Fragment recognition methods, devices, equipment, storage media, and software products

Abstract

This application provides a fragment recognition method, apparatus, device, storage medium, and program product, applicable to various fragment recognition scenarios such as cloud technology, artificial intelligence, smart transportation, and vehicle-mounted systems. The fragment recognition method includes: acquiring game scene information and at least one game event for a virtual object in a game fragment based on game data from a virtual scene, wherein the game event is determined based on multiple atomic game actions; acquiring game scene features corresponding to the game scene information, and acquiring at least one game event feature corresponding to at least one game event; determining a target deviation probability of at least one game event from the game scene information based on the matching features of the game scene features and the at least one game event feature; and determining whether the game fragment to be recognized is abnormal based on the comparison result of the target deviation probability and a probability threshold. This application can improve the accuracy of fragment recognition.

Description

Technical Field

[0001] This application relates to intelligent recognition technology in the field of artificial intelligence, and more particularly to a fragment recognition method, apparatus, device, storage medium, and program product. Background Technology

[0002] Currently, virtual matches based on virtual scenarios are widely used; however, abnormal applications that modify the data and match logic of virtual match applications have also emerged, which are referred to here as cheat applications. These cheat applications enable abnormal operations in virtual matches within virtual scenarios, posing security risks to virtual matches in virtual scenarios.

[0003] Generally, to identify abnormal actions in virtual games in a virtual setting, the usual method is to directly compare the received atomic game actions with normal atomic game actions; this is a data-level identification process with relatively low accuracy. Summary of the Invention

[0004] This application provides a method, apparatus, device, computer-readable storage medium, and computer program product for segment recognition based on virtual scenes, which can improve the recognition accuracy of game segments in virtual scenes.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides a fragment recognition method based on a virtual scene, including:

[0007] Based on the game data of the virtual scene, obtain the game scene information and at least one game event of the virtual object to be identified in the game segment to be identified;

[0008] Obtain the game scene features corresponding to the game scene information, and obtain at least one game event feature corresponding to at least one game event;

[0009] Based on the matching features of the game scene features and at least one of the game event features, determine the target deviation probability of at least one of the game events deviating from the game scene information;

[0010] Based on the comparison between the target deviation probability and the probability threshold, the identification result of whether the game segment to be identified is abnormal is determined.

[0011] This application provides a fragment recognition device based on a virtual scene, comprising:

[0012] The data acquisition module is used to acquire, based on the game data of the virtual scene, the game scene information of the virtual object to be identified in the game segment to be identified and at least one game event, wherein the game event is determined based on multiple atomic game actions;

[0013] The feature extraction module is used to obtain game scene features corresponding to the game scene information, and to obtain at least one game event feature corresponding to at least one game event;

[0014] The probability prediction module is used to determine the target deviation probability of at least one of the game events deviating from the game scene information based on the matching features of the game scene features and at least one of the game event features;

[0015] The anomaly detection module is used to determine whether the game segment to be identified is abnormal based on the comparison result between the target deviation probability and the probability threshold.

[0016] In this embodiment of the application, the game scene information includes at least one of global game scene information, local game scene information, and associated target game scene information;

[0017] Wherein, the global game scene information refers to the game information obtained based on the vision of the virtual object to be identified, the local game scene information refers to the information of the virtual object to be identified when it is playing a virtual game, and the associated target game scene information refers to the game information between the virtual object to be identified and the virtual object associated with the virtual object to be identified.

[0018] In this embodiment of the application, when the game scene information includes global game scene information, local game scene information, and associated target game scene information, the feature extraction module is further configured to obtain the intent feature corresponding to the global game scene information; obtain the action feature corresponding to the local game scene information; obtain the associated target feature corresponding to the associated target game scene information; and combine the intent feature, the action feature, and the associated target feature into the game scene feature.

[0019] In this embodiment of the application, the feature extraction module is further configured to extract game statistics from the game data, wherein the game statistics represent the statistical information of the virtual object to be identified in the virtual game process; and obtain auxiliary features corresponding to the game statistics.

[0020] In this embodiment of the application, the probability prediction module is further configured to concatenate the matching features of the game scene features and at least one of the game event features, as well as the auxiliary features, to obtain the features to be identified; and based on the features to be identified, to determine the target deviation probability of at least one of the game events deviating from the game scene information.

[0021] In this embodiment, the game event is determined based on multiple atomic game actions. The data acquisition module is further configured to extract whole game scene information and a set of atomic game actions from the game data of the virtual scene; determine the game scene information of the virtual object to be identified in the game segment to be identified from the whole game scene information, and determine the atomic game action sequence of the virtual object to be identified in the game segment to be identified from the set of atomic game actions; and determine at least one game event corresponding to the atomic game action sequence based on a preset relationship between events and atomic game actions, wherein the preset relationship is a one-to-many relationship between events and atomic game actions.

[0022] In this embodiment, the target deviation probability is obtained through a recognition model. The fragment recognition device further includes a model training module for acquiring target training data, wherein the target training data includes training samples and deviation labels. The training samples include game scene samples and at least one game event sample. The deviation label indicates whether at least one game event sample deviates from the game scene sample. Based on the recognition model to be trained, the features of the game scene sample and the features of at least one game event sample are obtained as first matching feature samples, and a first predicted deviation probability is determined based on the first matching feature samples. The recognition model to be trained is a neural network model to be trained for acquiring the deviation probability. Based on the loss between the first predicted deviation probability and the deviation label, the recognition model to be trained is trained to obtain the recognition model.

[0023] In this embodiment of the application, the model training module is further configured to generate at least one first game generation event corresponding to the first matching feature sample based on the recognition model to be trained; and to determine a first unsupervised loss based on at least one game event sample and at least one first game generation event.

[0024] In this embodiment of the application, the model training module is further configured to train the recognition model to be trained based on the loss between the first predicted deviation probability and the deviation label, and the combined loss of the first unsupervised loss, to obtain the recognition model.

[0025] In this embodiment, the model training module is further configured to acquire initial training data, wherein the initial training data includes initial game scene samples, at least one initial game event sample, and a first deviation label; based on the model to be trained, acquire second matching feature samples corresponding to the features of the initial game scene samples and the features of at least one initial game event sample, and generate at least one second game generation event corresponding to the second matching feature samples; determine a preset mismatch label as a second deviation label for the initial game scene samples and at least one second game generation event; and determine the combination of the initial game scene samples, at least one initial game event sample, and the first deviation label, and the combination of the initial game scene samples, at least one second game generation event, and the second deviation label as the target training data.

[0026] In this embodiment, the model training module is further configured to determine a second unsupervised loss based on the difference between at least one initial game event sample and at least one second game generation event; and to train the training generation model by combining the second unsupervised loss and the second predicted deviation probability; wherein the second predicted deviation probability refers to the probability that at least one second game generation event deviates from the initial game scene sample, and the second predicted deviation probability is obtained through the training recognition model; the trained training generation model is used to enhance the training data of the trained training recognition model.

[0027] In this embodiment, when at least one initial game event sample includes multiple initial game event samples, the model training module is further configured to: obtain the true deviation label corresponding to each initial game event sample; determine the number of samples for each deviation label combination based on the first deviation label and the true deviation label corresponding to the initial game event sample, and in combination with the deviation label category, wherein the deviation label category includes two types: deviation category and matching category, and the deviation label combination refers to a combination of the first deviation label and the true deviation label respectively in any one of the deviation label categories; determine the joint probability distribution of the first deviation label and the true deviation label based on the number of samples and the total number of initial game event samples; filter out the initial game event samples that are lower than the label probability threshold in the joint probability distribution from the multiple initial game event samples to obtain at least one target game event sample; and combine the initial game scene sample, at least one target game event sample, and the first deviation label into the target training data.

[0028] In this embodiment of the application, the model training module is further configured to obtain the category prediction probability of each initial game event sample corresponding to each deviation label category; determine the confidence probability corresponding to each deviation label category based on the category prediction probability; and determine the deviation label category corresponding to the largest category prediction probability greater than the confidence probability as the true deviation label of the initial game event sample.

[0029] This application provides a fragment recognition device based on a virtual scene, comprising:

[0030] Memory, used to store executable instructions;

[0031] The processor, when executing executable instructions stored in the memory, implements the virtual scene-based fragment recognition method provided in the embodiments of this application.

[0032] This application provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the virtual scene-based fragment recognition method provided in this application.

[0033] This application provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the virtual scene-based fragment recognition method provided in this application.

[0034] The embodiments of this application have at least the following beneficial effects: In the process of identifying the abnormality of the game segment to be identified in the virtual scene, the game scene information is extracted from the game data, and at least one game event corresponding to the game scene information is determined, and the deviation probability of at least one game event from the target of the game scene information is judged; since the game event is an abstraction of the game decision and represents the intention of the virtual game, by judging the matching of the game scene features and the features of at least one game event, it is possible to accurately determine whether the game scene information supports at least one game event, and thus accurately obtain the identification result of whether the game segment to be identified is abnormal. Therefore, the identification accuracy of game segments in the virtual scene can be improved. Attached Figure Description

[0035] Figure 1 This is an exemplary diagram illustrating an anomaly distribution provided in an embodiment of this application;

[0036] Figure 2 This is a schematic diagram of the architecture of the fragment recognition system provided in the embodiments of this application;

[0037] Figure 3 This is provided by the embodiments of this application. Figure 2 A schematic diagram of the component structure of a server in a computer.

[0038] Figure 4 This is a flowchart illustrating the fragment recognition method provided in an embodiment of this application;

[0039] Figure 5 This is an exemplary schematic diagram of a game scene provided in an embodiment of this application;

[0040] Figure 6 This is another exemplary schematic diagram of a game scene provided in an embodiment of this application;

[0041] Figure 7 This is a schematic diagram of the training process of the recognition model provided in the embodiments of this application;

[0042] Figure 8 This is an exemplary system architecture diagram provided in an embodiment of this application;

[0043] Figure 9 This is an exemplary model training diagram provided in an embodiment of this application;

[0044] Figure 10 This is a schematic diagram of an exemplary noise reduction process provided in an embodiment of this application;

[0045] Figure 11 This is a schematic diagram of an exemplary generator provided in an embodiment of this application;

[0046] Figure 12 This is a schematic diagram of the structure of an exemplary discriminator provided in an embodiment of this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0049] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0050] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0051] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0052] 1) Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0053] 2) Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills; and how to reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning typically includes techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, and inductive learning.

[0054] 3) Artificial neural networks are mathematical models that mimic the structure and function of biological neural networks. Exemplary structures of artificial neural networks in this application include Generative Adversarial Networks (GANs), Graph Convolutional Networks (GCNs), Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Neural State Machines (NSMs), and Phase-Functioned Neural Networks (PFNNs). The training model, recognition model, and training generative model involved in this application are the corresponding models of artificial neural networks.

[0055] 4) A virtual scene refers to a virtual scene displayed (or provided) by an application running on a terminal device. It can also refer to a virtual scene played by receiving audio and video information sent by a cloud server, where the application runs on a cloud server. Furthermore, a virtual scene can be a simulation of the real world, a semi-simulated / semi-fictional virtual environment, or a purely fictional virtual environment; and it can be any of a two-dimensional, 2.5-dimensional, or three-dimensional virtual scene. This application embodiment does not limit the dimension of the virtual scene. For example, a virtual scene may include a virtual sky, virtual land, and virtual ocean, etc. The virtual land may include environmental elements such as virtual deserts and virtual cities. Users can control virtual objects to move within the virtual scene. In this application embodiment, the virtual scene is used for virtual game sessions.

[0056] 5) Virtual objects: Images of various virtual people and virtual objects that can interact with within a virtual scene, or movable virtual objects within a virtual scene. These movable virtual objects can be virtual characters, virtual animals, or anime characters, such as virtual characters and virtual animals displayed in a virtual scene. The virtual object can be a virtual avatar representing the user within the virtual scene. A virtual scene can include multiple virtual objects, each with its own shape and volume, occupying a portion of the space within the virtual scene. In this embodiment, the virtual object to be identified can be either an abnormal virtual object or a normal virtual object.

[0057] 6) Cloud gaming, also known as gaming on demand, is an online gaming technology based on cloud computing. Cloud gaming technology enables thin clients with relatively limited graphics processing and data processing capabilities to run high-quality games. In a cloud gaming scenario, the game does not run on the player's gaming terminal but on a cloud server. The cloud server renders the game scene as an audio and video stream, which is then transmitted to the player's gaming terminal via the network. The player's gaming terminal does not need powerful graphics processing and data processing capabilities; it only needs basic streaming media playback capabilities and the ability to receive player input commands and send them to the cloud server. The virtual game provided in this application embodiment can be a game mechanism in a cloud gaming application.

[0058] It's important to note that virtual matches based on virtual environments (such as Multiplayer Online Battle Arena (MOBA) games in esports) are widely used. Users, as players, control virtual objects within a virtual environment to engage in virtual matches. However, abnormal applications that modify the data and match logic of these virtual match applications have emerged; these are referred to as cheat applications. Cheat applications enable abnormal operations in virtual matches, posing security risks and leading to high account churn rates, thus increasing resource consumption during application deployment. Furthermore, cheat applications consume system resources, slow down network speeds, and degrade application performance, such as rendering stuttering in virtual environments. Additionally, cheat applications may carry viruses or other malicious computer applications that illegally obtain account information, resulting in low system and information security.

[0059] Generally, esports games used to implement virtual matches are based on a client / server (C / S) model, employing the Transmission Control Protocol / Internet Protocol (TCP / IP) communication protocol; and data exchange is achieved through Internet Protocol data packet transmission. After a player's account logs into the game through the game client, the client sends movement, kill, and other operation commands to the server. The server receives these commands, processes them, and then sends them back to the client via the network, providing data such as movement coordinates and kill coordinates. To identify abnormal operations in virtual matches within a virtual environment, as shown in Table 1, detection typically falls into three categories: client-side detection, network-side detection, and server-side detection.

[0060] Table 1

[0061]

[0062]

[0063] See Table 1 for further details. Client-side detection includes CAPTCHA detection, which involves obtaining CAPTCHAs through the client. However, external applications can also perform CAPTCHA detection using image processing. Client-side detection also includes methods such as detecting the process of external applications. Due to limited coverage, its usability is low. Network-side detection includes traffic analysis, which involves analyzing the timing, data length, and service reception intervals of business network command packets (e.g., client commands, drift bursts at multiple time scales, and service sensitivity under various network conditions). Due to network fluctuations, its accuracy is low. Server-side detection includes user action analysis, path analysis, and observational evidence analysis. Among these, user action analysis involves directly comparing received atomic game actions with normal atomic game actions. This is a data-level identification process and cannot effectively simulate the characteristic data of virtual objects to detect abnormal behavior of external virtual objects, resulting in low identification accuracy. Path analysis refers to the detection method that identifies the differences between cheat-generated virtual objects and normal virtual objects. Since cheat-generated virtual objects use preset movement paths, while normal players have their own operational awareness, normal virtual objects correspond to various movement patterns. Therefore, detection can be achieved by statistically analyzing recurring path movement models, but this method is complex and has low feasibility. Observational analysis involves using automatic detection mechanisms to detect pop-up window events and obtaining input operations such as keyboard or mouse clicks. However, cheat applications can also achieve corresponding input operations by combining image processing, resulting in low detection accuracy.

[0064] It should also be noted that, see Figure 1 , Figure 1 This is an exemplary diagram illustrating anomaly distribution provided in an embodiment of this application; as shown... Figure 1 As shown, the distribution of methods corresponding to abnormal virtual games Figure 1-1 Among the various methods, perspective-based methods 1-11 accounted for the largest proportion (88%), followed by all-around perspective methods 1-12 (9%) and auto-aim methods 1-13 (3%). Therefore, the implementation of virtual matches differs from that of first-person shooter (FPS) games, and the detection methods used in FPS games cannot be applied to virtual match applications.

[0065] Based on this, embodiments of this application provide a method, apparatus, device, computer-readable storage medium, and computer program product for fragment recognition based on virtual scenes, which can improve the recognition accuracy of the game fragments to be identified. The following describes exemplary applications of the fragment recognition device based on virtual scenes (hereinafter referred to as the fragment recognition device) provided in embodiments of this application. The fragment recognition device provided in embodiments of this application can be implemented as various types of terminals such as smartphones, smartwatches, laptops, tablets, desktop computers, smart home appliances, set-top boxes, smart in-vehicle devices, portable music players, personal digital assistants, dedicated messaging devices, smart voice interaction devices, portable gaming devices, and smart speakers, or it can be implemented as a server. The following will describe exemplary applications when the fragment recognition device is implemented as a server.

[0066] See Figure 2 , Figure 2 This is a schematic diagram of the architecture of the fragment recognition system provided in the embodiments of this application; as shown Figure 2 As shown, to support a fragment recognition application based on a virtual scene, in the fragment recognition system 100, terminal 400 (terminals 400-1 and 400-2 are shown as examples) connects to server 200 (referred to as a fragment recognition device) via network 300. Network 300 can be a wide area network (WAN), a local area network (LAN), or a combination of both. Additionally, the fragment recognition system 100 also includes a database 500 for providing data support to server 200; and... Figure 2 The example shown illustrates a scenario where the database 500 is independent of the server 200. However, the database 500 can also be integrated into the server 200, and this embodiment does not limit this to any particular case.

[0067] Terminal 400 is used to display a virtual scene and, in response to virtual game operations for the virtual scene, interacts with server 200 via network 300 to display an animation of the virtual object to be identified playing a virtual game in the virtual scene in a graphical interface (graphical interfaces 410-1 and 410-2 are shown as examples).

[0068] Server 200 is used to interact with terminal 400 via network 300 to obtain game data of the virtual scene when terminal 400 responds to a virtual game operation for the virtual scene; it is also used to obtain game scene information and at least one game event of the virtual object to be identified in the game segment to be identified based on the game data of the virtual scene, wherein the game event is determined based on multiple atomic game actions; obtain game scene features corresponding to the game scene information, and obtain at least one game event feature corresponding to at least one game event; determine the target deviation probability of at least one game event from the game scene information based on the matching features of the game scene features and at least one game event feature; and determine the identification result of whether the game segment to be identified is abnormal based on the comparison result of the target deviation probability and the probability threshold.

[0069] In some embodiments, server 200 may be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Terminal 400 may be a smartphone, smartwatch, laptop, tablet, desktop computer, smart TV, set-top box, smart in-vehicle device, portable music player, personal digital assistant, dedicated messaging device, portable gaming device, and smart speaker, but is not limited to these. Terminals and servers can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.

[0070] See Figure 3 , Figure 3 This is provided by the embodiments of this application. Figure 2 A schematic diagram of the structure of a server in a computer. Figure 3 The server 200 shown includes at least one processor 210, memory 250, at least one network interface 220, and a user interface 230. The various components in server 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to implement communication between these components. In addition to a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 3 The general labeled all buses as Bus System 240.

[0071] Processor 210 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0072] User interface 230 includes one or more output devices 231 that enable the presentation of media content, including one or more speakers and / or one or more visual displays. User interface 230 also includes one or more input devices 232, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0073] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 250 may optionally include one or more storage devices physically located away from the processor 210.

[0074] The memory 250 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in this application embodiment is intended to include any suitable type of memory.

[0075] In some embodiments, memory 250 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0076] Operating system 251 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0077] The network communication module 252 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 including: Bluetooth, Wi-Fi, and Universal Serial Bus (USB), etc.

[0078] Presentation module 253 is configured to enable the presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 231 (e.g., a display screen, a speaker, etc.) associated with user interface 230;

[0079] The input processing module 254 is used to detect and translate one or more user inputs or interactions from one or more input devices 232.

[0080] In some embodiments, the virtual scene-based fragment recognition device (hereinafter referred to as the fragment recognition device) provided in this application can be implemented in software. Figure 3 A fragment recognition device 255 stored in memory 250 is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: a data acquisition module 2551, a feature extraction module 2552, a probability prediction module 2553, an anomaly recognition module 2554, and a model training module 2555. These modules are logically connected and can therefore be arbitrarily combined or further divided according to their implemented functions. The functions of each module will be described below.

[0081] In some embodiments, the fragment recognition device provided in this application can be implemented in hardware. As an example, the device provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the fragment recognition method based on a virtual scene provided in this application. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0082] In some embodiments, the terminal or server can implement the virtual scene-based fragment recognition method provided in this application by running a computer program. For example, the computer program can be a native program or software module in an operating system; it can be a native application (APP), i.e., a program that needs to be installed in the operating system to run, such as a plug-in detection APP; it can also be a mini-program, i.e., a program that only needs to be downloaded to a browser environment to run; or it can be a mini-program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module, or plug-in.

[0083] The following will describe the virtual scene-based segment recognition method provided in this application embodiment, with reference to exemplary applications and implementations of the segment recognition device provided in the embodiments of this application. Furthermore, the virtual scene-based segment recognition method provided in the embodiments of this application can be applied to various virtual scene-based segment recognition scenarios such as cloud technology, artificial intelligence, smart transportation, and vehicle-mounted systems.

[0084] See Figure 4 , Figure 4 This is a flowchart illustrating the fragment recognition method provided in the embodiments of this application, which will be combined with... Figure 4 The steps shown are explained.

[0085] Step 401: Based on the game data of the virtual scene, obtain the game scene information and at least one game event of the virtual object to be identified in the game segment to be identified.

[0086] In this embodiment, the segment recognition device determines whether a virtual object to be identified is abnormal by recognizing segments of virtual games played by the virtual object to be identified. Here, during the process of recognizing segments of virtual games played by the segment recognition device, it first acquires data records of the virtual game played by the virtual object to be identified, thus obtaining game data of the virtual scene, such as replay data used to reconstruct the virtual game operations performed by each virtual object during the virtual game. Next, the segment recognition device extracts game scene information of the virtual object to be identified in the game segment from the game data and obtains at least one game event corresponding to the game scene information.

[0087] It should be noted that the game scene information refers to the scene information of the virtual game, such as the status information of the virtual objects in the virtual game, virtual obstacles, virtual buildings in the virtual scene, and distance information (the distance between virtual objects and virtual objects, or between virtual objects, or between virtual objects), etc. A game event is determined based on multiple atomic game actions and represents the decision-making intent for conducting a virtual game. It can be a sub-virtual game event within a virtual game, such as entering a target area (virtual jungle area), virtually destroying virtual buildings, or following a target object. An atomic game action is a basic action in a virtual game; that is, an atomic game action is the smallest unit of action performed by a virtual object in a virtual scene, such as taking a step forward, turning around, or jumping. Furthermore, at least one game event refers to at least one decision made by the virtual object to be identified in response to game scene information. When at least one game event includes multiple game events, it forms a sequence of game events, meaning there is an order relationship between these multiple game events. This order relationship can correspond to one or more sequential branches. For example, if at least one game event includes game events 1 to 3, then game event 2 occurs after game event 1, game event 3 occurs after game event 3, or both game events 2 and 3 occur after game event 1.

[0088] For example, see Figure 5 , Figure 5 This is an exemplary schematic diagram of a game scene provided in an embodiment of this application; as shown... Figure 5 As shown, virtual scene 5-1 is a virtual game screen, including game scene information such as the status values ​​of various virtual objects (including our virtual objects and enemy virtual objects).

[0089] Step 402: Obtain the game scene features corresponding to the game scene information, and obtain at least one game event feature corresponding to at least one game event.

[0090] In this embodiment, after the fragment recognition device obtains the game scene information and at least one game event feature, it obtains the features of the game scene information, thus obtaining the game scene features, and obtains the features of the at least one game event, thus obtaining the at least one game event feature. Each at least one game event corresponds one-to-one with each at least one game event feature.

[0091] It should be noted that the game scene information includes at least one of the following: global game scene information, local game scene information, and associated target game scene information. Specifically, global game scene information refers to game information obtained based on the vision of the virtual object to be identified, such as the decision-making intent of the entire virtual game and field of view information; local game scene information refers to information about the virtual object to be identified during the virtual game, such as the virtual object's actions (virtual kills, moving forward, moving backward, etc.) and virtual health; associated target game scene information refers to game information between the virtual object to be identified and the virtual objects associated with it, such as the distance between the virtual object to be identified and the target virtual object.

[0092] For example, see Figure 6 , Figure 6 This is another exemplary schematic diagram of a game scene provided in an embodiment of this application; as shown... Figure 6 As shown, virtual scene 6-1 is the screen of a virtual game. Area 6-11 includes global game scene information, such as the position of our virtual object 6-111, the position of the enemy virtual object 6-112, the position of virtual buildings 6-113, etc. Area 6-12 includes local game scene information, such as the virtual state value of the virtual object to be identified 6-121. Area 6-12 also includes related target game scene information, such as the distance between the virtual object to be identified and the target virtual object 6-122.

[0093] In this embodiment of the application, when the game scene information includes global game scene information, local game scene information, and associated target game scene information, step 402, in which the segment recognition device acquires the game scene features corresponding to the game scene information, includes: the segment recognition device acquiring the intent features corresponding to the global game scene information; acquiring the action features corresponding to the local game scene information; and acquiring the associated target features corresponding to the associated target game scene information; finally, the intent features, action features, and associated target features are combined into game scene features. Here, intent features refer to the features of the global game scene information, action features refer to the features of the local game scene information, and associated target features refer to the features of the associated target game scene information. Here, the segment recognition device can determine at least one of the intent features, action features, and associated target features as game scene features.

[0094] Understandably, since game events represent decision-making intentions, acquiring at least one game event allows for the construction of a decision space, thus enabling the identification of abnormal virtual game operations such as perspective-based actions corresponding to decision-making intentions. Furthermore, when acquiring features, the fragment recognition device identifies at least one of the following as game scene features: intent features, action features, and associated target features. This process of determining game event features from at least one dimension of features enhances the comprehensiveness of game event features, thereby improving the accuracy of fragment recognition.

[0095] Step 403: Based on the matching features of the game scene features and at least one game event features, determine the target deviation probability of at least one game event deviating from the game scene information.

[0096] In this embodiment, after obtaining the local scene features and at least one game event feature, the fragment recognition device acquires a matching feature representing the degree of matching between the game scene information and at least one game event based on the game scene features and at least one game event feature. Then, based on this matching feature, it determines the target deviation probability of at least one game event deviating from the game scene information. The target deviation probability refers to the degree of deviation of at least one game event from the game scene information, and the degree of matching between the game scene information and at least one game event is negatively correlated with the target deviation probability.

[0097] It should be noted that the fragment recognition device can determine the matching feature by obtaining the difference between the game scene feature and at least one game event feature, or it can obtain the matching feature between the game scene feature and at least one game event feature based on preset mapping parameters, etc. The embodiments of this application do not limit this.

[0098] Step 404: Based on the comparison results of the target deviation probability and the probability threshold, determine the identification result of whether the game segment to be identified is abnormal.

[0099] In this embodiment, the fragment recognition device is equipped with a probability threshold, or the fragment recognition device can obtain the probability threshold from other devices or received information. The probability threshold represents the minimum probability of determining that at least one game event matches the game scene information. Thus, after obtaining the target deviation probability, the fragment recognition device compares the target deviation probability with the probability threshold. If the comparison result determines that the target deviation probability is less than the probability threshold, it indicates that at least one game event matches the game scene information, thereby indicating that the game fragment to be identified is normal. If the comparison result determines that the target deviation probability is greater than or equal to the probability threshold, it indicates that at least one game event deviates from the game scene information, thereby indicating that the game fragment to be identified is abnormal.

[0100] It should be noted that the identification results are used to determine whether the virtual object to be identified is abnormal; when the identification result indicates that the game segment to be identified is normal, it indicates that the virtual object to be identified is normal; when the identification result indicates that the game segment to be identified is abnormal, it indicates that the virtual object to be identified is abnormal.

[0101] Understandably, the process of identifying anomalies in game segments within a virtual scene involves extracting game scene information from game data, determining at least one game event corresponding to the game scene information, and judging the deviation probability of at least one game event from the target of the game scene information. Since game events are determined based on multiple atomic game actions, they are abstractions of game decisions and represent the intent of the virtual game. Therefore, by determining the matching between game scene features and at least one game event feature, it is possible to accurately determine whether the game scene information supports at least one game event, thus accurately obtaining the identification result of whether the game segment to be identified is abnormal. This improves the accuracy of segment identification in virtual scenes.

[0102] In this embodiment of the application, steps 405 and 406 (not shown in the figure) are included before step 403; that is, before the fragment recognition device determines the target deviation probability of at least one game event deviating from the game scene information based on the matching features of the game scene features and at least one game event features, the fragment recognition method based on the virtual scene further includes steps 405 and 406, and each step is described below.

[0103] Step 405: Extract game statistics from the game data.

[0104] In this embodiment of the application, the fragment recognition device extracts statistical information of the virtual object to be identified in the virtual game from the game data, thus obtaining game statistics information; such as whether it is in a virtual death state, the time without vision, the duration of low health, whether it returns to the destination, whether it is in the target virtual building, the number of virtual teammates, the number of virtual attacks without vision, etc.

[0105] Step 406: Obtain the auxiliary features corresponding to the game statistics.

[0106] In this embodiment, the fragment recognition device acquires features from the game statistics information, thus obtaining auxiliary features. Accordingly, in step S403, the fragment recognition device determines the target deviation probability of at least one game event from the game scene information based on the matching features of the game scene features and at least one game event feature. This includes: the fragment recognition device concatenates the matching features of the game scene features and at least one game event feature, as well as the auxiliary features, to obtain the feature to be recognized; and based on the feature to be recognized, determines the target deviation probability of at least one game event from the game scene information.

[0107] It should be noted that auxiliary features are used to enhance matching features. Thus, the fragment recognition device concatenates the auxiliary features and the matching features to obtain the feature to be recognized; that is, the feature to be recognized is the matching feature enhanced based on the auxiliary features.

[0108] It is understandable that by using auxiliary features corresponding to game statistics, the matching features of game scene features and at least one game event feature can be enhanced to obtain the features to be identified, so that the information represented by the features to be identified is comprehensive. Thus, the accuracy of the target deviation probability can be improved based on the features to be identified.

[0109] In this embodiment of the application, step 401 can be implemented by steps 4011 to 4013 (not shown in the figure); that is, the fragment recognition device obtains the game scene information of the virtual object to be identified in the game fragment to be identified and at least one game event based on the game data of the virtual scene, including steps 4011 to 4013. Each step is described below.

[0110] Step 4011: Extract the entire game scene information and the set of atomic game actions from the game data of the virtual scene.

[0111] In this embodiment of the application, the game data can be the data of virtual games conducted throughout the entire virtual game process. Thus, the segment recognition device can extract the scene information of virtual games conducted throughout the entire virtual game process from the game data, which is referred to here as the whole game scene information; it can also extract each atomic game action of virtual games conducted throughout the entire virtual game process, which is referred to here as the atomic game action set.

[0112] Step 4012: From the overall game scene information, determine the game scene information of the virtual object to be identified in the game segment to be identified, and from the atomic game action set, determine the atomic game action sequence of the virtual object to be identified in the game segment to be identified.

[0113] In this embodiment, after obtaining the overall game scene information and the set of atomic game actions, the fragment recognition device can extract the game scene information of the game fragment to be identified from the overall game scene information. Here, the game fragment to be identified can be any one of the preset game fragment sets, and each preset game fragment in the preset game fragment set has a higher probability of being abnormal compared to other game fragments in the virtual game process. In addition, the fragment recognition device is also used to obtain from the set of atomic game actions each atomic game action executed sequentially with the virtual object to be identified in the game fragment to be identified, which is referred to here as the atomic game action sequence.

[0114] Step 4013: Based on the preset relationship between events and atomic game actions, determine at least one game event corresponding to the atomic game action sequence.

[0115] In this embodiment, the fragment recognition device establishes a preset relationship between events and atomic game actions. Alternatively, the fragment recognition device can obtain the preset relationship between events and atomic game actions from other devices or received information. In this preset relationship, the relationship between events and atomic game actions is one-to-many, meaning that one event in the preset relationship corresponds to multiple atomic game actions. Furthermore, the game event is an event within the preset relationship and is therefore determined based on multiple atomic game actions. Here, the fragment recognition device sequentially determines the corresponding game events for multiple atomic game actions in the atomic game action sequence within the preset relationship. When the determination of the atomic game action sequence is completed, at least one game event is obtained.

[0116] Understandably, by identifying virtual fragments with a high probability of anomalies throughout the entire virtual game process, it is possible to determine whether the virtual objects to be identified are abnormal. This reduces the amount of data required for fragment identification, thereby improving the efficiency of fragment identification and reducing the resource consumption of fragment identification.

[0117] See Figure 7 , Figure 7 This is a schematic diagram of the recognition model training process provided in the embodiments of this application; as shown Figure 7 As shown in the embodiment of this application, the target deviation probability is obtained through an identification model, which is trained through steps 701 to 703. Each step is described below.

[0118] Step 701: Obtain target training data. Target training data includes training samples and off-labels. Training samples include game scene samples and at least one game event sample.

[0119] It should be noted that the target training data obtained by the fragment recognition device includes training samples and deviation labels. The training samples include game scene samples and at least one game event sample. The deviation label indicates whether at least one game event sample deviates from the game scene sample. Specifically, at least one game event sample can correspond to a normal game fragment or an abnormal game fragment. When at least one game event sample corresponds to a normal game fragment, the deviation label indicates that the game scene sample and at least one game event sample are not deviated, i.e., they are matched, for example, -1. When at least one game event sample can correspond to an abnormal game fragment, the deviation label indicates that the game scene sample and at least one game event sample are deviated, i.e., they are not matched, for example, 1.

[0120] Step 702: Based on the recognition model to be trained, obtain the first matching feature sample corresponding to the features of the game scene sample and the features of at least one game event sample, and determine the first prediction deviation probability based on the first matching feature sample.

[0121] In this embodiment, the fragment recognition device is equipped with a recognition model to be trained, or the fragment recognition device can obtain the recognition model to be trained from other devices or received information. This recognition model to be trained is a neural network model to be trained for obtaining deviation probabilities, such as a convolutional neural network model, a discriminator, a delay network model, a gated convolutional neural network model, etc. Thus, the fragment recognition device trains the recognition model to be trained based on the obtained target training data. During the training process, the fragment recognition device uses the recognition model to be trained to first obtain the features of the game scene samples and the features of at least one game event sample, then obtains the first matching feature samples corresponding to the features of the game scene samples and the features of at least one game event sample, and finally predicts the deviation probability based on the first matching feature samples to obtain the first predicted deviation probability of at least one game event sample deviating from the game scene sample.

[0122] It should be noted that the first matching feature sample and the matching feature are data of the same type and have similar meanings; the first predicted deviation probability and the target deviation probability are data of the same type and have similar meanings; the embodiments of this application will not be described again here.

[0123] Step 703: Based on the loss between the first predicted deviation probability and the deviation label, train the recognition model to be trained to obtain the recognition model.

[0124] In this embodiment of the application, after obtaining the first predicted conversion probability, the fragment recognition device compares the first predicted deviation probability with the deviation label, and performs backpropagation in the recognition model to be trained based on the difference between the first predicted deviation probability and the deviation label to adjust the model parameters in the recognition model to be trained, and obtains the recognition model based on the trained recognition model to be trained.

[0125] It should be noted that the training process of the recognition model to be trained is an iterative process. After each training iteration, it is determined whether the currently trained recognition model meets the training termination condition. If the training termination condition is met, training ends, and the currently trained recognition model is determined as the recognition model. If the training termination condition is not met, training continues until the training termination condition is met, and the recognition model trained at this point is determined as the recognition model. The training termination condition can be reaching an accuracy threshold, reaching a training count threshold, reaching a training duration threshold, or a combination of the above, etc. This application does not limit the specific criteria in this embodiment.

[0126] In this embodiment of the application, step 702 is followed by steps 704 and 705 (not shown in the figure); that is, after the fragment recognition device obtains the features of the game scene sample and the features of at least one game event sample based on the recognition model to be trained, the fragment recognition method further includes steps 704 and 705. Each step is described below.

[0127] Step 704: Based on the recognition model to be trained, generate at least one first game generation event corresponding to the first matching feature sample.

[0128] In this embodiment, the fragment recognition device can train the recognition model under test using supervised training, or it can combine supervised and unsupervised training methods to train the recognition model under test. Therefore, the fragment recognition device is also used to generate at least one first game generation event corresponding to a first matching feature sample based on the recognition model under test, so as to perform unsupervised training based on at least one game event sample and at least one first game generation event.

[0129] Step 705: Determine the first unsupervised loss based on at least one game event sample and at least one first game generation event.

[0130] In this embodiment, the fragment recognition device determines the difference between at least one game event sample and at least one first game generation event as the first unsupervised loss. The difference between the at least one game event sample and the at least one first game generation event can be a feature difference, a mean squared error, or a combination of both, etc., and this embodiment does not limit the specific difference.

[0131] Accordingly, in the embodiments of this application, step 703 can be implemented by step 7031 (not shown in the figure); that is, the fragment recognition method trains the recognition model to be trained based on the loss between the first predicted deviation probability and the deviation label to obtain the recognition model, including step 7031, which will be described below.

[0132] Step 7031: Based on the loss between the first predicted deviation probability and the deviation label, and the combined loss of the first unsupervised loss, train the recognition model to be trained to obtain the recognition model.

[0133] It should be noted that the fragment recognition device combines the loss between the first predicted deviation probability and the deviation label, as well as the first unsupervised loss, to obtain a combined loss. This combined loss is then used to train the recognition model to obtain the recognition model. Here, the fragment recognition device can directly sum the loss between the first predicted deviation probability and the deviation label, as well as the first unsupervised loss, or it can perform a weighted summation of the loss between the first predicted deviation probability and the deviation label, and the first unsupervised loss, etc. This embodiment does not limit the specific implementation.

[0134] Understandably, the fragment recognition device combines supervised and unsupervised training methods to train the recognition model, thus enabling multiple training methods for the model and improving its accuracy.

[0135] In the embodiments of this application, step 701 can be implemented by steps 7011 to 7014 (not shown in the figure); that is, the segment recognition device acquires target training data, including steps 7011 to 7014, and each step is described below.

[0136] Step 7011: Obtain initial training data.

[0137] It should be noted that the initial training data obtained by the fragment recognition device includes initial game scene samples corresponding to virtual object samples and at least one initial game event sample, as well as a first deviation label corresponding to the initial game scene sample and at least one initial game event sample. The virtual object sample can be a normal virtual object or an abnormal virtual object. When the virtual object sample is a normal virtual object, the first deviation label indicates that the initial game scene sample and at least one initial game event sample are not deviated, i.e., they are matched, for example, -1. When the virtual object sample is an abnormal virtual object, the first deviation label indicates that the initial game scene sample and at least one initial game event sample are deviated, i.e., they are not matched, for example, 1.

[0138] In this embodiment of the application, the fragment recognition device can directly determine the initial training data as the target training data. In this case, the initial game scene sample is the game scene sample, at least one initial game event sample is at least one game event sample, and the first deviation label is the deviation label. Alternatively, the initial training data can be augmented based on steps 7012 to 7014 to obtain the target training data. In this case, the target training data includes the initial training data and the training data augmented based on the initial training data.

[0139] Step 7012: Based on the generator model to be trained, obtain the features of the initial game scene sample and the features of at least one initial game event sample, and generate at least one second game generation event corresponding to the second matching feature sample.

[0140] In this embodiment, the fragment recognition device is equipped with a training generation model, or the fragment recognition device can obtain the training generation model from other devices or received information. This training generation model is a neural network model to be trained for generating game event samples, such as a convolutional neural network model, a generator, a delay network model, a gated convolutional neural network model, etc. Therefore, the fragment recognition device performs data augmentation on the initial training data based on the training generation model. During the augmentation process, the fragment recognition device uses the training generation model to first obtain the features of the initial game scene samples and the features of at least one initial game event sample, then obtains the second matching feature samples corresponding to the features of the initial game scene samples and the features of at least one initial game event sample, and finally generates an event sequence based on the second matching feature samples to obtain at least one second game generation event.

[0141] Step 7013: Determine the preset mismatch label as the second deviation label of the initial game scene sample and at least one second game generation event.

[0142] It should be noted that since at least one second game-generating event is generated by the model to be trained and is also abnormal, the fragment recognition device determines the second deviation label of the initial game scene sample and at least one second game-generating event as a preset mismatch label; wherein, the preset mismatch label indicates that the event sequence does not match the game scene and is deviated; that is, the initial game scene sample and at least one second game-generating event are negative samples.

[0143] Step 7014: Determine the initial game scene sample, at least one initial game event sample and the first deviation label combination, and the initial game scene sample, at least one second game generation event and the second deviation label combination as the target training data.

[0144] It should be noted that the combination of the initial game scene sample, at least one initial game event sample, and the first deviation label constitutes the initial training data; the combination of the initial game scene sample, at least one second game generation event, and the second deviation label constitutes the training data augmented based on the initial training data; after the fragment recognition device performs data augmentation on the initial training data, the target training data includes not only the initial training data but also the training data augmented based on the initial training data; thus, at this time, the game scene sample can be the initial game scene sample, and at least one game event sample can be at least one initial game event sample (at this time, the deviation label is the first deviation label) or at least one second game generation event (at this time, the deviation label is the second deviation label).

[0145] Understandably, after obtaining the initial training data, data augmentation is performed on the initial training data to expand the amount of training data corresponding to the recognition model to be trained. This can improve the training effect of the recognition model to be trained, and thus improve the recognition accuracy.

[0146] In this embodiment of the application, step 7012 is followed by steps 7015 and 7016 (not shown in the figure); that is, after the fragment recognition device generates at least one second game generation event corresponding to the second matching feature sample, the fragment recognition method further includes steps 7015 and 7016, and each step is described below.

[0147] Step 7015: Determine the second unsupervised loss based on the difference between at least one initial game event sample and at least one second game generation event.

[0148] In this embodiment, the fragment recognition device is further configured to determine the difference between at least one initial game event sample and at least one second game generation event as the loss for unsupervised training of the model to be trained, referred to here as the second unsupervised loss. The difference between at least one initial game event sample and at least one second game generation event can be a feature difference, a mean squared error, or a combination of both, etc., and this embodiment does not limit the specific differences.

[0149] Step 7016: Combine the second unsupervised loss and the second prediction deviation probability to train the generator model.

[0150] It should be noted that the second predicted deviation probability refers to the deviation probability jointly corresponding to the initial game scene sample obtained based on the recognition model to be trained and at least one second game generation event; that is, the second predicted deviation probability is the probability that at least one second game generation event deviates from the initial game scene sample, and the second predicted deviation probability is obtained through the recognition model to be trained; therefore, the second predicted deviation probability belongs to the category of second predicted deviation probabilities. The trained generation model is used to enhance the training data of the trained recognition model, while the trained recognition model is used to help determine the loss of the trained generation model, so as to realize the alternating training of the generation model and the recognition model.

[0151] In the embodiments of this application, when at least one initial game event sample includes multiple initial game event samples, step 7011 is followed by steps 7017 to 70111 (not shown in the figure); that is, after the fragment recognition device obtains the initial training data, the fragment recognition method further includes steps 7017 to 70111, and each step is described below.

[0152] Step 7017: Obtain the true deviation label corresponding to each initial game event sample.

[0153] It should be noted that the initial training samples obtained by the fragment recognition device include positive samples and negative samples. For positive samples, there are samples that match the initial game scene sample and at least one initial game event sample. However, at least one initial game event may also contain noise, that is, there are events that do not match the initial game scene sample. For negative samples, there are samples that do not match the initial game scene sample and at least one initial game event sample. However, at least one initial game event may also contain noise, that is, there are events that match the initial game scene sample. Therefore, the fragment recognition device performs noise reduction processing on the initial training data.

[0154] In this embodiment, the fragment recognition device acquires the true deviation label corresponding to each initial game event sample, including: acquiring the category prediction probability of each initial game event sample for each deviation label category; determining the confidence probability corresponding to each deviation label category based on the category prediction probability; and determining the deviation label category corresponding to the largest category prediction probability greater than the confidence probability as the true deviation label of the initial game event sample. Here, the fragment recognition can determine the category prediction probability through estimation, that is, it is estimated based on the features corresponding to the initial game event sample and the initial game scene sample respectively.

[0155] Step 7018: Based on the first deviation label and the true deviation label corresponding to the initial game event sample, and in combination with the deviation label category, determine the number of samples for each deviation label combination.

[0156] It should be noted that deviation label categories include two types: deviation category and matching category. Deviation label combination refers to a combination in which the first deviation label and the true deviation label are each from any of the deviation label categories, including: the first deviation label is a deviation category and the true deviation label is also a deviation category; the first deviation label is a deviation category and the true deviation label is a matching category; the first deviation label is a matching category and the true deviation label is also a matching category; and the first deviation label is a matching category and the true deviation label is a deviation category.

[0157] Step 7019: Based on the number of samples and the total number of initial game event samples, determine the joint probability distribution of the first deviation label and the true deviation label.

[0158] It should be noted that each element in the joint probability distribution represents the probability that the first deviation label is each deviation label category and the true deviation label is each deviation label category; for example, the probability that the first deviation label is a deviation category and the true deviation label is also a deviation category, the probability that the first deviation label is a deviation category and the true deviation label is also a deviation category, the probability that the first deviation label is a matching category and the true deviation label is also a deviation category, and the probability that the first deviation label is a matching category and the true deviation label is also a matching category.

[0159] Step 70110: Filter out initial game event samples that are lower than the label probability threshold in the joint probability distribution from multiple initial game event samples to obtain at least one target game event sample.

[0160] In this embodiment, the fragment recognition device is equipped with a label probability threshold, or the fragment recognition device can obtain the label probability threshold from other devices or received information. The label probability threshold represents the minimum probability of removing noise. Thus, the fragment recognition device filters out initial game event samples with a joint probability distribution that are lower than the label probability threshold from multiple initial game event samples, thereby obtaining at least one target game event sample. Among them, the initial game event samples that are greater than or equal to the label probability threshold are the target game event samples.

[0161] Step 70111: Combine the initial game scene samples, at least one target game event sample, and the first deviation label into target training data.

[0162] In this embodiment of the application, the combination of the initial game scene sample, at least one target game event sample, and the first deviation label is the initial training data after noise removal. Thus, the segment recognition identifies the combination of the initial game scene sample, at least one target game event sample, and the first deviation label as the target training data.

[0163] It should be noted that the process of model training for the segment recognition device is similar to the process of model application, and will not be described again in the embodiments of this application; the model training process can be implemented through various server devices, and the model application process can be implemented through various terminals that have deployed recognition models.

[0164] The following describes an exemplary application of the embodiments of this application in a real-world application scenario. This exemplary application describes the process of training a generator to identify whether a game segment in a MOBA contains cheats.

[0165] See Figure 8 , Figure 8 This is an exemplary system architecture diagram provided in the embodiments of this application; as shown below. Figure 8As shown, firstly, battlefield information 8-21 (referred to as the whole game scene information) and atomic actions 8-22 (referred to as the atomic game action set) of all game objects are extracted from the replay data 8-1. Then, from the battlefield information 8-21 and atomic actions 8-22 of all game objects, battlefield information 8-31 (referred to as game scene information) and decision information 8-32 (referred to as at least one game event) corresponding to suspicious segments (referred to as game segments to be identified) such as active virtual kills or active virtual attacks by a specified game object (referred to as game segments to be identified) are determined. Then, the battlefield information 8-31 and decision information 8-32 are input into the multi-task generative adversarial network (AGGAN) model 8-4. The perspective behavior classification module 8-42 (called the recognition model) in the multi-task generative adversarial network model 8-4 determines the deviation probability 8-6 (called the target deviation probability) of the decision information 8-32 from the battlefield information 8-31, thereby obtaining the abnormal probability of the suspicious segment. Based on the abnormal probability, it is then determined whether the specified game object is a virtual object that uses cheating to play the game (called an abnormal virtual object).

[0166] See also Figure 8 The behavior classification module 8-42 is obtained through training on the decision generation task 8-51 and decision rationality judgment task 8-52 in auxiliary task 8-5. Specifically, the decision generation task 8-51 is implemented through the decision sequence generation module in the multi-task generative adversarial network model 8-4, and the decision rationality judgment task 8-52 is implemented through the decision rationality judgment module in the multi-task generative adversarial network model 8-4. The decision generation task 8-51 is used to reproduce the behavior of normal game objects (i.e., virtual objects that play the game in a normal way; normal way refers to controlling game objects in the game scene by receiving user operations); the decision rationality judgment task 8-52 is used to measure whether the generated decision sequence is reasonable.

[0167] The training process of the multi-task generative adversarial network model 8-4 is described below.

[0168] See Figure 9 , Figure 9 This is an exemplary model training diagram provided in an embodiment of this application; as shown... Figure 9As shown, battlefield information 9-2 and decision information 9-3 of suspicious segment 9-1 are obtained. Battlefield information 9-2 includes macro-battlefield information 9-21, micro-battlefield information 9-22, and target battlefield information 9-23. Battlefield information 9-2 and decision information 9-3 are used by generator 9-4 (called the training generation model) to generate decision information 9-5. Finally, battlefield information 9-2 and decision information 9-5 are used as input to discriminator 9-6 (called the training recognition model) to output the deviation probabilities 9-7 of decision information 9-5 from battlefield feature 9-2. It is easy to see that since decision information 9-5 is generated decision information, the deviation probability 9-7 is higher than the probability threshold, indicating that decision information 9-5 does not match battlefield information 9-2, thus determining that suspicious segment 9-1 is an abnormal segment. Furthermore, when battlefield information 9-2 and decision information 9-5 are used as input to the discriminator 9-6, if the suspicious segment 9-1 is a segment of a normal game object playing the game, the output deviation probability is lower than the probability threshold; however, if the suspicious segment 9-1 is a segment of an abnormal game object playing the game, the output deviation probability is higher than the probability threshold. Figure 8 Multi-task generative adversarial network model 8-4 includes Figure 9 The generator 9-4 and discriminator 9-6 are in the example.

[0169] It should be noted that macro-battlefield information 9-21 refers to the intention information of macro-strategy, micro-battlefield characteristics 9-22 refers to the action information of micro-strategy, and target battlefield characteristics 9-23 represents the battlefield information associated with the target object.

[0170] In this embodiment of the application, the process of denoising the initial decision information of the suspicious segment 9-1 to obtain decision information 9-3 is also included. See Figure 10 , Figure 10 This is a schematic diagram of an exemplary noise reduction process provided in an embodiment of this application; as shown... Figure 10 As shown, for initial decision information 10-1 (referred to as at least one initial game event sample), the predicted probability P[i][j] (referred to as the category prediction probability) of the i-th decision in the initial decision information 10-1 under the j-th category (abnormal or normal) is obtained through model 10-2. Then, through cross-validation, the true category of each decision in the initial decision information 10-1 is determined based on the predicted probability P[i][j]. During the cross-validation process, the average predicted probability t[j] (referred to as the confidence probability) under category j corresponding to the initial decision information 10-1 is calculated, thus obtaining the confidence level for determining the category. Therefore, the true category y of the i-th decision can be obtained through equation (1). * (This is called the true deviation label), which is argmax. j P[i][j], Equation (1) is shown below:

[0171] argmaxj P[i][j],P[i][j]≥t[j] (1);

[0172] Among them, the noise category of the i-th decision in the initial decision information 10-1 This refers to the labeled category; thus, based on the true category and noise category of each decision, a counting matrix 10-3 can be obtained, where each element in the counting matrix 10-3 represents the number of decisions whose noise category is the u-th category (abnormal or normal, referred to as the off-label category or matching category) but whose true category is the v-th category (abnormal or normal).

[0173] Next, based on the counting matrix 10⁻³, the joint distribution matrix 10⁻⁴ (called the joint probability distribution) is calculated; the calculation process can be achieved through equations (2) and (3), which are shown below:

[0174]

[0175]

[0176] in, Noise category Let be the number of samples in the u-th category. The joint distribution matrix 10-4 reflects the distribution of noise and real categories. Based on the joint distribution matrix 10-4, decisions with incorrect category labeling are filtered out from black and white samples, resulting in decision information 9-3 (referred to as at least one target game event sample).

[0177] The following is about Figure 9 The training processes of generator 9-4 and discriminator 9-6 are explained separately.

[0178] See Figure 11 , Figure 11 This is a schematic diagram of an exemplary generator provided in an embodiment of this application; as shown... Figure 11 As shown, the exemplary generator 9-4 (referred to as the generative model to be trained) includes encoders 11-11 and decoders 11-12.

[0179] In encoder 11-11, the input information includes macro battlefield information 11-111, micro battlefield information 11-112, and target battlefield information 11-113 extracted from the replay data. By obtaining the embedding representations corresponding to macro battlefield information 11-111, micro battlefield information 11-112, and target battlefield information 11-113 respectively, macro battlefield features 11-114, micro battlefield features 11-115, and target battlefield features 11-116 are obtained respectively. Then, combined with position encoding 11-117, the final battlefield encoding features 11-118 (called the features of the initial game scene sample) are obtained, and the battlefield encoding features 11-118 are used as the input information of decoder 11-12.

[0180] In decoders 11-12, the input information also includes decision information 11-121 (referred to as at least one initial game event sample) extracted from the replay data; by obtaining the embedded representation of the decision information 11-121 and combining the embedded representation with the position encoding, decision features 11-122 (referred to as the features of at least one initial game event sample) are obtained; then, the decision features 11-122 and the battlefield encoding features 11-118 are translated by decoders 11-123 (including a normalization module and a multi-head attention module) and convolutional layers 11-124, and then connected to fully connected layers 11-127 (FC). The game statistics features 11-126 output by the Layer are concatenated (Concat), and the concatenated result is passed through the fully connected layers 11-128 to obtain the prediction decision sequence 11-129 (called at least one second game generation event); the input information of the decoder 11-12 also includes game statistics information 11-125 extracted from the replay data, such as whether the player died, the time without vision, ..., the duration of low health, whether the player returned to the virtual city, inside the virtual defense tower, ..., the number of teammates, etc.; the embedding representation of the game statistics information 11-125 This refers to the game statistics features 11-126. Furthermore, the predicted decision sequence 11-129, combined with game statistics 11-125, macro-battlefield information 11-111, micro-battlefield information 11-112, and target battlefield information 11-113, and processed by discriminator 9-6, can obtain the deviation probability 11-3 (called the second predicted deviation probability) corresponding to the predicted decision sequence 11-129. This also enhances the training data of discriminator 9-6, improving its discrimination performance. Generator 9-4 is the decision sequence generation module, used to implement... Figure 8 The decision generation task in section 8-51. Therefore, for generator 9-4, supervised loss can be obtained. As shown in equation (4).

[0181]

[0182] Among them, G s To predict decision sequences 11-129, D(G) s Let represent the deviation probability 11-3 corresponding to the predicted decision sequence 11-129. Furthermore, it is possible to obtain the unsupervised loss based on the decision information 11-121 and the predicted decision sequence 11-129. (referred to as the second unsupervised loss), as shown in equation (5).

[0183]

[0184] MSE() is used to calculate the mean squared error. This indicates decision information 11-121. This represents the prediction decision sequence 11-129.

[0185] Therefore, the loss function L of generator 9-4 G As shown in equation (6).

[0186]

[0187] See Figure 12 , Figure 12 This is a schematic diagram of an exemplary discriminator provided in an embodiment of this application; as shown... Figure 12 As shown, the exemplary discriminator 9-6 (referred to as the recognition model to be trained) includes an encoder 12-11 and a decoder 12-12.

[0188] In encoders 12-11, the input information may include... Figure 11 The macro battlefield information 11-111, micro battlefield information 11-112, and target battlefield information 11-113 are obtained. By acquiring the embedding representations corresponding to the macro battlefield information 11-111, micro battlefield information 11-112, and target battlefield information 11-113 respectively, the corresponding macro battlefield features 12-111, micro battlefield features 12-112, and target battlefield features 12-113 are obtained. Then, combined with the position encoding 12-114, the final battlefield encoding features 12-115 (called the features of the game scene sample) are obtained, and the battlefield encoding features 12-115 are used as the input information of the decoder 12-12.

[0189] In decoder 12-12, the input information also includes Figure 9The predicted decision sequences 11-129 (referred to as at least one game event sample) are obtained. Decision features 12-121 are obtained by acquiring the embedded representations corresponding to the predicted decision sequences 11-129. Then, decision features 12-121 and battlefield coding features 12-115 are passed through a decoder 12-122 and a convolutional layer 12-123, and then concatenated with the game statistics features 12-125 output by a fully connected layer 12-126. Based on the concatenation result, the predicted decision sequence 12-127 (referred to as at least one first game generation event) is obtained, along with... Figure 11 The deviation probability 11-3 (referred to as the first prediction deviation probability) is also included in the input information of decoder 12-12. Figure 11 The game statistics information 11-125 is processed through fully connected layers 12-126 to obtain the embedded representation of the game statistics information 11-125, thus yielding the game statistics features 12-125. The discriminator 9-64 is the decision rationality judgment module, used to implement... Figure 8 The task of determining the rationality of decisions is 8-52. Therefore, for the discriminator 9-6, based on the predicted decision sequences 11-129 and 12-127, the unsupervised loss can be obtained. As shown in equation (7).

[0190]

[0191] in, and Obtained through equation (8), and The corresponding y is either no deviation or a match (e.g., -1). The corresponding y is deviation or mismatch (e.g., 1), as shown in equation (8).

[0192]

[0193] A(x) is obtained through equation (9), which is shown below.

[0194]

[0195] Furthermore, the supervised loss of discriminator 9-6 can be obtained based on the deviation probability 11-3. As shown in equation (10).

[0196]

[0197] in, s The deviation probability is 11-3. Since the input prediction decision sequence 11-129 is generated, y is 1.

[0198] Therefore, the loss L of discriminator 9-6 DIt can be obtained through equation (11), which is shown below.

[0199]

[0200] Where α is a hyperparameter.

[0201] It should be noted that the input to the discriminator can be information about normal game objects (in which case, y takes the value of -1), or information generated by the generator based on information about normal or abnormal game objects (in which case, y takes the value of 1). Furthermore, during training, the generator and discriminator are trained alternately, and an optimizer (e.g., RMSprop) is also used for training, with a learning rate set to, for example, 1e-4; see Table 2 for a description of hyperparameter-related information.

[0202] Table 2

[0203] Hyperparameters Value Learning rate (Lr) 1e-4 Number of layers in translation (num_layers) 1 Translation parameter 1 (d_model) 128 Dropout rate 0.1 Length of suspicious segment (time_len) 20 seconds Translation parameter 1 (num_heads) 8

[0204] The following describes the beneficial effects of the fragment recognition method provided in the embodiments of this application.

[0205] See Table 3, which describes the indicator results for recognition based on confidence learning.

[0206] Table 3

[0207]

[0208] Referring to Table 3, after combining confidence learning, the F1 score of the external fragment can reach 84.18%, which is 14.91% higher than the F1 score combined with confidence learning. This shows that confidence learning can effectively remove noisy samples in sample selection.

[0209] See Table 4, which describes the metrics for different fragment recognition methods on different datasets.

[0210] Table 4

[0211]

[0212]

[0213] Referring to Table 4, the multi-task generative adversarial network (GAN) model achieves an F1 score of 85.45% when identifying anomalous segments and 82.62% when identifying anomalous games, which is higher than the F1 scores of the base models (translation model and GAN model). This indicates that the multi-task GAN model, as well as the multi-task GAN model combined with belief learning, can accurately achieve identification.

[0214] The following description continues to illustrate the exemplary structure of the fragment recognition device 255 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 3 As shown, the software module stored in the fragment recognition device 255 of the memory 250 may include:

[0215] The data acquisition module 2551 is used to acquire, based on the game data of the virtual scene, the game scene information of the virtual object to be identified in the game segment to be identified and at least one game event.

[0216] The feature extraction module 2552 is used to obtain the game scene features corresponding to the game scene information, and to obtain at least one game event feature corresponding to at least one game event;

[0217] The probability prediction module 2553 is used to determine the target deviation probability of at least one of the game events deviating from the game scene information based on the matching features of the game scene features and at least one of the game event features.

[0218] The anomaly identification module 2554 is used to determine whether the game segment to be identified is abnormal based on the comparison result between the target deviation probability and the probability threshold.

[0219] In this embodiment of the application, the game scene information includes at least one of global game scene information, local game scene information, and associated target game scene information;

[0220] Wherein, the global game scene information refers to the game information obtained based on the vision of the virtual object to be identified, the local game scene information refers to the information of the virtual object to be identified when it is playing a virtual game, and the associated target game scene information refers to the game information between the virtual object to be identified and the virtual object associated with the virtual object to be identified.

[0221] In this embodiment of the application, when the game scene information includes global game scene information, local game scene information, and associated target game scene information, the feature extraction module 2552 is further configured to obtain the intent feature corresponding to the global game scene information; obtain the action feature corresponding to the local game scene information; obtain the associated target feature corresponding to the associated target game scene information; and combine the intent feature, the action feature, and the associated target feature into the game scene feature.

[0222] In this embodiment of the application, the feature extraction module 2552 is further configured to extract game statistics from the game data, wherein the game statistics represent the statistical information of the virtual object to be identified in the virtual game process; and obtain auxiliary features corresponding to the game statistics.

[0223] In this embodiment of the application, the probability prediction module 2553 is further configured to concatenate the matching features of the game scene features and at least one of the game event features, as well as the auxiliary features, to obtain the features to be identified; and based on the features to be identified, to determine the target deviation probability of at least one of the game events deviating from the game scene information.

[0224] In this embodiment, the game event is determined based on multiple atomic game actions. The data acquisition module 2551 is further configured to extract whole game scene information and a set of atomic game actions from the game data of the virtual scene; determine the game scene information of the virtual object to be identified in the game segment to be identified from the whole game scene information, and determine the atomic game action sequence of the virtual object to be identified in the game segment to be identified from the set of atomic game actions; and determine at least one game event corresponding to the atomic game action sequence based on a preset relationship between events and atomic game actions, wherein the preset relationship is a one-to-many relationship between events and atomic game actions.

[0225] In this embodiment, the target deviation probability is obtained through a recognition model. The fragment recognition device 255 further includes a model training module 2555 for acquiring target training data. The target training data includes training samples and deviation labels. The training samples include game scene samples and at least one game event sample. The deviation label indicates whether at least one game event sample deviates from the game scene sample. Based on the recognition model to be trained, the features of the game scene sample and the features of at least one game event sample are acquired, and a first predicted deviation probability is determined based on the first matched feature sample. The recognition model to be trained is a neural network model to be trained for acquiring the deviation probability. Based on the loss between the first predicted deviation probability and the deviation label, the recognition model to be trained is trained to obtain the recognition model.

[0226] In this embodiment of the application, the model training module 2555 is further configured to generate at least one first game generation event corresponding to the first matching feature sample based on the recognition model to be trained; and to determine a first unsupervised loss based on at least one game event sample and at least one first game generation event.

[0227] In this embodiment of the application, the model training module 2555 is further configured to train the recognition model to be trained based on the loss between the first predicted deviation probability and the deviation label, and the combined loss of the first unsupervised loss, to obtain the recognition model.

[0228] In this embodiment of the application, the model training module 2555 is further configured to acquire initial training data, wherein the initial training data includes initial game scene samples, at least one initial game event sample, and a first deviation label; based on the model to be trained, acquire the features of the initial game scene samples and the features of at least one initial game event sample corresponding to a second matching feature sample, and generate at least one second game generation event corresponding to the second matching feature sample; determine a preset mismatch label as a second deviation label for the initial game scene sample and at least one second game generation event; and determine the combination of the initial game scene sample, at least one initial game event sample, and the first deviation label, and the combination of the initial game scene sample, at least one second game generation event, and the second deviation label as the target training data.

[0229] In this embodiment of the application, the model training module 2555 is further configured to determine a second unsupervised loss based on the difference between at least one initial game event sample and at least one second game generation event; and to train the training generation model by combining the second unsupervised loss and the second predicted deviation probability; wherein, the second predicted deviation probability refers to the probability that at least one second game generation event deviates from the initial game scene sample, and the second predicted deviation probability is obtained by the training recognition model; the trained training generation model is used to enhance the training data of the trained training recognition model.

[0230] In this embodiment of the application, when at least one initial game event sample includes multiple initial game event samples, the model training module 2555 is further configured to obtain the true deviation label corresponding to each initial game event sample; based on the first deviation label and the true deviation label corresponding to the initial game event sample, and combined with the deviation label category, determine the number of samples for each deviation label combination, wherein the deviation label category includes two types: deviation category and matching category, and the deviation label combination refers to a combination of the first deviation label and the true deviation label respectively in any of the deviation label categories; based on the number of samples and the total number of initial game event samples, determine the joint probability distribution of the first deviation label and the true deviation label; filter out the initial game event samples below the label probability threshold in the joint probability distribution from the multiple initial game event samples to obtain at least one target game event sample; and combine the initial game scene sample, at least one target game event sample, and the first deviation label into the target training data.

[0231] In this embodiment of the application, the model training module 2555 is further configured to obtain the category prediction probability of each initial game event sample corresponding to each deviation label category; determine the confidence probability corresponding to each deviation label category based on the category prediction probability; and determine the deviation label category corresponding to the largest category prediction probability greater than the confidence probability as the true deviation label of the initial game event sample.

[0232] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device (referred to as a fragment recognition device) reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the fragment recognition method described above in this application.

[0233] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, they cause the processor to execute the fragment recognition method provided in this application. For example, ... Figure 4 The fragment recognition method is shown.

[0234] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0235] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0236] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0237] As an example, executable instructions can be deployed to execute on a single electronic device (in this case, the single electronic device is a fragment recognition device), or to execute on multiple electronic devices located at one location (in this case, the multiple electronic devices located at one location are fragment recognition devices), or to execute on multiple electronic devices distributed across multiple locations and interconnected via a communication network (in this case, the multiple electronic devices distributed across multiple locations and interconnected via a communication network are fragment recognition devices).

[0238] It is understood that in the embodiments of this application, data related to game scene information is involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0239] In summary, in this embodiment, the process of identifying the anomalies of game segments to be identified in a virtual scene is achieved by extracting game scene information from game data, determining at least one game event corresponding to the game scene information, and judging the deviation probability of at least one game event from the target of the game scene information. Since game events are determined based on multiple atomic game actions, they are abstractions of game decisions and represent the intent of the virtual game. Therefore, by judging the matching between game scene features and at least one game event feature, it is possible to accurately determine whether the game scene information supports at least one game event, and thus accurately obtain the identification result of whether the game segment to be identified is abnormal. Therefore, the accuracy of segment identification in virtual scenes can be improved.

[0240] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A fragment recognition method based on virtual scenes, characterized in that, The method includes: Based on the game data of the virtual scene, obtain the game scene information and at least one game event of the virtual object to be identified in the game segment to be identified; Obtain the game scene features corresponding to the game scene information, and obtain at least one game event feature corresponding to at least one game event; Based on the matching features of the game scene features and at least one game event feature, a target deviation probability of at least one game event deviating from the game scene information is determined; the target deviation probability refers to the degree of deviation of at least one game event from the game scene information; the target deviation probability is obtained through a recognition model; the recognition model is a model corresponding to an artificial neural network. Based on the comparison between the target deviation probability and the probability threshold, the identification result of whether the game segment to be identified is abnormal is determined; the probability threshold represents the minimum probability of determining that at least one game event matches the game scene information.

2. The method according to claim 1, characterized in that, The game scenario information includes at least one of global game scenario information, local game scenario information, and associated target game scenario information; Wherein, the global game scene information refers to the game information obtained based on the vision of the virtual object to be identified, the local game scene information refers to the information of the virtual object to be identified when it is playing a virtual game, and the associated target game scene information refers to the game information between the virtual object to be identified and the virtual object associated with the virtual object to be identified.

3. The method according to claim 1 or 2, characterized in that, When the game scene information includes global game scene information, local game scene information, and associated target game scene information, obtaining the game scene features corresponding to the game scene information includes: Obtain the intent features corresponding to the global game scene information; Obtain the action features corresponding to the local game scene information; Obtain the associated target features corresponding to the associated target game scene information; The intent feature, the action feature, and the associated target feature are combined to form the game scenario feature.

4. The method according to claim 1 or 2, characterized in that, Before determining the target deviation probability of at least one of the game events deviating from the game scene information based on the matching features of the game scene features and at least one of the game event features, the method further includes: From the game data, game statistics are extracted, wherein the game statistics represent the statistical information of the virtual object to be identified in the virtual game process. Obtain the auxiliary features corresponding to the game statistics; The step of determining the target deviation probability of at least one of the game events from the game scene information based on the matching features of the game scene features and at least one of the game event features includes: The matching features of the game scene features and at least one of the game event features, as well as the auxiliary features, are concatenated to obtain the features to be identified; Based on the features to be identified, determine the target deviation probability of at least one of the game events deviating from the game scene information.

5. The method according to claim 1 or 2, characterized in that, The game events are determined based on multiple atomic game actions; The game data based on the virtual scene, obtaining game scene information and at least one game event for the virtual object to be identified in the game segment to be identified, includes: Extract the entire game scene information and the set of atomic game actions from the game data of the virtual scene; From the overall game scene information, determine the game scene information of the virtual object to be identified in the game segment to be identified, and from the atomic game action set, determine the atomic game action sequence of the virtual object to be identified in the game segment to be identified; Based on the preset relationship between events and atomic game actions, at least one game event corresponding to the atomic game action sequence is determined, wherein the preset relationship between events and atomic game actions is a one-to-many relationship.

6. The method according to claim 1, characterized in that, The recognition model is obtained through the following steps: Obtain target training data, wherein the target training data includes training samples and deviation labels, the training samples include game scene samples and at least one game event sample, and the deviation labels indicate whether at least one game event sample deviates from the game scene sample; Based on the recognition model to be trained, the features of the game scene sample and the features of at least one game event sample are obtained as first matching feature samples, and a first prediction deviation probability is determined based on the first matching feature samples, wherein the recognition model to be trained is a neural network model to be trained for obtaining the deviation probability. Based on the loss between the first predicted deviation probability and the deviation label, the recognition model to be trained is trained to obtain the recognition model.

7. The method according to claim 6, characterized in that, After obtaining the features of the game scene sample and the features of at least one game event sample based on the recognition model to be trained, the method further includes: Based on the recognition model to be trained, at least one first game generation event corresponding to the first matching feature sample is generated; A first unsupervised loss is determined based on at least one of the game event samples and at least one of the first game generation events; The step of training the recognition model based on the loss between the first predicted deviation probability and the deviation label to obtain the recognition model includes: The recognition model is trained based on the loss between the first predicted deviation probability and the deviation label, and the combined loss of the first unsupervised loss, to obtain the recognition model.

8. The method according to claim 6 or 7, characterized in that, The acquisition of target training data includes: Acquire initial training data, wherein the initial training data includes initial game scenario samples, at least one initial game event sample, and a first deviation label; Based on the model to be trained, the features of the initial game scene sample and the features of at least one initial game event sample are obtained as a second matching feature sample, and at least one second game generation event corresponding to the second matching feature sample is generated. The preset mismatch label is determined as the second deviation label of the initial game scene sample and at least one of the second game generation events; The initial game scene sample, at least one initial game event sample and the first deviation label combination, and the initial game scene sample, at least one second game generation event and the second deviation label combination are determined as the target training data.

9. The method according to claim 8, characterized in that, After generating at least one second game generation event corresponding to the second matching feature sample, the method further includes: The second unsupervised loss is determined based on the difference between at least one initial game event sample and at least one second game generation event; The trainable generative model is trained by combining the second unsupervised loss and the second prediction deviation probability. Wherein, the second predicted deviation probability refers to the probability that at least one of the second game generation events deviates from the initial game scene sample, and the second predicted deviation probability is obtained through the recognition model to be trained; the trained generation model to be trained is used to enhance the training data of the trained recognition model to be trained.

10. The method according to claim 8, characterized in that, When at least one of the initial game event samples includes multiple initial game event samples, after obtaining the initial training data, the method further includes: Obtain the true deviation label corresponding to each of the initial game event samples; Based on the first deviation label and the true deviation label corresponding to the initial game event sample, and combined with the deviation label category, the number of samples for each deviation label combination is determined. The deviation label category includes two types: deviation category and matching category. The deviation label combination refers to a combination in which the first deviation label and the true deviation label are each one of the deviation label categories. Based on the number of samples and the total number of initial game event samples, determine the joint probability distribution of the first deviation label and the true deviation label; From the multiple initial game event samples, filter out the initial game event samples that are below the label probability threshold in the joint probability distribution to obtain at least one target game event sample; The initial game scene sample, at least one target game event sample, and the first deviation label are combined to form the target training data.

11. The method according to claim 10, characterized in that, The step of obtaining the true deviation label corresponding to each initial game event sample includes: Obtain the category prediction probability for each of the initial game event samples corresponding to each deviation label category; Based on the predicted category probabilities, determine the confidence probability corresponding to each of the deviation label categories; The deviation label category corresponding to the largest predicted category probability greater than the confidence probability is determined as the true deviation label of the initial game event sample.

12. A fragment recognition device based on a virtual scene, characterized in that, The fragment recognition device includes: The data acquisition module is used to acquire, based on the game data of the virtual scene, the game scene information of the virtual object to be identified in the game segment to be identified and at least one game event. The feature extraction module is used to obtain game scene features corresponding to the game scene information, and to obtain at least one game event feature corresponding to at least one game event; The probability prediction module is used to determine the target deviation probability of at least one game event deviating from the game scene information based on the matching features of the game scene features and at least one game event feature; the target deviation probability refers to the degree of deviation of at least one game event from the game scene information; the target deviation probability is obtained through a recognition model; the recognition model is a model corresponding to an artificial neural network; An anomaly detection module is used to determine whether the game segment to be identified is abnormal based on the comparison result between the target deviation probability and the probability threshold; the probability threshold represents the minimum probability of determining that at least one game event matches the game scene information.

13. A fragment recognition device based on a virtual scene, characterized in that, The fragment recognition device includes: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the method according to any one of claims 1 to 11.

14. A computer-readable storage medium storing executable instructions, characterized in that, The executable instructions, when executed by a processor, implement the method described in any one of claims 1 to 11.

15. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method described in any one of claims 1 to 11.

Images