models. Based on the fourth probability, the third detection result of the first image to be detected is determined. Based on the difference between the third detection result and the verification result, the model parameters of the target multi-type general classification model are adjusted to obtain a multi-type special classification model. The multi-type special classification model is then used to replace the multi-type general classification model. Based on the fifth probability, the fourth detection result of the first image to be detected is determined. Based on the difference between the fourth detection result and the verification result, the model parameters of the easily identifiable general classification model are adjusted to obtain an easily identifiable special classification model. The easily identifiable special classification model is then used to replace the easily identifiable general classification model.

9. The method according to claim 8, characterized in that, The method further includes: A third image to be detected is obtained, which is an image captured during the operation of the target application. The third image to be detected is detected by multiple specialized classification models of the same type, and the sixth probability that the third image to be detected is the target type is obtained by the output of each specialized classification model of the same type. Based on multiple sixth probabilities, a second intermediate pending result is determined for the third image to be detected; If the second intermediate pending result identifies the third image to be detected as the target type, then the text recognition model is used to perform text recognition on the third image to be detected to obtain a second text recognition result; if the second text recognition result is the same as the keyword in the keyword set, then a fifth pending result is determined to identify the third image to be detected as the target type; if the second text recognition result is not the same as the keyword in the keyword set, then the fifth pending result is determined to identify the third image to be detected as not being the target type. The third image to be detected is detected using the easily identifiable dedicated classification model to obtain a third intermediate pending result; If the third intermediate pending result identifies the third image to be detected as belonging to the first behavior category, then the text recognition model is used to perform text recognition on the third image to be detected to obtain a third text recognition result; if the third text recognition result is the same as a keyword in the keyword set, then a sixth pending result is determined to identify the third image to be detected as the target type; if the third text recognition result is not the same as a keyword in the keyword set, then the sixth pending result is determined to identify the third image to be detected as not belonging to the target type. If the fifth pending result identifies the third image to be detected as the target type, and the sixth pending result identifies the third image to be detected as the target type, then the third target detection result of the third image to be detected is determined to identify the third image to be detected as the target type, and the account corresponding to the third image to be detected is processed.

10. The method according to claim 9, characterized in that, The method further includes: If the fifth pending result indicates that the third image to be detected is not the target type, and the sixth pending result indicates that the third image to be detected is not the target type, then the third target detection result of the third image to be detected is determined to be used to indicate that the third image to be detected is not the target type; If the fifth pending result identifies the third image to be detected as the target type, and the sixth pending result identifies the third image to be detected as not being the target type, then the third target detection result of the third image to be detected is determined to be used to identify the third image to be detected as a type to be verified. If the fifth pending result indicates that the third image to be detected is not the target type, and the sixth pending result indicates that the third image to be detected is the target type, then the third target detection result of the third image to be detected is determined to identify the third image to be detected as the type to be verified.

11. The method according to claim 9, characterized in that, The method further includes: If the average of the multiple sixth probabilities is greater than or equal to the first probability threshold, then the second intermediate pending result is determined to identify the third image to be detected as the target type; If the average of the plurality of sixth probabilities is less than the first probability threshold, and the average of the plurality of sixth probabilities is greater than or equal to the second probability threshold, then the second intermediate pending result is determined to identify the third image to be detected as a type to be verified, and the second probability threshold is less than the first probability threshold. If the average of the multiple sixth probabilities is less than the second probability threshold, then the second intermediate pending result is determined to identify that the third image to be detected is not the target type.

12. The method according to claim 9, characterized in that, After determining the second intermediate pending result of the third image to be detected based on a plurality of the sixth probabilities, the method further includes: If the second intermediate pending result identifies the third image to be detected as the target type, then the third image to be detected is detected by the whole image classification model to obtain the seventh pending result of the third image to be detected. The whole image classification model is used to identify other behavior categories besides the first behavior category among the multiple behavior categories included in the target type. The whole image classification model is trained based on image samples with a resolution greater than the resolution threshold. If the seventh pending result identifies the third image to be detected as the target type, and the third intermediate pending result identifies the third image to be detected as the target type, then the third target detection result of the third image to be detected is determined to be the target type, and the account corresponding to the third image to be detected is processed.

13. The method according to claim 12, characterized in that, The whole image classification model includes a first whole image classification model and a second whole image classification model. The first whole image classification model is used to identify a second behavior category among multiple behavior categories included in the target type, and the second whole image classification model is used to identify a third behavior category among multiple behavior categories included in the target type. The step of detecting the third image to be detected using the whole image classification model to obtain the seventh pending result of the third image to be detected includes: The third image to be detected is detected by the first whole image classification model to obtain the fourth intermediate undetermined result of the third image to be detected; The third image to be detected is detected by the second whole image classification model to obtain the fifth intermediate undetermined result of the third image to be detected; Based on the fourth and fifth intermediate pending results, the seventh pending result of the third image to be detected is obtained; Wherein, if the third intermediate pending result identifies the third image to be detected as the target type, and either the fourth intermediate pending result or the fifth intermediate pending result identifies the third image to be detected as the target type, then the seventh pending result identifies the third image to be detected as the target type; If the third intermediate pending result indicates that the third image to be detected is not the target type, and both the fourth intermediate pending result and the fifth intermediate pending result indicate that the third image to be detected is the target type, then the seventh pending result indicates that the third image to be detected is the target type.

14. The method according to claim 9, characterized in that, After determining the second intermediate pending result of the third image to be detected based on a plurality of the sixth probabilities, the method further includes: If the second intermediate pending result identifies the third image to be detected as the target type, then based on the center of the application interface of the target application, the third image to be detected is cropped to obtain the third local image to be detected. The third local image to be detected is detected by a local image-specific classification model to obtain the eighth pending result of the third local image to be detected. The local image-specific classification model is used to identify other behavior categories besides the first behavior category among the multiple behavior categories included in the target type. The local image-specific classification model is trained based on image samples with a resolution less than or equal to a resolution threshold. If the eighth pending result identifies the third image to be detected as the target type, then the third target detection result of the third image to be detected is determined to be the target type, and the account corresponding to the third image to be detected is processed.

15. The method according to claim 14, characterized in that, The local image-specific classification model includes a first local image-specific classification model and a second local image-specific classification model. The first local image-specific classification model is used to identify a second behavior category among multiple behavior categories included in the target type, and the second local image-specific classification model is used to identify a third behavior category among multiple behavior categories included in the target type. The step of detecting the third image to be detected using the local image-specific classification model to obtain the eighth pending result of the third image to be detected includes: The third local image to be detected is detected by the first local image-specific classification model to obtain the sixth intermediate pending result of the third local image to be detected. The third local image to be detected is detected by the second local image-specific classification model to obtain the seventh intermediate pending result of the third local image to be detected; If the sixth intermediate pending result identifies the third image to be detected as the second behavior category, or the seventh intermediate pending result identifies the third image to be detected as the third behavior category, then the eighth pending result is determined to identify the third image to be detected as the target type.

16. The method according to any one of claims 1-15, characterized in that, The image to be detected is an image obtained based on a pre-filtering classification model, which is used to filter the image to be detected. Wherein, if the probability that the image to be detected is of the target type is greater than the third probability threshold, then the image to be detected is used as the image to be detected; if the probability that the image to be detected is of the target type is less than or equal to the third probability threshold, then the image to be detected is deleted, and the image to be detected is the first image to be detected, the second image to be detected, or the third image to be detected.

17. A data detection device, characterized in that, The device includes: an acquisition unit, a first detection unit, a second detection unit, and a determination unit; The acquisition unit is used to acquire a first image to be detected and a scene identifier. The first image to be detected is an image obtained by taking a screenshot during the operation of the target application. The scene identifier is used to identify the detection scene type corresponding to the target application. The first detection unit is configured to detect the first image to be detected by multiple multi-type general classification models respectively, obtain the first probability that the first image to be detected is the target type output by each of the multi-type general classification models, and determine the first pending result of the first image to be detected based on the multiple first probabilities, wherein the multiple multi-type general classification models are different models; The second detection unit is used to determine a target scene supplementary model suitable for the target application from multiple scene supplementary models based on the scene identifier, and to detect the first image to be detected by the target scene supplementary model to obtain a second pending result. Different scene supplementary models are applicable to different detection scene types. The determining unit is configured to determine a first target detection result of the first image to be detected based on the first pending result and the second pending result, wherein the first target detection result is used to identify whether the type of the first image to be detected is the target type.

18. A computer device, characterized in that, The computer device includes a processor and memory: The memory is used to store computer programs and to transfer the computer programs to the processor; The processor is configured to perform the method according to any one of claims 1-16 according to the computer program.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method according to any one of claims 1-16.

20. A computer program product comprising a computer program, characterized in that, When it is run on a computer device, it causes the computer device to perform the method described in any one of claims 1-16.

Images