Object recognition method, device, computer equipment and storage medium
By distributing analysis of resource transfer characteristics of network objects and statistics of the results of multiple anomaly detection models, the problems of label errors and low recognition accuracy in traditional technology are solved, and efficient object recognition and abnormal detection are achieved.
Patent Information
- Application Number
- CN202110475178.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-04-29
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-04-29
AI Technical Summary
Traditional network detection models require a large number of label samples training in object recognition. The existence of label errors leads to a reduced recognition accuracy and lacks effective anomaly detection strategies.
By obtaining the resource transfer feature set of the target object, using multiple anomaly detection models for feature distribution analysis, combining the results of each model for statistics, determining the abnormality detection results of the target object, and improving the recognition accuracy.
It improves the accuracy and efficiency of object recognition, reduces dependence on sample tags, and can perform accurate abnormality detection without tag information.
Smart Images

Figure CN113761292B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of network identification technology, and in particular to an object identification method, device, computer equipment and storage medium. Background Art
[0002] With the development of network recognition technology, a technology for identifying objects through network detection models has emerged. In this technology, the input features are identified through the network detection model to obtain the corresponding model results.
[0003] In traditional technology, a network detection model is trained based on labeled samples, and objects are identified based on the trained network detection model. However, this method requires a lot of time to collect a large number of labeled samples for model training, and there are cases of label errors, which leads to reduced accuracy of object recognition.
[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present application, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention
[0005] Based on this, it is necessary to provide an object recognition method, device, computer equipment and storage medium to address the above technical problems.
[0006] An object recognition method, the method comprising: obtaining a target object set to be identified for anomalies, the target object set comprising multiple target objects; obtaining a resource transfer feature set corresponding to each of the target objects, the resource transfer feature set comprising resource transfer features corresponding to multiple feature dimensions; determining an anomaly detection model set, the anomaly detection model set comprising multiple anomaly detection models; the anomaly detection models in the anomaly detection model set adopt different anomaly detection strategies; determining, through each anomaly detection model, the distribution result of each resource transfer feature in the resource transfer feature set corresponding to the target object in the target feature set corresponding to the feature dimension; determining the model detection result of the anomaly detection model on the target object based on the distribution result obtained by the anomaly detection model; and statistically analyzing the model detection results of each anomaly detection model in the anomaly detection model set on the target object to obtain the anomaly detection result of the target object.
[0007] An object recognition device, the device comprising: an object set acquisition module, used to acquire a target object set to be identified for anomaly, the target object set comprising a plurality of target objects; a feature set acquisition module, used to acquire a resource transfer feature set corresponding to each of the target objects, the resource transfer feature set comprising resource transfer features corresponding to a plurality of feature dimensions; a model set determination module, used to determine an anomaly detection model set, the anomaly detection model set comprising a plurality of anomaly detection models; the anomaly detection models in the anomaly detection model set adopt different anomaly detection strategies; a feature distribution determination module, used to determine, through each of the anomaly detection models, the distribution results of each resource transfer feature in the resource transfer feature set corresponding to the target object in the target feature set corresponding to the feature dimension; a model detection result determination module, used to determine the model detection result of the anomaly detection model on the target object based on the distribution result obtained by the anomaly detection model; an anomaly detection result determination module, used to perform statistics on the model detection results of each anomaly detection model in the anomaly detection model set on the target object, to obtain an anomaly detection result of the target object.
[0008] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the following steps when executing the computer program: obtaining a target object set to be identified for anomaly, wherein the target object set comprises a plurality of target objects; obtaining a resource transfer feature set corresponding to each of the target objects, wherein the resource transfer feature set comprises resource transfer features corresponding to a plurality of feature dimensions; determining an anomaly detection model set, wherein the anomaly detection model set comprises a plurality of anomaly detection models; the anomaly detection models in the anomaly detection model set adopt different anomaly detection strategies; determining, through each of the anomaly detection models, a distribution result of each resource transfer feature in the resource transfer feature set corresponding to the target object in the target feature set corresponding to the feature dimension; determining a model detection result of the anomaly detection model on the target object based on the distribution result obtained by the anomaly detection model; and performing statistics on the model detection results of each anomaly detection model in the anomaly detection model set on the target object to obtain an anomaly detection result of the target object.
[0009] A computer-readable storage medium stores a computer program, which implements the following steps when executed by a processor: obtaining a target object set to be identified for anomaly, the target object set including multiple target objects; obtaining a resource transfer feature set corresponding to each of the target objects, the resource transfer feature set including resource transfer features corresponding to multiple feature dimensions; determining an anomaly detection model set, the anomaly detection model set including multiple anomaly detection models; the anomaly detection models in the anomaly detection model set adopt different anomaly detection strategies; determining, through each of the anomaly detection models, the distribution result of each resource transfer feature in the resource transfer feature set corresponding to the target object in the target feature set corresponding to the feature dimension; determining the model detection result of the anomaly detection model on the target object based on the distribution result obtained by the anomaly detection model; and statistically analyzing the model detection results of each anomaly detection model in the anomaly detection model set on the target object to obtain the anomaly detection result of the target object.
[0010] The object recognition method, device, computer equipment and storage medium described above obtain a target object set, obtain a resource transfer feature set corresponding to each target object in the target object set, analyze the distribution of resource transfer features in the target feature set of the feature dimension based on the anomaly detection model, and determine the model detection result corresponding to the target object based on the distribution results of each anomaly detection model. Since the distribution of resource transfer features is analyzed in units of feature dimensions, and the model detection result corresponding to the object is obtained by combining the distribution results of each feature dimension, it is possible to comprehensively detect anomalies of the object by combining the resource transfer features of the object in each feature dimension, so that the model detection result output by the model has high accuracy. Moreover, the anomaly detection strategies adopted by the anomaly detection models are different, and the model detection results corresponding to these anomaly detection models are comprehensively counted to obtain the anomaly detection result of the target object. The anomaly detection result of the target object can be determined based on multiple distribution results, which can effectively improve the accuracy of object recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] Figure 1 A diagram of an application environment of an object recognition method in an embodiment;
[0012] Figure 2 is a schematic flow chart of an object recognition method in one embodiment;
[0013] Figure 3 A schematic diagram of dividing resource transfer features in a feature set in one embodiment;
[0014] Figure 4A schematic diagram of comparing anomaly detection model set with a label database in one embodiment;
[0015] Figure 5 A schematic diagram of obtaining resource transfer features and performing anomaly detection in one embodiment;
[0016] Figure 6 is a schematic diagram of a model detection result in one embodiment;
[0017] Figure 7 is a flow chart of an object recognition method in another embodiment;
[0018] Figure 8 A schematic diagram of the proportion of objects of various subdivided types after abnormal objects are refined in one embodiment;
[0019] Fig. 9 is a structural block diagram of an object recognition device in one embodiment;
[0020] Fig.10 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0021] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0022] For the object recognition method disclosed in the present application, the data generated during the operation of the method can be stored on the blockchain. For example, the resource transfer feature set is recorded in the blockchain, the server obtains the resource transfer feature set corresponding to each target object from the blockchain, and the classification analysis is performed based on the resource transfer features in the resource transfer feature set through each anomaly detection model in the anomaly detection model set, and the corresponding distribution results are obtained. The model detection results for the target object are determined based on the distribution results of each anomaly detection model, and the model detection results of each anomaly detection model are counted. The anomaly detection results of the target object are obtained based on the statistical results of the model detection results, and the anomaly detection results are recorded in the blockchain. The terminal can obtain the anomaly detection results from the blockchain. At the same time, the target object set, resource transfer feature set, distribution results, model detection results, etc. generated during the operation process can all be recorded in the blockchain. By recording the data in the blockchain, it is possible to prevent the data from being tampered with, ensure the storage reliability of the data, and thus ensure the reliability of the determined anomaly detection results.
[0023] In one embodiment, the object recognition method, device, computer equipment and storage medium provided by the present application can be implemented through artificial intelligence technology. Among them, 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. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that the machines have the functions of perception, reasoning and decision-making.
[0024] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, mechatronics, and other technologies. Artificial intelligence software technologies mainly include computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.
[0025] Among them, Machine Learning (ML) is a multi-disciplinary interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning. Based on the classification of learning methods, machine learning includes at least one of supervised learning, unsupervised learning, or reinforcement learning.
[0026] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, drones, robots, smart medical care, smart customer service, etc. I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.
[0027] The solution provided in the embodiments of the present application involves artificial intelligence, machine learning and other technologies, which are specifically described by the following embodiments:
[0028] The object recognition method provided in this application can be applied to Figure 1 In the application environment shown. The application environment includes a terminal 102, a server 104 and a terminal 106. Among them, the terminal 102 and the terminal 106 communicate with the server 104 through the network respectively. Among them, the server 104 obtains the resource transfer feature set corresponding to each target object from the terminal 102, obtains each anomaly detection model in the anomaly detection (Fraud detection) model set from the terminal 106, and performs classification analysis based on the resource transfer features in the resource transfer feature set through each anomaly detection model to obtain a distribution result, determines the model detection result of the target object based on the distribution result of each anomaly detection model, obtains the anomaly detection result of the target object based on the statistical result of the model detection result, and returns the anomaly detection result to the terminal 102. Further, the terminal 102 and the terminal 106 can be, but are not limited to, various personal computers, laptops, smart phones, tablet computers and portable wearable devices. In addition, the terminal 102 and the terminal 106 can also be implemented by the same terminal device. Further, a client can be installed in the terminal 102 or the terminal 106, for example: a client for resource transfer. The server 102 may be implemented as an independent server or a server cluster consisting of multiple servers.
[0029] In one embodiment, Figure 2 As shown, an object recognition method is provided. This embodiment uses the method applied to a server as an example for illustration. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The object recognition method includes the following steps:
[0030] S202, obtaining a target object set for abnormality identification, where the target object set includes multiple target objects.
[0031] The target object refers to the subject that needs to be identified for anomalies, which can be a network user, for example, a registered user on a specific application or network platform. The application can be an application with resource processing functions, and can be an application provided by a fund company or a financial management platform, for example: financial management and trading applications such as stock trading applications and fund trading applications. Furthermore, the target object can be a stock account, a fund account, etc.
[0032] It should be noted that, in the embodiment of the present application, the plurality may be at least two. Further, the target object set includes at least two target objects. The object types of these target objects may be the same or different. Among them, the object type may be a stock account type, a fund account type, etc.
[0033] Furthermore, the target object set may be a set of objects that meet the screening conditions on a specific application. The screening conditions may be a match between an object feature and a target object feature, and the target object feature may be a predetermined object attribute or an operation feature determined based on an object operation. Examples of target object sets are as follows: 1. Determine the target object set based on object attributes, and the target object set is a set consisting of objects with a registration period of 1-3 years; 2. Determine the target object set based on the network operations performed, and the target object set is a set consisting of objects whose number of resource transfers between 0:00-06:00 in the morning exceeds a set number threshold.
[0034] In one embodiment, the target object set may be stored in the terminal. The implementation process of S202 may be: the server reads the target object set from the terminal.
[0035] S204, obtaining a resource transfer feature set corresponding to each target object, wherein the resource transfer feature set includes resource transfer features corresponding to a plurality of feature dimensions.
[0036] Among them, resources refer to items that can be owned through the network, including at least one of virtual items and physical items. Among them, virtual items can specifically include but are not limited to at least one of account values, funds, stocks, bonds, virtual image products, virtual recharge cards, game equipment, etc. Physical items can be any items with actual form that can be owned by users, and can specifically include but are not limited to electronic products, toys, crafts, or signed photos, etc. In one embodiment, resources can be resources of a specific type of network account, such as at least one of funds, stocks, or bonds, etc.
[0037] Resource transfer can be transferring resources from one storage medium to another storage medium, or transferring the resource value corresponding to the resource from a certain account, or transferring the resource value corresponding to the resource to a certain account. The storage medium can be a computer device with resource storage function, or a server or storage device corresponding to a bank.
[0038] The resource transfer feature is a feature corresponding to the resource transfer, which can be a feature under multiple feature dimensions, and each feature dimension can correspond to at least one resource transfer feature. Among them, the feature dimension can include at least one of the following: time dimension, frequency dimension, object dimension or amount dimension. Further, under the time dimension, the resource transfer feature can be the time of the resource transfer operation; under the frequency dimension, the resource transfer feature can be at least one of the following: the frequency of the resource transfer operation or the resource transfer frequency in a set time period; under the object dimension, the resource transfer feature can be at least one of the following: the receiving object for the resource transfer operation, the number of objects of the receiving object, the object feature of the receiving object, the sending object for the resource transfer operation, the number of objects of the sending object or the object feature of the sending object. In the transaction scenario, the receiving object and the sending object can also be referred to as counterparties; under the amount dimension, the resource transfer feature can be at least one of the following: the transferred resource value or the resource transfer value in a set time period, etc. In some embodiments, each resource transfer feature can also be used as a feature dimension, that is, one resource transfer feature is used as a feature dimension.
[0039] Furthermore, the resource transfer feature set is a set consisting of multiple resource transfer features. A target object may correspond to at least one resource transfer feature set.
[0040] The resource transfer feature set may be stored in the terminal, and the server may obtain the resource transfer feature set from the terminal. Specifically, the server sends a feature set acquisition request to the terminal, and the feature set acquisition request carries the object identifier of each target object to trigger the terminal to return the resource transfer feature set corresponding to each object identifier to the server. In addition, the resource transfer feature set may also be stored in a storage space corresponding to the server, and the server may obtain the resource transfer feature set corresponding to each target object by reading the storage space.
[0041] The implementation process of S204 may be: the server obtains resource transfer features of each target object under multiple feature dimensions from the terminal device respectively, and the resource transfer features corresponding to each target object constitute a resource transfer feature set, thereby obtaining the resource transfer feature set corresponding to each target object.
[0042] S206, determining an anomaly detection model set, where the anomaly detection model set includes multiple anomaly detection models; the anomaly detection models in the anomaly detection model set adopt different anomaly detection strategies.
[0043] Among them, the anomaly detection model is a network model that can realize abnormal object detection according to the anomaly detection strategy of the feature. Further, the anomaly detection model can be a machine learning model, such as a supervised learning model, a semi-supervised learning model or an unsupervised learning model. Specifically, the anomaly detection model can be at least one of the models such as random forest, isolation forest (iForest), HBOS (Histogram-based Outlier Score, anomaly score based on histogram), COPOD (Copula-Based Outlier Detection, anomaly detection based on coupling), Auto Encoder, CBLOF (Cluster-Based Local Outlier Factor, local anomaly factor based on clustering). In some embodiments, the anomaly detection model can also be called a weak learner.
[0044] The anomaly detection strategy is the method used for anomaly detection. Furthermore, the anomaly detection strategy may be a strategy corresponding to the classification or feature clustering of resource transfer features. Different anomaly detection models correspond to different anomaly detection strategies.
[0045] In one embodiment, each anomaly detection model may be a functional module configured in a server, and the server calls each anomaly detection model to divide the resource transfer features in the resource transfer feature set to obtain a corresponding distribution result. In addition, each anomaly detection model may also be a functional module configured in different terminals or servers, and the server sends a division request to these terminals or servers to trigger these terminals or servers to divide the resource transfer features in the resource transfer feature set based on the corresponding anomaly detection model to obtain a corresponding distribution result.
[0046] S208, determining, through each anomaly detection model, a distribution result of each resource transfer feature in a resource transfer feature set corresponding to the target object in a target feature set corresponding to the feature dimension.
[0047] The distribution result is the distribution classification result of each resource transfer feature in the corresponding target feature set. Furthermore, the distribution result can be represented by a classification mark, such as abnormal, normal, etc.; it can also be represented by the probability value corresponding to each category, such as 80%, 90%, etc. In addition, the distribution result can also be information such as the distribution density of the classification probability.
[0048] The target feature set is a feature set consisting of all or part of the resource transfer features under a certain feature dimension. The resource transfer features in the target feature set may correspond to a specific feature distribution state, and the resource transfer features may be divided based on the feature distribution state to obtain a distribution result. The feature distribution state may be represented by at least one of a tree diagram, a histogram, a normal distribution diagram, and the like.
[0049] As mentioned above, different anomaly detection models can classify resource transfer features in different ways, and then determine the distribution results of each resource transfer feature in the target feature set.
[0050] In one embodiment, the anomaly detection model can determine the target feature set corresponding to a certain feature dimension, analyze and classify the distribution state of the target feature set corresponding to the feature dimension according to the corresponding anomaly detection strategy to obtain the distribution result corresponding to the feature dimension; then integrate the distribution results under each feature dimension to obtain the distribution result of each resource transfer feature in the target feature set corresponding to the feature dimension. In addition, the anomaly detection model can also perform an overall analysis of the resource transfer features under each feature dimension, for example, determine the overall distribution state of the resource transfer features under all feature dimensions, and classify based on the distribution state to obtain the overall distribution results corresponding to all feature dimensions.
[0051] S210, determining a model detection result of the anomaly detection model for the target object based on the distribution result obtained by the anomaly detection model.
[0052] The model detection result is the detection result of the anomaly detection model on the category to which the target object belongs. The model detection result of the anomaly detection model on the target object can be represented by the category identifier of the category to which it belongs. In addition, the model detection result can also carry the model identifier of each anomaly detection model. Based on this, the server can know which anomaly detection model outputs each model detection result.
[0053] The distribution results of each anomaly detection model can characterize the classification results of each anomaly detection model for each resource transfer feature. On this basis, the distribution results can be analyzed to determine the model detection results of each anomaly detection model for the target object. Furthermore, the distribution results expressed in the form of classification identification or probability value can be converted into numerical form, and the converted results are determined as the corresponding model detection results. For example: compare the probability value with the probability value threshold. When the probability value is larger, the model detection result is expressed as 1, and when the probability value is smaller, the model detection result is expressed as 0. Among them, the probability value threshold can be determined according to the actual scenario.
[0054] In one embodiment, the implementation process of S210 may be: the server determines the distribution results of the resource transfer features determined by the anomaly detection model in the target feature set corresponding to the feature dimension, performs statistics on the distribution results corresponding to each feature dimension, and obtains the distribution results of the target object under the corresponding anomaly detection model; transforms the distribution results of the target object under the corresponding anomaly detection model, and uses the transformed results as the model detection results of the anomaly detection model for the target object.
[0055] S212, collecting statistics on the model detection results of the target object by each anomaly detection model in the anomaly detection model set to obtain an anomaly detection result of the target object.
[0056] The abnormal detection result may be whether the target object is an abnormal object, and thus the abnormal detection result may be that the target object is an abnormal object, a normal object, or a suspicious object. Furthermore, the abnormal object may also be called an abnormal resource transfer object.
[0057] Among them, performing statistics on the model detection results may refer to performing statistical operations on the number of model detection results, etc., and using the results of the statistical operations as the anomaly detection results. Specifically, the anomaly detection results of the target object may be determined as an abnormal object when the number of model detection results meets the quantity condition. Among them, the quantity condition may be greater than a set quantity threshold, and the quantity threshold may be determined according to actual conditions. The process of obtaining anomaly detection results based on the model detection results corresponding to each anomaly detection model can be considered as an integrated learning process of each anomaly detection model. When the anomaly detection model is called a weak learner, the set of anomaly detection models obtained by the integrated learning of these weak learners can be called a strong learner.
[0058] In one embodiment, statistics on the model detection results can be implemented by at least one of the following methods: bagging (self-help aggregation method), boosting (boosting method), stacking (stacking method), etc.
[0059] In the above object recognition method, since the distribution of resource transfer features is analyzed in units of feature dimensions, and the model detection results corresponding to the object are obtained by combining the distribution results of each feature dimension, the resource transfer features of the object in each feature dimension can be integrated to detect anomalies of the object, so that the model detection results output by the model have high accuracy. In addition, the anomaly detection strategies adopted by the anomaly detection models are different. The model detection results corresponding to these anomaly detection models are comprehensively counted to obtain the anomaly detection results of the target object. The anomaly detection results of the target object can be determined based on multiple distribution results, which can effectively improve the accuracy of object recognition. In addition, the distribution of resource transfer features in the target feature set of the feature dimension is analyzed by the anomaly detection model. Even if there is no sample label, each anomaly detection model can obtain the distribution result. While ensuring accurate anomaly detection results, the time of obtaining sample labels is saved, which can effectively improve the efficiency of object recognition.
[0060] In one embodiment, the distribution results of each resource transfer feature in the target feature set corresponding to the feature dimension in the resource transfer feature set corresponding to the target object are determined through each anomaly detection model, including: obtaining the resource transfer features corresponding to the feature dimension in the resource transfer feature set corresponding to the target object, and obtaining the target feature sets corresponding to each feature dimension respectively; obtaining the distribution partitioning method corresponding to the target feature set in the anomaly detection model; based on the distribution partitioning method, dividing the resource transfer features in the target feature set of the feature dimension, and obtaining the distribution results of the resource transfer features in the target feature set of the feature dimension.
[0061] Among them, the distribution division method is a method of dividing the resource transfer features in the target feature set into corresponding feature intervals. The distribution division method can be a method of division based on a threshold value, or a method of division based on a distribution interval. Among them, the method of division based on a threshold value can be: comparing the resource transfer feature with the feature division threshold value, when the resource transfer feature is less than the feature division threshold value, the resource transfer feature is divided into a feature interval A, and when the resource transfer feature is greater than or equal to the feature division threshold value, the resource transfer feature is divided into another feature interval B. The method of division based on the distribution interval can be: determining multiple feature division intervals, each feature division interval corresponds to a feature value range corresponding to the resource transfer feature, comparing the feature value of the resource transfer feature with the feature value range, and then allocating the resource transfer feature to the feature division interval of the feature value. Among them, the feature value of the resource transfer feature can be a specific value corresponding to the resource transfer feature, such as: the time value corresponding to the resource transfer operation, the resource value transferred, the specific number of resource transfer operations, etc.
[0062] In one embodiment, resource transfer features corresponding to each feature dimension are selected from the resource transfer feature set corresponding to the target object. The resource transfer features under one feature dimension constitute a target feature set. The resource transfer features in the target feature set are divided based on the distribution partitioning method to divide each resource transfer feature into a corresponding feature interval. One feature interval can correspond to one distribution result, thereby obtaining the distribution result of the resource transfer feature in the target feature set of the feature dimension.
[0063] In one embodiment, the distribution partitioning method can be determined based on the anomaly detection strategy of each anomaly detection model. Further, through each anomaly detection model, the distribution result of each resource transfer feature in the target feature set corresponding to the feature dimension in the resource transfer feature set corresponding to the target object is determined, including: obtaining the resource transfer feature corresponding to the feature dimension in the resource transfer feature set corresponding to the target object, and obtaining the target feature set corresponding to each feature dimension; determining the distribution partitioning method corresponding to the target feature set in the anomaly detection model based on the anomaly detection strategy corresponding to each anomaly detection model, and dividing the resource transfer feature in the target feature set of the feature dimension based on the distribution partitioning method, and obtaining the distribution result of the resource transfer feature in the target feature set of the feature dimension.
[0064] In the above embodiment, the resource transfer features are divided into a target feature set corresponding to the feature dimension, and the distribution results under each feature dimension can be determined. Even if there is no distance between the resource transfer features, the distribution results corresponding to the resource transfer features can be accurately determined, and the distribution results of the transfer features can be quickly determined, thereby effectively improving the efficiency of object recognition.
[0065] In one embodiment, the distribution partitioning method corresponding to the target feature set includes a partitioning method based on a threshold value, and the distribution partitioning method corresponding to the target feature set in the anomaly detection model is obtained; based on the distribution partitioning method, the resource transfer feature is partitioned in the target feature set of the feature dimension, and the distribution result of the resource transfer feature in the target feature set of the feature dimension is obtained, including: obtaining a feature distribution structure tree, the feature distribution structure tree includes multiple child nodes; taking the initial node of the feature distribution structure tree as the current child node corresponding to the target object, obtaining the current feature dimension corresponding to the current child node, and obtaining the current feature partitioning threshold of the current target feature set corresponding to the current feature dimension; based on the current feature partitioning threshold and the resource transfer feature of the target object in the current feature dimension, determining the distribution result of the resource transfer feature in the current target feature set; determining the next child node corresponding to the target object based on the distribution result, taking the next node as the updated current child node, returning to the step of obtaining the current feature dimension corresponding to the current child node, and obtaining the current feature partitioning threshold of the current target feature set corresponding to the current feature dimension, until the child node corresponding to the target object is updated.
[0066] The feature distribution structure tree is a bifurcated tree constructed based on the resource transfer feature, and can be at least one of an isolated tree or a random tree. Each resource transfer feature can be used as a node in the feature distribution structure tree. The feature distribution structure tree can be used as an anomaly detection model.
[0067] The number of feature distribution structure trees can be at least one. When the number of feature distribution structure trees is multiple, all or part of the feature distribution structure trees can be used together as an anomaly detection model. These feature distribution structure trees in the anomaly detection model can divide the resource transfer features in parallel to divide the resource transfer features corresponding to the target object into corresponding child nodes. Further, the resource transfer feature set corresponding to a certain target object can be input into each feature distribution structure tree respectively, and each feature distribution structure tree outputs the distribution result of the target object, and the distribution results of each feature distribution structure tree are integrated to obtain the overall distribution result corresponding to the feature distribution structure tree as the distribution result of the corresponding anomaly detection model. In the same way, the resource transfer feature sets corresponding to other target objects are input into each feature distribution structure tree respectively to obtain the corresponding distribution results.
[0068] The current feature partitioning threshold may be predetermined or determined based on the feature value of the resource transfer feature of the target object under the current feature dimension. For example, at least one of the mean, median or variance of the feature value of the resource transfer feature of the target object under the current feature dimension may be used as the current feature partitioning threshold.
[0069] The judgment node after the subnode update is completed may be that the subnode corresponding to the resource transfer feature is a leaf node, that is, there is no next node. Further, the distribution result corresponding to the leaf node where the resource transfer feature is located can be determined as the distribution result in the target feature set of the feature dimension where the resource transfer feature is located.
[0070] In one embodiment, based on the current feature partitioning threshold and the resource transfer feature of the target object in the current feature dimension, the implementation process of determining the distribution result of the resource transfer feature in the current target feature set can be: comparing the resource transfer feature of the target object in the current feature dimension with the current feature partitioning threshold; when the resource transfer feature of the target object in the current feature dimension is less than the current feature partitioning threshold, the resource transfer feature of the target object in the current feature dimension is divided into the first node; when the resource transfer feature of the target object in the current feature dimension is greater than or equal to the current feature partitioning threshold, the resource transfer feature of the target object in the current feature dimension is divided into the second node. Afterwards, taking the first node as an example, the next child node of the first node is used as the updated current child node, and the above process is repeated. The second node is the same, and it will not be repeated here.
[0071] Specifically, taking resource transfer characteristics as the frequency of resource transfer operations, the object characteristics of the receiving object, and the transferred resource value as examples, each resource transfer characteristic is used as a characteristic dimension; in the first-level division process, the frequency of the resource transfer operation is compared with the frequency threshold to divide the resource transfer characteristics into the corresponding first node and the second node; in the second-level division process, taking one side as an example, the object characteristics of the receiving object are compared with the object characteristic attributes to divide the resource transfer characteristics in the first node into the third node and the fourth node; in the third-level division process, taking one side as an example, the transferred resource value is compared with the resource value threshold to divide the resource transfer characteristics in the third node into the fifth node and the sixth node. At this time, the child node corresponding to the resource transfer characteristic is a leaf node, and it is determined that the node update is completed.
[0072] In the above embodiment, the current feature dimension corresponding to the current child node is gradually obtained, and then the resource transfer characteristics of the target object in the current feature dimension are divided based on the current feature division threshold. Each division of the resource transfer characteristics can be considered as the completion of a level of division. The feature dimensions of each level are interrelated and progressive, so accurate distribution results can be obtained, and then accurate object recognition results can be obtained.
[0073] In one embodiment, a model detection result of the anomaly detection model for a target object is determined based on a distribution result obtained by the anomaly detection model, including: determining a child node corresponding to the target object based on the distribution result; counting the number of child nodes corresponding to the target object to obtain a path length of the target object in a feature distribution structure tree; determining a first anomaly detection value corresponding to the target object based on the path length, the anomaly detection value being negatively correlated with the path length; and determining the model detection result of the anomaly detection model for the target object based on the first anomaly detection value.
[0074] The child nodes corresponding to the target object may be nodes from the root node to the leaf nodes of the resource transfer feature of the target object, and the number of these child nodes may be determined as the path length of the target object in the feature distribution structure tree.
[0075] The first abnormality detection value is a detection value capable of evaluating whether the target object is an abnormal resource transfer object.
[0076] In one embodiment, the implementation process of determining the first abnormality detection value corresponding to the target object based on the path length may be: when there is one feature distribution structure tree, determine the path length corresponding to the target object determined by the feature distribution structure tree, construct an exponential function with the path length as the exponent and a preset constant value as the base, substitute the path length corresponding to the target object into the exponential function, and the obtained function value is the first abnormality detection value. When there are multiple feature distribution structure trees, determine the expected value of the path length corresponding to the target object based on the path length, construct an exponential function with the expected value of the path length as the exponent and a preset constant value as the base, substitute the expected value of the path length corresponding to the target object into the exponential function, and the obtained function value is the first abnormality detection value.
[0077] The corresponding anomaly detection value is obtained based on the path length of the subnode corresponding to the target object in the feature distribution structure tree. The shorter the path length, that is, the fewer times it needs to be segmented, the farther the resource transfer feature corresponding to the target object is from the normal data point, and the larger the anomaly detection value. Furthermore, in order to ensure that the anomaly detection value is negatively correlated with the path length, the exponent in the exponential function can be negative.
[0078] In one embodiment, the implementation process of determining the model detection result of the target object by the anomaly detection model based on the first anomaly detection value may be: comparing the first anomaly detection value with the first anomaly detection value threshold, and when the first anomaly detection value is greater than the first anomaly detection value threshold, determining the model detection result of the target object as an abnormal resource transfer object. The first anomaly detection value threshold may be a predetermined fixed value, or may be calculated based on the feature value of the resource transfer feature in the target feature set.
[0079] Specifically, the implementation process of determining the first anomaly detection value corresponding to the target object based on the path length may be:
[0080] A feature distribution tree is constructed based on n training samples, and the average path length of the feature distribution tree is determined by the following formula:
[0081] c(n)=2H(n-1)-(2(n-1) / n)
[0082] Where H(i) is the harmonic number, which can be estimated as ln(i)+0.5772156649.
[0083] For the resource transfer feature x of the target object, the first anomaly detection value corresponding to the target object is determined by the following formula:
[0084]
[0085] Where E(h(x)) is the expected path length of sample x in a batch of feature distribution structure trees.
[0086] The method for determining anomaly detection values in the above-mentioned embodiment can achieve reliable anomaly detection based on the node division of the feature distribution structure tree even if the distance or density between resource transfer features is not determined. Compared with distance and density calculations, the computational consumption is greatly reduced, and it has the advantages of near-linear complexity and low memory consumption.
[0087] In one embodiment, the process of constructing the feature distribution structure tree may be to construct multiple feature distribution structure trees based on multiple training samples. The training samples may not have corresponding labels, and the feature distribution structure tree is constructed in an unsupervised manner. The process of constructing the feature distribution structure tree is specifically described as follows:
[0088] Given n sample data X = {x_1, x_2, ..., x_n}, these n sample data are resource transfer features under d dimensions. Randomly select a resource transfer feature q and its split value p, recursively split the data set X, that is, divide the sample data corresponding to the current child node into two or more nodes based on the split value p, until any of the following conditions is met: 1. The tree reaches the limit height, 2. There is only one sample on the node, 3. All features of the samples on the node are the same.
[0089] Further, assuming that T is a node of the feature distribution structure tree, T can be a leaf node or an internal node with child nodes (T_l, T_r). In one embodiment, the process of gradually segmenting the resource transfer features in a feature set can be as follows: Figure 3As shown. For each step of segmentation, the resource transfer feature q and the segmentation value p in the feature dimension are determined. Based on the segmentation value p, the corresponding resource transfer feature q is divided into the corresponding interval, and the resource transfer features within the interval are further segmented. The segmentation line corresponding to p is as shown in Figure 3 the dashed line in. If a certain resource transfer feature q < p, then this resource transfer feature is assigned to T_l. If a certain resource transfer feature q ≥ p, then this resource transfer feature is assigned to T_r. The process of successive segmentation can be as shown in Figure 3 shown. To save space, Figure 3 only the partitioning process on one side is shown in.
[0090] Furthermore, for the resource transfer feature x of the target object, the path length of the sample point x is determined, that is, the number of edges passed from the root node to the leaf node of the feature distribution structure tree is determined. The path length can be determined by using the binary search method. Based on the constructed feature distribution structure tree, a set of feature distribution structure trees is output, which is the feature distribution structure forest. Then, the first anomaly detection value corresponding to the target object can be determined based on the feature distribution structure forest.
[0091] In one embodiment, the distribution partitioning method includes a method of partitioning based on distribution intervals, obtaining the distribution partitioning method corresponding to the target feature set in the anomaly detection model; based on the distribution partitioning method, partitioning the resource transfer features in the target feature set of the corresponding feature dimension to obtain the distribution result of the resource transfer features in the target feature set of the corresponding feature dimension, including: obtaining the set of feature partitioning intervals corresponding to the target feature set in the anomaly detection model, where the set of feature partitioning intervals includes multiple feature partitioning intervals; obtaining the number of features of the resource transfer features in each feature partitioning interval in the target feature set; determining the distribution density corresponding to the feature partitioning interval based on the number of features, and using the distribution density as the distribution result of the resource transfer features in the target feature set of the corresponding feature dimension.
[0092] Among them, each feature partitioning interval corresponds to a range of feature values of the resource transfer feature. The widths of the ranges of feature values of each feature partitioning interval can be the same or different. The distribution density can include probability density, etc.
[0093] In one embodiment, the target feature sets corresponding to each feature dimension can be determined respectively. Based on the resource transfer features in the target feature set, the distribution density corresponding to the feature partitioning interval is determined, and the respective distribution results are obtained. After obtaining the distribution results corresponding to each feature dimension, these distribution results can be statistically analyzed, and the overall distribution result of the resource transfer features of the target object can be obtained based on the statistical results.
[0094] In the above-mentioned embodiment, the resource transfer features in the target feature set are divided based on multiple feature division intervals, and the corresponding distribution density is determined based on the number of features of the resource transfer features in each feature division interval, thereby obtaining a distribution result. Even without label information, the resource transfer features of the target object can be accurately divided based on the distribution density, thereby obtaining an accurate distribution result.
[0095] In one embodiment, a model detection result of the anomaly detection model for the target object is determined based on the distribution result obtained by the anomaly detection model, including: determining a feature anomaly detection value corresponding to the resource transfer feature based on the distribution density, the feature anomaly detection value is negatively correlated with the distribution density; statistically analyzing the feature anomaly detection values corresponding to each resource transfer feature in the resource transfer feature set corresponding to the target object to obtain a second anomaly detection value corresponding to the target object; and determining the model detection result of the anomaly detection model for the target object based on the second anomaly detection value.
[0096] Among them, the characteristic anomaly detection value is a detection value that can evaluate whether the target object is an abnormal resource transfer object. Further, the characteristic anomaly detection value can be obtained by performing a specific statistical operation on the characteristic anomaly detection value corresponding to the resource transfer feature determined by the distribution density, for example: performing a reciprocal processing, performing an exponential operation, etc. Specifically, in order to ensure that the characteristic anomaly detection value is negatively correlated with the distribution density, the reciprocal of the distribution density can be determined as the characteristic anomaly detection value.
[0097] The second abnormality detection value is a detection value capable of evaluating whether the target object is an abnormal resource transfer object.
[0098] Furthermore, after determining the feature anomaly detection value corresponding to each resource transfer feature in the resource transfer feature set corresponding to the target object, the feature anomaly detection value corresponding to each feature dimension can be determined, and the feature anomaly detection values under each feature dimension are counted to obtain a second anomaly detection value corresponding to the target object.
[0099] Specifically, for the resource transfer feature set p corresponding to the target object, the feature anomaly detection value can be expressed as the probability density hist i (p). When the feature dimension is d, the second anomaly detection value corresponding to the target object can be determined by the following formula:
[0100]
[0101] In one embodiment, the implementation process of determining the model detection result of the target object by the anomaly detection model based on the second anomaly detection value may be: comparing the second anomaly detection value with the second anomaly detection value threshold, and when the second anomaly detection value is greater than the second anomaly detection value threshold, determining the model detection result of the target object as an abnormal resource transfer object. The second anomaly detection value threshold may be a predetermined fixed value, or may be calculated based on the feature value of the resource transfer feature in the target feature set.
[0102] In the above embodiment, an anomaly detection value is obtained based on the distribution density, and then a distribution result corresponding to the anomaly detection model is obtained based on the anomaly detection value. Even without label information, the resource transfer characteristics of the target object can be accurately divided based on the distribution density, and the resource transfer characteristics with a small distribution density can be determined, thereby obtaining an accurate distribution result.
[0103] In one embodiment, a distribution partitioning method corresponding to a target feature set in an anomaly detection model is obtained; based on the distribution partitioning method, a resource transfer feature is partitioned in a target feature set of a feature dimension, and a distribution result of the resource transfer feature in the target feature set of the feature dimension is obtained, including: determining an end partitioning threshold; determining a tail feature set corresponding to a feature tail and a head feature set corresponding to a feature head in the target feature set based on the end partitioning threshold; performing statistics on the tail feature set to obtain a tail feature statistic corresponding to the feature dimension set; determining a tail distribution possibility of the resource transfer feature based on the tail feature statistic corresponding to the feature dimension set; performing statistics on the head feature set to obtain a head feature statistic corresponding to the feature dimension set; determining a head distribution possibility of the resource transfer feature based on the head feature statistic corresponding to the feature dimension set; determining a target distribution possibility corresponding to the resource transfer feature based on the tail distribution possibility and the head distribution possibility, and taking the target distribution possibility as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
[0104] The end division threshold may be a threshold for dividing the resource transfer features at the edge of the target feature set. It may include a left division threshold and a right division threshold. The left division threshold may also be called a head division threshold, which is used to divide the resource transfer features at the left end; the right division threshold may also be called a tail division threshold, which is used to divide the resource transfer features at the right end. Furthermore, the end division threshold may be a predetermined fixed value, or may be calculated based on the feature value of the resource transfer features in the target feature set.
[0105] In one embodiment, resource transfer features in the target feature set that are less than the head partition threshold may be used as the head feature set, and resource transfer features in the target feature set that are greater than or equal to the tail partition threshold may be used as the tail feature set.
[0106] In one embodiment, the tail feature statistic may be an average value calculation of the feature values of the resource transfer features in the tail feature set, and the result of the average value calculation is used as the tail feature statistic. The head feature statistic may be an average value calculation of the feature values of the resource transfer features in the head feature set, and the result of the average value calculation is used as the tail feature statistic. In addition, when calculating the tail feature statistic and the head feature statistic, other operations may also be performed on the feature values of the resource transfer features in the tail feature set and the head feature set.
[0107] The tail distribution possibility is the possibility of a resource transfer feature being distributed to the tail feature set, which can be expressed as a probability value; the head distribution possibility is the possibility of a resource transfer feature being distributed to the head feature set, which can also be expressed as a probability value. In addition, the head distribution possibility and the tail distribution possibility can be collectively referred to as the tail probability.
[0108] In one embodiment, the implementation process of determining the target distribution possibility corresponding to the resource transfer feature based on the tail distribution possibility and the head distribution possibility can be: performing certain operations on the tail distribution possibility and the head distribution possibility, and determining the operation result as the target distribution possibility, for example: performing average value operations on the tail distribution possibility and the head distribution possibility, and using the calculated average value as the target distribution possibility. In another embodiment, it is also possible to select one from the tail distribution possibility and the head distribution possibility according to certain rules as the target distribution possibility, for example: determining the larger value of the tail distribution possibility and the head distribution possibility as the target distribution possibility.
[0109] In the above embodiment, an abnormal detection value is obtained based on the end distribution, and then a distribution result is obtained. Even without label information, the resource transfer characteristics of the target object can be accurately divided based on the end distribution in the normal distribution, and the resource transfer characteristics with low probability can be determined, thereby obtaining an accurate distribution result.
[0110] In one embodiment, a target distribution possibility corresponding to a resource transfer feature is determined based on a tail distribution possibility and a head distribution possibility, and the target distribution possibility is used as a distribution result of the resource transfer feature in a target feature set of the feature dimension in which the resource transfer feature is located, including: obtaining an average value of each resource transfer feature in the resource transfer feature set, and determining a skewness factor corresponding to the target object based on the average value corresponding to the resource transfer feature set; obtaining a skewness factor reference value, and determining a relative size of the skewness factor and a skewness factor reference value; determining a skewness distribution possibility from a tail distribution possibility and a head distribution possibility based on a relative size of the skewness factor and a skewness factor reference value; comparing the tail distribution possibility, the head distribution possibility, and the skewness distribution possibility, and determining a target distribution possibility corresponding to the resource transfer feature from the tail distribution possibility, the head distribution possibility, and the skewness distribution possibility according to the size comparison result; and using the target distribution possibility as a distribution result of the resource transfer feature in the target feature set of the feature dimension in which the resource transfer feature is located.
[0111] The skewness factor reference value is a value used for reference evaluation of the skewness factor, and its value can be determined according to actual conditions and can be 0 or other values.
[0112] In one embodiment, when the feature dimension has d dimensions, the skewness factor corresponding to a resource transfer feature of the target object may be calculated by the following formula:
[0113]
[0114] Where n represents the number of resource transfer features in the resource transfer feature set. i Represents a resource transfer feature in the resource transfer feature set. It is the average value corresponding to each resource transfer feature in the resource transfer feature set.
[0115] In one embodiment, the process of determining the tail distribution possibility, the head distribution possibility and the skewness distribution possibility may be:
[0116] For a resource transfer feature x in the resource transfer feature set i , the header feature statistics are expressed as F dl (x i ), the tail feature statistics are expressed as F dr (x i ). The coupling detection value is determined by the following formula:
[0117] U d,i =F dl (x i )
[0118] Vd,i =F dr (x i )
[0119] W d,i =U d,i , b d <0;W d,i =V d,i , b d ≥0
[0120] Among them, b d is the skewness factor corresponding to the dth feature dimension. d,i When the skewness factor is less than 0, it means that the distribution of the target feature set is skewed to the left. In this case, the first feature statistics are used as W d,i ; When the skewness factor is greater than or equal to 0, it means that the distribution of the target feature set is skewed to the right. In this case, the tail feature statistics are used as W d,i .
[0121] Based on the coupling detection value, the first distribution probability p is determined by the following formula l , tail distribution possibility p r and the probability of skewness distribution p s :
[0122]
[0123]
[0124]
[0125] In one embodiment, the implementation process of determining the target distribution possibility corresponding to the resource transfer feature from the tail distribution possibility, the head distribution possibility and the skewness distribution possibility according to the size comparison result can be: determining the maximum value among the tail distribution possibility, the head distribution possibility and the skewness distribution possibility as the target distribution possibility corresponding to the resource transfer feature.
[0126] Furthermore, the first distribution probability p l , tail distribution possibility p r and the probability of skewness distribution p s Compare the size to distribute the first part to the possible degree p l , tail distribution possibility p r and the probability of skewness distribution p s The maximum value among is taken as the target distribution possibility.
[0127] In the above embodiment, the corresponding distribution possibility is determined based on the relative size of the skewness factor and the skewness factor reference value on each feature dimension, and the accurate distribution possibility can be determined based on the center deviation of the target feature set, thereby obtaining an accurate and reliable distribution result.
[0128] In one embodiment, the model detection results of each anomaly detection model in the anomaly detection model set on the target object are counted to obtain the anomaly detection result of the target object, including: determining the number of abnormal results whose model detection results are abnormal according to the model detection results of the target object; when the number of abnormal results exceeds the abnormal number threshold, determining the target object as an abnormal resource transfer object, and obtaining the anomaly detection result of the target object. The abnormal model detection result means that the model detection result shows that the target object is an abnormal object.
[0129] The size of the abnormal quantity threshold can be determined according to actual conditions, and can be a pre-set fixed threshold, or can be determined according to the number of target objects, for example: multiplying the number of target objects by a fixed coefficient, and using the product as the abnormal quantity threshold.
[0130] In one embodiment, when the number of abnormal results exceeds the abnormal number threshold, the server may determine that the abnormal detection model exceeding the abnormal number threshold identifies the target object as an abnormal object, and then determine that the target object is an abnormal resource transfer object.
[0131] In one embodiment, when there are multiple target objects, if the number of abnormal results exceeds the abnormal number threshold, all target objects can be determined as abnormal resource transfer objects, or the target objects corresponding to the number of abnormal results exceeding the abnormal number threshold can be determined as abnormal resource transfer objects.
[0132] In the above embodiment, when the model detection result that exceeds the anomaly number threshold is determined to be abnormal, the target object is determined as an abnormal resource transfer object, and the results of multiple anomaly detection models are integrated to obtain an anomaly detection result. Compared with the anomaly detection result obtained by a single anomaly detection model, the result obtained in this way has higher accuracy.
[0133] In one embodiment, the model detection result can be represented by a probability value that the target object is an abnormal object. Further, the model detection results of each abnormal detection model in the abnormal detection model set for the target object are counted to obtain the abnormal detection result of the target object, including: based on the model detection results of each abnormal detection model for the target object, determining the model probability of each abnormal detection model for the target object as an abnormal object; respectively obtaining the comparison information of the model probability of each abnormal detection model relative to the corresponding probability threshold, and converting the model probability of each abnormal detection model into a voting result that the target object is an abnormal object based on the comparison information; and counting the voting results corresponding to each abnormal detection model to obtain the abnormal detection result of the target object.
[0134] Among them, the size of the probability threshold can be determined according to the actual situation, which can be a pre-set fixed threshold, or it can be determined according to the model probability corresponding to each anomaly detection model, for example: the average value of the model probability corresponding to each anomaly detection model is used as the probability threshold. Further, the rough range of the probability threshold can be determined, and each probability threshold within the rough range of the probability threshold constitutes a candidate probability threshold, and then the candidate probability thresholds are traversed one by one, and the anomaly detection result of the object is determined based on the selected candidate probability threshold, and the anomaly detection result is compared with the label of the corresponding object in the label database. If the comparison between the two is consistent, for example: both represent that the corresponding object is an abnormal resource transfer object, then it is determined that the selected candidate probability threshold is selected appropriately and is used as the probability threshold of the anomaly detection model; and if the comparison between the two is inconsistent, it is determined that the selected candidate probability threshold is inappropriate, and the next candidate probability threshold is traversed until the selected candidate probability threshold is appropriate. By traversing, a suitable probability threshold can be determined from multiple candidate probability thresholds, and then an accurate and reliable anomaly detection model can be obtained based on the selected probability threshold.
[0135] In one embodiment, counting the voting results may be determining the number of votes for abnormal objects. When the number of votes is greater than a voting number threshold, the abnormal detection result of the target object is determined as an abnormal resource transfer object. When the number of votes is less than or equal to the voting number threshold, the abnormal detection result of the target object is determined as a normal object. The voting number threshold may be determined based on actual conditions. It may be a pre-set fixed threshold or may be determined based on the number of abnormal detection models. For example, the sum of the number of abnormal detection models is determined as the voting number threshold. In this case, the server will determine the target object as an abnormal resource transfer object only when the voting results corresponding to all abnormal detection models are abnormal objects.
[0136] In the above embodiment, the voting results corresponding to each anomaly detection model are determined, and these voting results are counted, so that the voting information of multiple anomaly detection models can be integrated to obtain accurate anomaly detection results.
[0137] In one embodiment, after the model detection results of the target object by each anomaly detection model in the anomaly detection model set are counted to obtain the anomaly detection result of the target object, the method further includes: obtaining the anomaly label corresponding to the target object in the label database; comparing the anomaly detection result of the target object with the anomaly label corresponding to the target object, and determining the detection gain result corresponding to the anomaly detection model set according to the comparison result; the detection gain result includes candidate objects whose anomaly detection results are abnormal and whose anomaly labels are normal; determining the object classification result of the candidate object based on the resource transfer feature set corresponding to the candidate object through the object classification model; and determining the detection gain evaluation result of the anomaly detection model set relative to the label database based on the object classification result of the candidate object.
[0138] The label database contains labels of whether each object in the application is an abnormal object, that is, whether each object is an abnormal resource transfer object. For abnormal objects, the corresponding labels can be called black labels, and for normal objects, the corresponding labels can be called white labels.
[0139] In some embodiments, the label database may also only include labels corresponding to abnormal objects. For example: when an object is marked as an abnormal object, the object identifier corresponding to the object is added to the label database, and each object identifier is used as an abnormal label. In this case, the anomaly detection result can be compared with each object identifier in the label database. When it is determined that there is a matching object identifier, it is determined that both the anomaly detection model set and the label database identify the corresponding object as an abnormal object.
[0140] In one embodiment, the process of determining abnormal labels in the label database can be achieved in at least the following ways: 1. When a complaint operation against a reference object is received, the label of the reference object is replaced from normal to abnormal; 2. Feedback information of each object in the application is obtained, and when it is determined that there is report information against the reference object in the feedback information, the label of the reference object is replaced from normal to abnormal; 3. Operation records of each object in the application are obtained, abnormal object identification rules are obtained, and the operation records of each object are compared with the abnormal object identification rules. When the operation record of the object matches the abnormal object identification rules, the label of the corresponding object is replaced from normal to abnormal.
[0141] In some embodiments, the resource transfer features of the reference object can also be classified based on a supervised learning neural network model, etc., and the label of the reference object whose classification result is an abnormal object is determined as an abnormal object and added to the label database. Furthermore, after the abnormal detection model set obtains the abnormal detection result, the abnormal label in the label database is obtained, and the detection accuracy of the abnormal detection model set is evaluated based on the abnormal label. Among them, the abnormal detection model set is obtained based on unsupervised learning, and the abnormal label in the label database is obtained based on supervised learning. The abnormal label obtained based on supervised learning has a higher accuracy. In this case, the accuracy of the abnormal detection model based on unsupervised learning can be verified based on the label obtained by supervised learning, and a reliable accuracy evaluation result can be obtained.
[0142] The detection gain evaluation result can characterize the accuracy of the anomaly detection model set relative to the label database. The process of determining the detection gain evaluation result can be: determining the number of objects accurately identified by the anomaly detection model set but not accurately identified by the label database, determining the ratio of the number to the total number of target objects, and determining the ratio as the detection gain evaluation result.
[0143] The object classification model is a model that can determine the category of an object, and can be a neural network model, etc. Furthermore, the object classification model can be a supervised learning model. In one embodiment, the resource transfer feature set corresponding to the candidate object is input into the object classification model, and the object classification model extracts and processes the resource transfer feature set, and outputs the object classification result.
[0144] In one embodiment, the server can output a classification request to the terminal device, which carries a resource transfer feature set corresponding to the candidate object, so as to prompt the object using the terminal device to review, check, classify, etc. the candidate object based on the resource transfer feature set corresponding to the candidate object, and then the terminal device receives the object classification result of the candidate object input by the object.
[0145] In one embodiment, the anomaly detection results obtained by the anomaly detection model set are compared with the anomaly labels in the label database to obtain candidate objects that are identified as abnormal by the anomaly detection model but normal by the label database, and the candidate objects are further analyzed. If it is determined that the candidate object is indeed an abnormal object, it is determined that the anomaly detection model set more accurately determines the abnormal detection result of the candidate object, and there is a detection gain relative to the label database. Further, when there are multiple candidate objects, the number of objects that are indeed abnormal among the candidate objects is determined, and a ratio operation is performed on the number of objects and the total number of target objects, and the ratio operation result is determined as the detection gain evaluation result.
[0146] Specifically, the process of comparing the anomaly detection model set with the label database can be as follows: Figure 4 shown. Figure 4 In the example, the anomaly detection model set includes a first anomaly detection model, a second anomaly detection model, and a third anomaly detection model. These anomaly detection models output corresponding model detection results respectively. After that, these model detection results are integrated and learned, that is, these model detection results are statistically analyzed, and anomaly detection results are obtained according to the statistical results. In addition, the anomaly labels in the label database are compared with the anomaly detection results (versus, VS). If the comparison result is consistent, it means that both the anomaly detection result and the label database consider the corresponding object to be an abnormal resource transfer object. At this time, the anomaly detection result can be determined to be an accurate result.
[0147] In the above embodiment, based on the comparison result of the anomaly detection result of the anomaly detection model set with the anomaly label in the label database, the detection gain evaluation result of the anomaly detection model set relative to the label database is determined. Based on the detection gain evaluation result of the actual experiment, it can be found that the anomaly detection model set can obtain more accurate anomaly detection results relative to the label database.
[0148] In one embodiment, after the model detection results of each anomaly detection model in the anomaly detection model set on the target object are counted and the anomaly detection result of the target object is obtained, the method further includes: determining a first time when the anomaly detection result of the target object is obtained; obtaining a target anomaly label that matches the anomaly detection result in a label database; the anomaly label included in the label database is an object label generated based on response information of a resource transfer feature set of the target object, and the anomaly label is added to the label database after generation; determining a second time when the target anomaly label is added to the label database; when the first time is less than the second time, determining an early detection evaluation result of the anomaly detection model set relative to the label database based on the difference between the first time and the second time.
[0149] The difference between the first time and the second time may be the time difference between the first time and the second time. In addition, the number of times the first time is earlier than the second time may be determined, and the ratio operation may be performed on the number and the total number of target objects, and the ratio operation result may be used as the early detection evaluation result.
[0150] When the first time is less than the second time, it is determined that the anomaly detection model set detects the corresponding abnormal object earlier than the label database.
[0151] In one embodiment, the object identification method also includes: obtaining the abnormal label marking time of the target object, the abnormal label marking time is the operation time of the label marking object performing the abnormal label marking operation on the target object; when it is determined that the abnormal detection result of the target object is an abnormal resource transfer object, determining the abnormal detection time of obtaining the abnormal detection result of the target object; determining the time difference between the abnormal label marking time and the abnormal detection time; based on the time difference, obtaining the resource transfer permission restriction time, the resource transfer permission restriction time is when the object is detected as an abnormal resource transfer object using an abnormal detection model, the resource transfer permission restriction time is the restriction time of the resource transfer permission of the abnormal resource transfer object.
[0152] The label annotation object is an object that adds abnormal labels to various objects in the application. When the abnormal label is generated based on a complaint operation, the label annotation object may be the user who triggers the complaint operation; when the abnormal label is generated based on feedback information, the label annotation object may be the user who triggers the generation of feedback information. Furthermore, when the label annotation object triggers a complaint operation, the time of the complaint operation is determined as the abnormal label annotation time; when the label annotation object triggers a feedback information generation operation, the time of the feedback information generation operation is determined as the abnormal label annotation time.
[0153] The time difference may be a difference operation between the anomaly labeling time and the anomaly detection time, and the obtained difference is used as the time difference.
[0154] The resource transfer permission restriction duration is the duration of resource transfer permission restriction on the abnormal resource transfer object. Resource transfer permission restriction can be a restriction on all operations of the abnormal resource transfer object in the application; it can also be a restriction on sensitive operations involving account security of the abnormal resource transfer object in the application, for example: restricting the transfer of resource values from the account, and restricting other accounts from transferring resource values to the account corresponding to the abnormal resource transfer object.
[0155] In one embodiment, the implementation process of obtaining the resource transfer permission restriction duration based on the time difference may be: obtaining a time coefficient, multiplying the time difference by the time coefficient, and obtaining the product as the resource transfer permission restriction duration. The size of the time coefficient may be greater than or equal to 1. In addition, the time difference corresponding to each object in the application may be determined, and these time differences may be statistically calculated, for example, the average value of these time differences may be calculated, and the calculated result may be used as the resource transfer permission restriction duration.
[0156] In one embodiment, the time difference between the abnormal label marking time and the abnormal detection time is determined; the implementation process of obtaining the resource transfer permission restriction time based on the time difference can be: when the abnormal detection time is earlier than the abnormal label marking time, the time difference between the abnormal label marking time and the abnormal detection time is determined; the resource transfer permission restriction time is obtained based on the time difference. In this way, when the abnormal detection model set detects the corresponding abnormal resource transfer object earlier than the label database, the resource transfer permission of the abnormal resource transfer object can be restricted. It can effectively solve the lag problem caused by permission restriction based on the label database and effectively ensure the security of the account.
[0157] In one embodiment, when the abnormal detection model determines that an abnormal resource transfer object has been identified and there is no abnormal label corresponding to the abnormal resource transfer object in the label database, the resource transfer authority of the abnormal resource transfer object is restricted, and a timer corresponding to the resource transfer authority restriction duration is set. The abnormal labels in the label database are detected within the timer. When it is determined that the label database has added an abnormal label corresponding to the abnormal resource transfer object during this period of time, the timer is canceled, and the resource transfer authority of the abnormal resource transfer object continues to be restricted; when it is determined that the label database has not added an abnormal label corresponding to the abnormal resource transfer object during this period of time, when the timer's timing is reached, the restriction on the resource transfer authority of the abnormal resource transfer object is canceled, and the corresponding object restores the corresponding resource transfer authority.
[0158] In the above embodiment, the resource transfer permission limit duration is set based on the time difference between the abnormal label marking time and the abnormal detection time, so as to limit the resource transfer permission of the abnormal resource transfer object within the resource transfer permission limit duration, which can effectively ensure the security of resource transfer.
[0159] In one embodiment, a resource transfer feature set corresponding to each target object is obtained, including: obtaining a resource transfer record set corresponding to the target object, the resource transfer record set including multiple resource transfer records; taking the resource transfer direction as a statistical dimension, determining feature dimensions corresponding to multiple resource transfer directions; and performing statistics on resource transfer attribute information in the same feature dimension in the resource transfer record set to obtain resource transfer features corresponding to each feature dimension, wherein the resource transfer features corresponding to each feature dimension constitute the resource transfer feature set corresponding to the target object.
[0160] The resource transfer direction is resource transfer-in or resource transfer-out. The resource transfer direction is relative to the target object. For example, if the corresponding resource value is transferred from the account corresponding to the target object to another account, the resource transfer direction is transfer-out, which can also be called outgoing payment; if the corresponding resource value is transferred from another account to the account corresponding to the target object, the resource transfer direction is transfer-in, which can also be called incoming payment.
[0161] Resource transfer attribute information is attribute information related to resource transfer, and may include at least one of transfer account information, transfer time, transfer frequency, or the value of the transferred resource.
[0162] Furthermore, for the outgoing direction, the resource transfer feature corresponding to each feature dimension may be at least one of the following: the number of outgoing transactions, the outgoing amount, the number of outgoing transactions by time period, the outgoing amount by time period, or the characteristic information of the counterparty. For the incoming direction, the resource transfer feature corresponding to each feature dimension may be at least one of the following: the number of incoming transactions, the incoming amount, the number of incoming transactions by time period, the incoming amount by time period, or the characteristic information of the counterparty.
[0163] In one embodiment, the process of obtaining resource transfer characteristics and performing anomaly detection can be as follows: Figure 5 shown. Figure 5 In the feature engineering part, the server determines the outgoing and incoming features as the resource transfer feature set. The resource transfer feature set is input into each anomaly detection model in the anomaly detection model set. These anomaly detection models output corresponding model detection results respectively, and then these model detection results are integrated and learned to obtain the anomaly detection result.
[0164] Furthermore, if Figure 6 The following is a schematic diagram of the model detection results of an embodiment. The model detection results obtained by inputting the outgoing account features into the anomaly detection model set can be as follows: Figure 6 As shown in (a), the server identified 74 network accounts as abnormal resource transfer network accounts based on the anomaly detection model set. After comparing these abnormal resource transfer network accounts with the abnormal labels in the label database, the results are as follows: Figure 6As shown in the box on the right of (a), the black squares are the number and proportion of black labels, the white squares are the number and proportion of white labels, and the gray ones are the number of labels that have not yet been matched. Among them, 5 network accounts are considered to be abnormal resource transfer network accounts by the anomaly detection results, and there are corresponding abnormal labels in the label database, that is, 6.7% of the anomaly detection results match the label database; 41 network accounts are considered to be abnormal resource transfer network accounts by both the anomaly detection results and the label database, that is, 55.4% of the anomaly detection results are results whose accuracy is pending; 28 network accounts are not considered to be abnormal resource transfer network accounts by the anomaly detection results, but there are corresponding abnormal labels in the label database, that is, 37.8% of the anomaly detection results do not match the label database. Afterwards, the 41 network accounts in the results whose accuracy is pending are further analyzed to determine the categories of these network accounts. The model detection results obtained by inputting the account features into the anomaly detection model set can be as follows Figure 6 As shown in (b) of the figure, the model detection result obtained by inputting the outgoing and incoming features into the anomaly detection model set can be as follows: Figure 6 As shown in (c), the processing process is as described above and will not be repeated here.
[0165] In one embodiment, after obtaining the resource transfer features corresponding to each feature dimension, the resource transfer features under each feature dimension can be transformed into feature vectors, and the transformed feature vectors can be determined as resource transfer feature sets. Alternatively, each resource transfer feature can be transformed into a feature vector, and the transformed feature vectors can be determined as resource transfer feature vectors, and then the resource transfer feature set can be obtained based on each resource transfer feature vector.
[0166] In the above embodiment, resource transfer features are obtained based on resource transfer direction and feature dimension, and then a resource transfer feature set is obtained. More comprehensive resource transfer features can be obtained from multiple directions and dimensions, and accurate and reliable anomaly detection results can be obtained based on the comprehensive resource transfer features.
[0167] In one embodiment, determining an anomaly detection model set includes: obtaining a model parameter set for constructing an anomaly detection model; the model parameter set includes multiple model parameters; obtaining a parameter value interval corresponding to each model parameter in the model parameter set; obtaining a candidate parameter combination based on the parameter value interval corresponding to each model parameter in the model parameter set; traversing the candidate parameter combinations, and performing an accuracy evaluation on the anomaly detection model based on the selected parameter combination to obtain a first evaluation result; determining a target parameter combination from the candidate parameter combination based on the first evaluation result; adjusting the model parameters of the anomaly detection model based on the target parameter combination, and obtaining an anomaly detection model set based on the anomaly detection model after the parameter adjustment. Wherein, the model parameters may be hyperparameters corresponding to each anomaly detection model.
[0168] In one embodiment, obtaining a parameter value range corresponding to each model parameter in a model parameter set includes: obtaining multiple initial parameter values corresponding to each model parameter in the model parameter set to obtain multiple initial parameter combinations; performing accuracy evaluation on the anomaly detection model based on the multiple initial parameter combinations to obtain a second evaluation result; and determining the parameter value range corresponding to each model parameter in the model parameter set based on the second evaluation result.
[0169] In one embodiment, an accuracy evaluation of an anomaly detection model is performed based on a selected parameter combination to obtain a first evaluation result, including: performing an anomaly detection result of a sample object based on the selected parameter combination through the anomaly detection model; obtaining an anomaly label of the sample object from a label database; and determining an accuracy evaluation result corresponding to the anomaly detection model based on a match between the anomaly detection result of the sample object and the anomaly label of the sample object to obtain a first evaluation result.
[0170] In one embodiment, the parameters of the three anomaly detection models are tuned using a grid search method. Before using the grid tuning scheme, some parameter combinations are first tried manually, and the reasonable value range of each parameter is determined based on the trial results. After the reasonable value range of the parameter is determined, all parameter combinations are traversed using grid tuning, and finally the parameter combination with the best precision is selected as the optimal parameter.
[0171] Among them, the precision rate refers to the prediction results. It indicates how many samples with positive detection results of the model are truly positive samples. There are two possibilities for predicting positive: one is to predict the positive class as positive (TP), and the other is to predict the negative class as positive (FP). Based on this, the calculation formula of the precision rate P can be:
[0172] Taking the anomaly detection model HBOS as an example, the parameters in the model may include alpha, contamination, n-bins, and tol. Set the range of alpha, contamination, n-bins, and tol parameters respectively, match each model detection result with the label table to calculate the accuracy, and finally select the parameter combination corresponding to the highest accuracy. The obtained HBOS model parameters are as follows:
[0173] Parameter fields Parameter Value Field meaning alpha 0.1 Regularization parameter contamination 0.0001 Percentage of outliers n_bins 10 Number of boxes tol 0.5 The probability that a sample falls outside a bin
[0174] Taking the anomaly detection model as COPOD as an example, the model parameters of COPOD determined in the same way are as follows:
[0175] Parameter fields Parameter Value Field meaning contamination 0.0001 Percentage of outliers
[0176] Taking the anomaly detection model iForest as an example, the model parameters of iForest determined in the same way are as follows:
[0177] Parameter fields Parameter Value Field meaning max_features 1 The number of features input to each weak learner contamination 0.0001 Percentage of outliers n_estimators 100 Number of weak learners max_samples Auto The number of samples for each weak learner training; Auto: min(256, number of samples)
[0178] In the above embodiment, a parameter value interval is determined, and then various parameter combinations are traversed based on the parameter value interval, so that reliable model parameters can be obtained, thereby ensuring the reliability of the anomaly detection model.
[0179] In one embodiment, Figure 7 As shown, an object recognition method is provided, which is applied to Figure 1 The server in the example is used to illustrate the following steps:
[0180] S702, obtaining a target object set for abnormality identification, where the target object set includes multiple target objects.
[0181] S704, determine the characteristic dimensions corresponding to multiple resource transfer directions respectively; count the resource transfer attribute information in the same characteristic dimension in the resource transfer record set to obtain the resource transfer characteristics corresponding to each characteristic dimension respectively, and the resource transfer characteristics corresponding to each characteristic dimension constitute the resource transfer characteristic set corresponding to the target object.
[0182] S706, determining an anomaly detection model set, where the anomaly detection model set includes multiple anomaly detection models.
[0183] S708, determining the target feature sets corresponding to the respective feature dimensions.
[0184] S710, obtaining a feature distribution structure tree, and taking an initial node of the feature distribution structure tree as a current child node corresponding to the target object.
[0185] S712, obtaining a current feature dimension corresponding to the current child node, and obtaining a current feature partitioning threshold of a current target feature set corresponding to the current feature dimension.
[0186] S714, based on the current feature division threshold and the resource transfer feature of the target object in the current feature dimension, determine the distribution result of the resource transfer feature in the current target feature set.
[0187] S716, determining the next child node corresponding to the target object based on the distribution result, taking the next node as the updated current child node, and returning to S712 until the child node corresponding to the target object is updated.
[0188] S718, counting the number of child nodes corresponding to the target object to obtain the path length of the target object in the feature distribution structure tree.
[0189] S720, determining a first anomaly detection value corresponding to the target object based on the path length; and determining a model detection result of the anomaly detection model for the target object based on the first anomaly detection value.
[0190] S722, obtaining a feature partition interval set corresponding to a target feature set in an anomaly detection model.
[0191] S724, determining the distribution density corresponding to the feature partition interval, and using the distribution density as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
[0192] S726, determining a feature anomaly detection value corresponding to the resource transfer feature based on the distribution density; performing statistics on the feature anomaly detection value to obtain a second anomaly detection value corresponding to the target object; and determining a model detection result of the anomaly detection model for the target object based on the second anomaly detection value.
[0193] S728, determining an end division threshold; determining a tail feature set and a head feature set in the target feature set based on the end division threshold.
[0194] S730, performing statistics on the tail feature set to obtain a tail feature statistical value corresponding to the feature dimension set; and determining a tail distribution possibility corresponding to the resource transfer feature based on the tail feature statistical value.
[0195] S732, perform statistics on the header feature set to obtain header feature statistics corresponding to the feature dimension set; and determine the header distribution possibility corresponding to the resource transfer feature based on the header feature statistics.
[0196] S734, determining the skewness factor corresponding to the target object; and determining the skewness distribution possibility from the tail distribution possibility and the head distribution possibility based on the skewness factor.
[0197] S736, selecting a target distribution possibility corresponding to the resource transfer feature from the tail distribution possibility, the head distribution possibility, and the skewness distribution possibility.
[0198] S738, determining the model detection result of the anomaly detection model for the target object based on the target distribution possibility.
[0199] S740, collecting statistics on the model detection results of the target object by each anomaly detection model in the anomaly detection model set to obtain an anomaly detection result of the target object.
[0200] The present application also provides an application scenario, and the application scenario applies the above-mentioned object recognition method. Specifically, the application of the object recognition method in the application scenario is as follows:
[0201] The server obtains data of 3Minion (3 million) network accounts containing 77 transaction features from the terminal for experiment. The overall experiment is divided into 3 groups from the perspective of feature dimension, namely, account entry features, account exit features, and account exit & account entry features.
[0202] The server performs a grid search from the candidate model parameters to obtain the model parameters of each anomaly detection model, and then constructs the corresponding anomaly detection model. The account-in feature, account-out feature, and account-out & account-in feature are respectively input into each anomaly detection model to obtain the model detection result. The model detection results of each anomaly detection model are voted, and when all model detection results are abnormal network accounts, the corresponding network account is determined to be an abnormal network account. Then the anomaly detection results of all network accounts are obtained. The anomaly detection results of abnormal network accounts are matched with the black labels and white labels in the label library. Based on the matching results, the effect of the anomaly detection model set composed of anomaly detection models can be evaluated.
[0203] Furthermore, after the anomaly detection model set performed anomaly detection on three million sample data, 74 (about 0.00256% of the sample set) abnormal network accounts were obtained from the outgoing feature, 77 (about 0.00256% of the sample set) abnormal network accounts were obtained from the incoming feature, and 59 (about 0.00197% of the sample set) abnormal network accounts were obtained from the outgoing and incoming features. Figure 6 As shown in the effect diagram:
[0204] (1) Excluding the number of accounts on the whitelist, the anomaly detection model set has the highest recognition accuracy for inbound and outbound features.
[0205] (2) The anomaly detection model set can identify abnormal network accounts with a higher suspicion rate, which is higher than the suspicion rate of the overall label database, with an improvement rate of about 2%.
[0206] Based on this, the anomaly detection model set in this application scenario can obtain accurate anomaly detection results. Through the integrated learning of multiple unsupervised models, the accuracy of a single unsupervised model can be greatly improved, while ensuring the recall rate and taking into account the improvement of accuracy. In addition, the use of multiple unsupervised model combinations for anomaly detection can adapt to changes in data distribution without frequent parameter adjustments and model updates. The model does not need to collect a large number of labels during the detection process, which can greatly reduce the investment in labor costs, and the recall rate of the anomaly detection model set is high.
[0207] Furthermore, the transaction text of the abnormal network account whose abnormal detection result is abnormal network account is obtained, and the transaction text is further analyzed to determine the subdivision type of the abnormal network account. The subdivision type may include at least one of abnormal type A, abnormal type B, abnormal type C, abnormal type D, or abnormal type E. After the abnormal network account is refined, the proportion of network accounts of various subdivision types is as follows: Figure 8 shown.
[0208] Compare the anomaly detection time with the anomaly labeling time. Figure 8 802 is the percentage of network accounts of the subdivided type whose abnormal detection time is earlier than the abnormal label marking time. According to the comparison results, it can be determined that the recognition time of half of the abnormal network accounts in the anomaly detection model set in the embodiment of the present application is earlier than the recognition time of the label database. Therefore, the anomaly detection model can obtain accurate anomaly detection results in the shortest possible time. Based on this, the server can perceive unknown risks in advance, and can effectively and quickly provide early warnings and reminders to reduce platform risks.
[0209] It should be understood that, although the various steps in the above-mentioned flow chart are sequentially displayed according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.
[0210] Based on the same idea as the object recognition method in the above embodiment, the present application also provides an object recognition device, which can be used to perform the above object recognition method. For ease of explanation, the structural diagram of the object recognition device embodiment only shows the parts related to the embodiment of the present application. Those skilled in the art can understand that the illustrated structure does not constitute a limitation on the device, and may include more or fewer components than shown in the diagram, or combine certain components, or arrange the components differently.
[0211] In one embodiment, Fig. 9 As shown, an object recognition device 900 is provided. The device can adopt a software module or a hardware module, or a combination of the two to become a part of a computer device. The device specifically includes: an object set acquisition module 902, a feature set acquisition module 904, a model set determination module 906, a feature distribution determination module 908, a model detection result determination module 910 and an anomaly detection result determination module 912, wherein:
[0212] The object set acquisition module 902 is used to acquire a target object set for abnormality identification, where the target object set includes multiple target objects.
[0213] The feature set acquisition module 904 is used to acquire resource transfer feature sets corresponding to each of the target objects, wherein the resource transfer feature sets include resource transfer features corresponding to multiple feature dimensions.
[0214] The model set determination module 906 is used to determine an anomaly detection model set, wherein the anomaly detection model set includes multiple anomaly detection models; the anomaly detection models in the anomaly detection model set adopt different anomaly detection strategies.
[0215] The feature distribution determination module 908 is used to determine the distribution result of each resource transfer feature in the resource transfer feature set corresponding to the target object in the target feature set corresponding to the feature dimension through each of the anomaly detection models.
[0216] The model detection result determination module 910 is used to determine the model detection result of the anomaly detection model on the target object based on the distribution result obtained by the anomaly detection model.
[0217] The anomaly detection result determination module 912 is used to collect statistics on the model detection results of each anomaly detection model in the anomaly detection model set on the target object to obtain the anomaly detection result of the target object.
[0218] In the above-mentioned object recognition device, the distribution of resource transfer features in the target feature set of the feature dimension is analyzed through the anomaly detection model. Even if there is no sample label, each anomaly detection model can obtain the distribution result. The model detection result corresponding to the target object is determined based on the distribution result of each anomaly detection model, and then the model detection results corresponding to these anomaly detection models are comprehensively counted to obtain the anomaly detection result of the target object, which can effectively improve the efficiency of object recognition.
[0219] In one embodiment, the feature distribution determination module includes: a feature set determination submodule, which is used to obtain the resource transfer features corresponding to the feature dimensions in the resource transfer feature set corresponding to the target object, and obtain the target feature sets corresponding to each feature dimension; a feature partitioning submodule, which is used to obtain the distribution partitioning method corresponding to the target feature set in the anomaly detection model; based on the distribution partitioning method, the resource transfer features are divided in the target feature set of the feature dimension, and the distribution results of the resource transfer features in the target feature set of the feature dimension are obtained.
[0220] In one embodiment, the distribution partitioning method corresponding to the target feature set includes a partitioning method based on a threshold value, and the feature partitioning submodule includes: a structure tree acquisition unit, which is used to obtain a feature distribution structure tree, and the feature distribution structure tree includes multiple child nodes; a feature partitioning threshold determination unit, which is used to use the initial node of the feature distribution structure tree as the current child node corresponding to the target object, obtain the current feature dimension corresponding to the current child node, and obtain the current feature partitioning threshold of the current target feature set corresponding to the current feature dimension; a first distribution result determination unit, which is used to determine the distribution result of the resource transfer feature in the current target feature set based on the current feature partitioning threshold and the resource transfer feature of the target object in the current feature dimension; a child node updating unit, which is used to determine the next child node corresponding to the target object based on the distribution result, use the next node as the updated current child node, return to the step of obtaining the current feature dimension corresponding to the current child node, and obtaining the current feature partitioning threshold of the current target feature set corresponding to the current feature dimension, until the child node corresponding to the target object is updated.
[0221] In one embodiment, the model detection result determination module includes: a sub-node determination sub-module, which is used to determine the sub-node corresponding to the target object based on the distribution result; a path length determination sub-module, which is used to count the number of sub-nodes corresponding to the target object to obtain the path length of the target object in the feature distribution structure tree; a first detection value determination sub-module, which is used to determine the first abnormal detection value corresponding to the target object based on the path length, and the abnormal detection value is negatively correlated with the path length; a first detection result determination sub-module, which is used to determine the model detection result of the abnormal detection model for the target object based on the first abnormal detection value.
[0222] In one embodiment, the distribution partitioning method includes a partitioning method based on distribution intervals, and the feature partitioning submodule includes: a partitioning interval acquisition unit, used to obtain a feature partitioning interval set corresponding to the target feature set in the anomaly detection model, and the feature partitioning interval set includes multiple feature partitioning intervals; a feature quantity acquisition unit, used to obtain the feature quantity of the resource transfer feature in the target feature set in each feature partitioning interval; a second distribution result determination unit, used to determine the distribution density corresponding to the feature partitioning interval based on the feature quantity, and use the distribution density as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
[0223] In one embodiment, the model detection result determination module includes: a feature anomaly detection value determination submodule, which is used to determine the feature anomaly detection value corresponding to the resource transfer feature based on the distribution density, and the feature anomaly detection value is negatively correlated with the distribution density; a second detection value determination submodule, which is used to count the feature anomaly detection values corresponding to each resource transfer feature in the resource transfer feature set corresponding to the target object to obtain a second anomaly detection value corresponding to the target object; and a second detection result determination submodule, which is used to determine the model detection result of the anomaly detection model for the target object based on the second anomaly detection value.
[0224] In one embodiment, the feature partitioning submodule includes: an end partitioning threshold determination unit, which is used to determine the end partitioning threshold; an end feature set determination unit, which is used to determine the tail feature set corresponding to the feature tail and the head feature set corresponding to the feature head in the target feature set based on the end partitioning threshold; a tail feature set statistics unit, which is used to perform statistics on the tail feature set to obtain the tail feature statistics corresponding to the feature dimension set; a tail distribution possibility determination unit, which is used to determine the tail distribution possibility corresponding to the resource transfer feature based on the tail feature statistics corresponding to the feature dimension set; a head feature set statistics unit, which is used to perform statistics on the head feature set to obtain the head feature statistics corresponding to the feature dimension set; a head distribution possibility determination unit, which is used to determine the head distribution possibility corresponding to the resource transfer feature based on the head feature statistics corresponding to the feature dimension set; a third distribution result determination unit, which is used to determine the target distribution possibility corresponding to the resource transfer feature based on the tail distribution possibility and the head distribution possibility, and use the target distribution possibility as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
[0225] In one embodiment, the third distribution result determination unit includes: a skewness factor determination subunit, which is used to obtain the average value of each resource transfer feature in the resource transfer feature set, and determine the skewness factor corresponding to the target object based on the average value corresponding to the resource transfer feature set; a factor relative size determination subunit, which is used to obtain a skewness factor reference value and determine the relative size of the skewness factor and the skewness factor reference value; a skewness distribution possibility determination subunit, which is used to determine the skewness distribution possibility from the tail distribution possibility and the head distribution possibility based on the relative size of the skewness factor and the skewness factor reference value; a target distribution possibility determination subunit, which is used to compare the tail distribution possibility, the head distribution possibility and the skewness distribution possibility, and determine the target distribution possibility corresponding to the resource transfer feature from the tail distribution possibility, the head distribution possibility and the skewness distribution possibility according to the size comparison result; a distribution result determination subunit, which is used to use the target distribution possibility as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
[0226] In one embodiment, the abnormal detection result determination module includes: an abnormal result quantity determination submodule, which is used to determine the number of abnormal results whose model detection results are abnormal based on the model detection results of the target object; an abnormal resource transfer object determination submodule, which is used to determine that the target object is an abnormal resource transfer object when the number of abnormal results exceeds an abnormal quantity threshold, and obtain the abnormal detection result of the target object.
[0227] In one embodiment, the device also includes: a mark time acquisition module, which is used to acquire the abnormal label marking time of the target object, and the abnormal label marking time is the operation time of the label marking object performing the abnormal label marking operation on the target object; a detection time determination module, which is used to determine the abnormal detection time of the abnormal detection result of the target object when it is determined that the abnormal detection result of the target object is an abnormal resource transfer object; a time difference determination module, which is used to determine the time difference between the abnormal label marking time and the abnormal detection time; a limit time determination module, which is used to obtain the resource transfer permission limit time based on the time difference, and the resource transfer permission limit time is the limit time of the resource transfer permission of the abnormal resource transfer object when the object is detected as an abnormal resource transfer object using the abnormal detection model.
[0228] In one embodiment, the feature set acquisition module includes: a transfer record set acquisition submodule, used to acquire a resource transfer record set corresponding to the target object, wherein the resource transfer record set includes multiple resource transfer records; a feature dimension determination submodule, used to determine the feature dimensions corresponding to multiple resource transfer directions based on the resource transfer direction as the statistical dimension; a transfer feature determination submodule, used to count the resource transfer attribute information in the same feature dimension in the resource transfer record set, and obtain the resource transfer features corresponding to each of the feature dimensions, and the resource transfer features corresponding to each of the feature dimensions constitute the resource transfer feature set corresponding to the target object.
[0229] For the specific definition of the object recognition device, please refer to the definition of the object recognition method above, which will not be repeated here. Each module in the above object recognition device can be implemented in whole or in part by software, hardware and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0230] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.10As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as model detection results and anomaly detection results. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an object recognition method is implemented.
[0231] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0232] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.
[0233] In one embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0234] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-mentioned method embodiments.
[0235] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0236] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0237] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A method for classifying object accounts, characterized in that: The method comprises: Acquire a target object account set to be subjected to anomaly identification, wherein the target object account set includes a plurality of target object accounts; Acquire a resource transfer feature set corresponding to each of the target object accounts, the resource transfer feature set including resource transfer features corresponding to a plurality of feature dimensions, the plurality of feature dimensions including at least: a time dimension of a resource transfer operation and a frequency dimension of a resource transfer operation, the resource transfer features under the time dimension of the resource transfer operation at least including: the time of the resource transfer operation, and the resource transfer features under the frequency dimension of the resource transfer operation at least including: the frequency of the resource transfer operation; Determine an anomaly detection model set, wherein the anomaly detection model set includes multiple anomaly detection models; each anomaly detection model in the anomaly detection model set performs anomaly detection in a different manner; By using each of the anomaly detection models and adopting respective anomaly detection methods, each resource transfer feature in the resource transfer feature set corresponding to the target object account is classified respectively, and a distribution result of each resource transfer feature in the target feature set corresponding to the feature dimension is obtained; Determining a model detection result of the anomaly detection model on the target object account based on the distribution result obtained by the anomaly detection model, wherein the model detection result is a detection result of the anomaly detection model on the category to which the target object account belongs; The model detection results of each anomaly detection model in the anomaly detection model set on the target object account are counted to obtain an anomaly detection result of the target object account, and the anomaly detection result is used to indicate whether the target object account is an abnormal object account.
2. The method according to claim 1, characterized in that The determining, by using each of the anomaly detection models, a distribution result of each resource transfer feature in the resource transfer feature set corresponding to the target object account in the target feature set corresponding to the feature dimension, includes: Obtaining resource transfer features corresponding to feature dimensions from the resource transfer feature set corresponding to the target object account, and obtaining target feature sets corresponding to each feature dimension; Obtain a distribution partitioning method corresponding to the target feature set in the anomaly detection model; based on the distribution partitioning method, divide the resource transfer feature in the target feature set of the feature dimension to obtain a distribution result of the resource transfer feature in the target feature set of the feature dimension.
3. The method according to claim 2, characterized in that The distribution division method corresponding to the target feature set includes a division method based on a threshold value, and the distribution division method corresponding to the target feature set in the anomaly detection model is obtained; based on the distribution division method, the resource transfer feature is divided in the target feature set of the feature dimension, and the distribution result of the resource transfer feature in the target feature set of the feature dimension is obtained, including: Acquire a feature distribution structure tree, wherein the feature distribution structure tree includes a plurality of child nodes; Taking the initial node of the feature distribution structure tree as the current child node corresponding to the target object account, obtaining the current feature dimension corresponding to the current child node, and obtaining the current feature partitioning threshold of the current target feature set corresponding to the current feature dimension; Based on the current feature division threshold and the resource transfer feature of the target object account in the current feature dimension, determining a distribution result of the resource transfer feature in the current target feature set; Based on the distribution result, determine the next child node corresponding to the target object account, use the next node as the updated current child node, return to the step of obtaining the current feature dimension corresponding to the current child node, and obtain the current feature partition threshold of the current target feature set corresponding to the current feature dimension, until the child node corresponding to the target object account is updated.
4. The method according to claim 3, characterized in that The determining, based on the distribution result obtained by the anomaly detection model, a model detection result of the anomaly detection model on the target object account includes: Determine the child node corresponding to the target object account based on the distribution result; Counting the number of child nodes corresponding to the target object account to obtain the path length of the target object account in the feature distribution structure tree; Determine a first anomaly detection value corresponding to the target object account based on the path length, wherein the anomaly detection value is negatively correlated with the path length; A model detection result of the anomaly detection model on the target object account is determined based on the first anomaly detection value.
5. The method according to claim 2, characterized in that: The distribution partitioning method includes a partitioning method based on a distribution interval, and the distribution partitioning method corresponding to the target feature set in the anomaly detection model is obtained; based on the distribution partitioning method, the resource transfer feature is partitioned in the target feature set of the feature dimension, and the distribution result of the resource transfer feature in the target feature set of the feature dimension is obtained, including: Acquire a feature partition interval set corresponding to the target feature set in the anomaly detection model, wherein the feature partition interval set includes multiple feature partition intervals; Obtaining the number of features of the resource transfer feature in the target feature set in each feature division interval; The distribution density corresponding to the feature division interval is determined based on the feature quantity, and the distribution density is used as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
6. The method according to claim 5, characterized in that The determining, based on the distribution result obtained by the anomaly detection model, a model detection result of the anomaly detection model on the target object account includes: Determine a feature anomaly detection value corresponding to the resource transfer feature based on the distribution density, wherein the feature anomaly detection value is negatively correlated with the distribution density; The feature anomaly detection value corresponding to each resource transfer feature in the resource transfer feature set corresponding to the target object account is counted to obtain a second anomaly detection value corresponding to the target object account; A model detection result of the anomaly detection model on the target object account is determined based on the second anomaly detection value.
7. The method according to claim 2, characterized in that The method of obtaining the distribution partitioning method corresponding to the target feature set in the anomaly detection model; dividing the resource transfer feature in the target feature set of the feature dimension based on the distribution partitioning method to obtain the distribution result of the resource transfer feature in the target feature set of the feature dimension, including: Determine the end division threshold; Determine, based on the end division threshold, a tail feature set corresponding to the feature tail and a head feature set corresponding to the feature head in the target feature set; Performing statistics on the tail feature set to obtain a tail feature statistical value corresponding to the feature dimension set; Determining the possibility of tail distribution corresponding to the resource transfer feature based on the tail feature statistics corresponding to the feature dimension set; Performing statistics on the header feature set to obtain header feature statistics corresponding to the feature dimension set; Determine the header distribution possibility corresponding to the resource transfer feature based on the header feature statistics corresponding to the feature dimension set; The target distribution possibility corresponding to the resource transfer feature is determined based on the tail distribution possibility and the head distribution possibility, and the target distribution possibility is used as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
8. The method according to claim 7, characterized in that The determining the target distribution possibility corresponding to the resource transfer feature based on the tail distribution possibility and the head distribution possibility, and using the target distribution possibility as the distribution result of the resource transfer feature in the target feature set of the feature dimension, includes: Obtaining an average value of each resource transfer feature in the resource transfer feature set, and determining a skewness factor corresponding to the target object account based on the average value corresponding to the resource transfer feature set; Obtaining a skewness factor reference value, and determining a relative magnitude of the skewness factor and the skewness factor reference value; Determining the skewness distribution possibility from the tail distribution possibility and the head distribution possibility based on the relative size of the skewness factor and the skewness factor reference value; Compare the tail distribution possibility, the head distribution possibility and the skewness distribution possibility, and determine the target distribution possibility corresponding to the resource transfer feature from the tail distribution possibility, the head distribution possibility and the skewness distribution possibility according to the comparison result; The target distribution possibility is taken as the distribution result of the resource transfer feature in the target feature set of the feature dimension.
9. The method according to claim 1, characterized in that: The step of collecting statistics on the model detection results of the target object account by each anomaly detection model in the anomaly detection model set to obtain the anomaly detection result of the target object account includes: Determine the number of abnormal results of which the model detection results are abnormal according to the model detection results of the target object account; When the number of abnormal results exceeds the abnormal number threshold, the target object account is determined to be an abnormal resource transfer object, and an abnormal detection result of the target object account is obtained.
10. The method according to claim 1, characterized in that The method further comprises: Acquire the abnormal label marking time of the target object account, where the abnormal label marking time is the operation time when the label marking object performs the abnormal label marking operation on the target object account; When it is determined that the abnormality detection result of the target object account is an abnormal resource transfer object, determining the abnormality detection time at which the abnormality detection result of the target object account is obtained; Determine the time difference between the anomaly labeling time and the anomaly detection time; The resource transfer permission restriction duration is obtained based on the time difference, and the resource transfer permission restriction duration is the restriction duration of the resource transfer permission of the abnormal resource transfer object when the object is detected as an abnormal resource transfer object by using the abnormality detection model.
11. The method according to claim 1, characterized in that: The obtaining of resource transfer feature sets corresponding to each of the target object accounts includes: Acquire a resource transfer record set corresponding to the target object account, wherein the resource transfer record set includes multiple resource transfer records; Taking the resource transfer direction as the statistical dimension, determining the characteristic dimensions corresponding to the multiple resource transfer directions respectively; The resource transfer attribute information in the same feature dimension in the resource transfer record set is counted to obtain resource transfer features corresponding to each of the feature dimensions, and the resource transfer features corresponding to each of the feature dimensions constitute the resource transfer feature set corresponding to the target object account.
12. An object account classification device, characterized in that: The device comprises: An object set acquisition module is used to acquire a target object account set to be identified for anomaly, wherein the target object account set includes multiple target object accounts; A feature set acquisition module is used to acquire a resource transfer feature set corresponding to each of the target object accounts, wherein the resource transfer feature set includes resource transfer features corresponding to a plurality of feature dimensions, wherein the plurality of feature dimensions include at least: a time dimension of a resource transfer operation and a frequency dimension of a resource transfer operation, wherein the resource transfer features under the time dimension of the resource transfer operation include at least: a time of the resource transfer operation, and the resource transfer features under the frequency dimension of the resource transfer operation include at least: a frequency of the resource transfer operation; A model set determination module, used to determine an anomaly detection model set, wherein the anomaly detection model set includes a plurality of anomaly detection models; each anomaly detection model in the anomaly detection model set performs anomaly detection in a different manner; A feature distribution determination module, used to classify each resource transfer feature in the resource transfer feature set corresponding to the target object account by using each of the anomaly detection models and adopting respective anomaly detection methods, and obtain the distribution result of each resource transfer feature in the target feature set corresponding to the feature dimension; A model detection result determination module, used to determine a model detection result of the anomaly detection model on the target object account based on the distribution result obtained by the anomaly detection model, wherein the model detection result is a detection result of the anomaly detection model on the category to which the target object account belongs; The anomaly detection result determination module is used to collect statistics on the model detection results of each anomaly detection model in the anomaly detection model set on the target object account to obtain the anomaly detection result of the target object account, and the anomaly detection result is used to indicate whether the target object account is an abnormal object account.
13. The device according to claim 12, characterized in that The feature distribution determination module comprises: A feature set determination submodule is used to obtain resource transfer features corresponding to feature dimensions in the resource transfer feature set corresponding to the target object account, and obtain target feature sets corresponding to each feature dimension; The feature partitioning submodule is used to obtain the distribution partitioning method corresponding to the target feature set in the anomaly detection model; based on the distribution partitioning method, the resource transfer feature is partitioned in the target feature set of the feature dimension to obtain the distribution result of the resource transfer feature in the target feature set of the feature dimension.
14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 11 are implemented.
15. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.
