Account anomaly identification methods, devices, equipment, media and products

By identifying the account security status, selecting the corresponding anomaly detection model, and calculating and deducting the offset attributable to the status change, the problems of high false alarm rate and inability to continuously identify abnormal access in the existing technology are solved, and more efficient account security protection is achieved.

CN123122937APending Publication Date: 2026-11-17INDUSTRIAL AND COMMERCIAL BANK OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611353812.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to distinguish between changes in normal access scenarios and abnormal account behavior, resulting in a high false alarm rate and an inability to continuously identify abnormal access behavior.

Method used

By acquiring the target user's account environment information, identifying its current security status, selecting the corresponding anomaly detection model, obtaining access feature data, calculating the initial feature offset, deducting the contribution offset attributable to state changes, and using the remaining feature offset to determine whether the account is abnormal.

Benefits of technology

It effectively distinguishes between changes in normal access scenarios and abnormal real accounts, reduces false alarm rates, continuously identifies abnormal access behavior, and improves the integrity and reliability of account security protection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122937A_ABST
    Figure CN123122937A_ABST
Patent Text Reader

Abstract

This invention discloses a method, apparatus, device, medium, and product for identifying account anomalies, relating to the field of network security technology. The method includes: responding to a target user's current access and identifying their account security status; selecting a corresponding anomaly detection model based on the account security status to perform anomaly detection on access feature data during the current access process, obtaining an initial feature offset; determining the state-associated access feature corresponding to the account security status and its actual feature value; determining the expected distribution of the feature value when the target user is in that account security status, and determining a contribution offset based on the degree of deviation of the actual feature value from the expected distribution; subtracting the contribution offset from the initial feature offset, and determining whether the account is abnormal based on the remaining feature offset. This invention can effectively distinguish between normal account access scenario changes and potential abnormal account behavior, reducing the false alarm rate, and can continuously identify anomalies during the access process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of network security technology, and in particular to a method, apparatus, device, medium, and product for identifying abnormal accounts. Background Technology

[0002] Account anomaly detection is used to detect unauthorized account activities such as theft and impersonation. With the diversification of attack methods, account theft often continues after a user logs in normally, making single-login verification insufficient for security needs.

[0003] Existing technologies typically trigger alarms based on static rules such as device fingerprints and IP addresses, or employ one-time verification during the login phase. However, normal changes in access scenarios, such as users traveling to different locations or changing devices, can alter these characteristics, making it difficult for static rules to distinguish them from genuine account anomalies and leading to false alarms. Furthermore, one-time verification only takes effect during login and cannot continuously identify anomalies during subsequent access, creating a significant security blind spot. Summary of the Invention

[0004] This invention provides a method, apparatus, device, medium, and product for identifying abnormal accounts, in order to solve the problem in the prior art that it is difficult to distinguish between changes in normal access scenarios and abnormal account behavior.

[0005] According to one aspect of the present invention, a method for identifying abnormal accounts is provided, the method comprising: In response to a target user's current access to the business system, the system obtains the target user's account environment information and identifies the target user's current account security status based on the account environment information. Based on the account security status, select the corresponding target anomaly detection model from the candidate anomaly detection models, and obtain the access feature data generated by the target user during the current access process; The access feature data is input into the target anomaly detection model for anomaly detection to obtain an initial feature offset; wherein, the initial feature offset is used to characterize the feature offset of the access feature data relative to the normal access feature baseline of the target user; Determine the state-related access feature corresponding to the account security status from the access feature data, and obtain the actual feature value corresponding to the state-related access feature; wherein, the state-related access feature is a feature that is expected to change with the change of the account security status; Determine the expected distribution of the feature values ​​of the state-associated access feature when the target user is in the account security state, and determine the contribution offset of the state-associated access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected distribution of the feature value; The contribution offset is subtracted from the initial feature offset to obtain the remaining feature offset, and the target user's account is determined to be abnormal based on the remaining feature offset.

[0006] According to another aspect of the present invention, an account anomaly identification device is provided, the device comprising: The access response module is used to respond to the target user's current access to the business system, obtain the target user's account environment information, and identify the target user's current account security status based on the account environment information. The feature data acquisition module is used to select the corresponding target anomaly detection model from the candidate anomaly detection models according to the account security status, and acquire the access feature data generated by the target user during the current access process. The feature offset determination module is used to input the access feature data into the target anomaly detection model for anomaly detection to obtain an initial feature offset; wherein, the initial feature offset is used to characterize the feature offset of the access feature data relative to the normal access feature baseline of the target user. The feature value acquisition module is used to determine the state-related access feature corresponding to the account security status from the access feature data, and to acquire the actual feature value corresponding to the state-related access feature; wherein, the state-related access feature is a feature that is expected to change with the change of the account security status. The offset calculation module is used to determine the expected distribution of the feature values ​​of the state-related access feature when the target user is in the account security state, and to determine the contribution offset of the state-related access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected distribution of the feature value. The account anomaly detection module is used to subtract the contribution offset from the initial feature offset to obtain the remaining feature offset, and to determine whether the target user's account is abnormal based on the remaining feature offset.

[0007] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method described in any one of the present invention.

[0008] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the method described in any one of the present invention.

[0009] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in any one of the present invention.

[0010] The beneficial effects of this invention are as follows: First, based on the account security status, determine the expected changes in the status associated with access features and their corresponding contribution offsets, and deduct the contribution offsets from the initial feature offsets, using only the remaining feature offsets as the basis for determining whether the account is abnormal. This effectively distinguishes changes in account security status such as login status from different locations and login status from new devices from real account anomalies, reducing the false alarm rate.

[0011] Second, the entire anomaly identification process is triggered by the user's current access to the business system and can be continuously executed. It no longer relies on the one-time verification during the login stage, effectively making up for the security blind spot of existing technologies that cannot continuously identify abnormal access behavior after the user has passed login verification, and improving the integrity and reliability of account security protection.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of an account anomaly identification method provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of an initial feature offset determination method provided in Embodiment 2 of the present invention; Figure 3 This is a flowchart of a contribution offset determination method provided in Embodiment 3 of the present invention; Figure 4 This is a flowchart of a comprehensive feature offset determination method provided in Embodiment 4 of the present invention; Figure 5This is a schematic diagram of the structure of an account anomaly identification device provided in Embodiment 5 of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device that implements the account anomaly identification method of this invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0016] It should be noted that the terms "first," "second," "third," "fourth," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0017] Example 1 Figure 1 This is a flowchart of an account anomaly identification method provided in Embodiment 1 of the present invention. This embodiment is applicable to scenarios that involve real-time analysis of user access behavior in a business system to identify whether an account has been stolen or misused, or other abnormal risks. This method can be executed by an account anomaly identification device, which can be implemented in hardware and / or software. Figure 1 As shown, the method includes: S101. In response to the target user's current access to the business system, obtain the target user's account environment information, and identify the target user's current account security status based on the account environment information.

[0018] In this context, the target user refers to the entity currently accessing the business system and requiring anomaly detection. In a banking scenario, the target user is typically a customer holding a bank account.

[0019] A business system refers to a system platform that provides business functions and is accessed by target users through terminal devices, including online banking, mobile banking, and corporate online banking.

[0020] Current access refers to an access behavior initiated by a target user to the business system at the current moment. In this embodiment, "current access" is the trigger condition for the entire anomaly identification method, indicating that the system responds in real time to every access behavior of the user, pointing to an access event that is "occurring" and is different from historical access data.

[0021] Account environment information refers to a collection of environmental data related to the target user's current access, such as device, network, and time, used to characterize the physical or logical environment of the access. This information covers various environmental data generated when a user accesses a business system, including but not limited to terminal device information, network address information, and access time information.

[0022] Account security status refers to the category label obtained by classifying the current access scenario of a target user based on account environment information. It is used to characterize the environmental scenario category to which the current access belongs, such as whether the access occurred in an unusual geographical location, whether an unregistered device was used, and whether the account has recently experienced any security incidents. The purpose of identifying account security status is to perceive the scenario category to which the current access belongs, providing a basis for subsequent judgment on whether feature changes are within normal expectations, thereby distinguishing feature shifts attributable to status changes from genuine abnormal shifts.

[0023] Optionally, account security status includes normal status, security information change status, login status from a different location, and login status from a new device.

[0024] The "normal" state indicates that the device fingerprint, IP address, and login timestamp of the currently accessed device all match the baseline of normal access characteristics of the target user, and the account has no recent security event records. In this state, there are no additional offsets caused by changes in the scenario.

[0025] A security information change status indicates that the target user's account has recently undergone a security information change event, such as changing the linked mobile phone number or resetting the password. After such an event occurs, the user's access may trigger an additional security verification process, resulting in differences in access characteristics such as page access sequence compared to the normal pattern.

[0026] A login from a different location indicates that the location of the currently accessing IP address is inconsistent with the commonly used IP range recorded in the target user's normal access characteristic baseline. This type of status typically corresponds to cross-regional access scenarios such as the target user being on a business trip or traveling.

[0027] The "new device login" status indicates that the device fingerprint used for the current access does not appear in the list of frequently used devices recorded in the target user's normal access characteristic baseline. This type of status typically corresponds to the target user's first access scenario after changing terminal devices.

[0028] In one implementation, in response to a target user's current access to the business system, the system obtains the account environment information corresponding to that access, including device fingerprint, IP address, and login timestamp. The system compares the device fingerprint of the current access with the list of commonly used devices recorded in the target user's normal access characteristic baseline, compares the IP address of the current access with the commonly used IP range recorded in the normal access characteristic baseline, compares the login timestamp of the current access with the commonly used time period recorded in the normal access characteristic baseline, and checks whether there have been any recent security information change events for the account.

[0029] If the device fingerprint, IP address, and login timestamp all meet the baseline characteristics of normal access, and the account has no recent security event records, then the current access is determined to be in a normal state; if the account has recently experienced a security information change event, then the current access is determined to be in a security information change state; if the IP address does not meet the baseline characteristics of normal access, then the current access is determined to be in a remote login state; if the device fingerprint does not meet the baseline characteristics of normal access, then the current access is determined to be in a new device login state.

[0030] This state recognition process is executed in real time when an access occurs, and its recognition results are used to indicate the scenario category to which the current access belongs, providing a basis for selecting the corresponding anomaly detection model for different scenarios in subsequent steps.

[0031] S102. Based on the account security status, select the corresponding target anomaly detection model from the candidate anomaly detection models, and obtain the access feature data generated by the target user during the current access process.

[0032] Among them, candidate anomaly detection models refer to multiple pre-configured alternative anomaly detection models used to detect anomalies in access behavior under different scenario conditions. Different candidate anomaly detection models are trained based on different data sources and are suitable for different access scenarios.

[0033] A target anomaly detection model refers to the anomaly detection model selected from candidate anomaly detection models based on the current account security status of the target user. Different target anomaly detection models respond differently to the same access feature data, thus outputting different initial feature offsets in different scenario states. By selecting the target anomaly detection model corresponding to the current state, the system can adopt differentiated detection benchmarks for different access scenarios.

[0034] Access feature data refers to data generated by a target user during their current access process and used for anomaly detection. This data reflects the environmental and interaction characteristics of the current access and serves as the data foundation for subsequent anomaly detection and deducting contributing offsets from the initial feature offsets.

[0035] In one implementation, after identifying the current account security status of the target user, the system selects the model corresponding to that status from the candidate anomaly detection models as the target anomaly detection model.

[0036] During a target user's access to the business system, the system collects access characteristic data in real time. This data includes at least login environment data and interaction behavior data. Login environment data describes the device, network, and time of the current access, while interaction behavior data describes the user's page navigation, dwell time, and operational rhythm during the access process.

[0037] It should be noted that the above-mentioned access feature data was collected during the user's normal use of the business system, and the scope of collection was limited to the access behavior itself, without involving the specific content input by the user. The collected access feature data is used as input to the target anomaly detection model in subsequent steps to participate in anomaly detection calculations.

[0038] The relevant information (including but not limited to device fingerprints, IP addresses, login timestamps, page access sequences, page dwell time, operation interval duration, etc.) and data involved in the embodiments of this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant regions.

[0039] S103. Input the access feature data into the target anomaly detection model to perform anomaly detection and obtain the initial feature offset.

[0040] The initial feature offset is used to characterize the access feature data, which is the feature offset relative to the normal access feature baseline of the target user.

[0041] The initial feature offset refers to the quantitative result output by the target anomaly detection model after inputting the access feature data generated by the target user during the current access process. This offset represents the degree of deviation of the current access feature data from the normal access feature baseline. This offset is the original input for subsequent operations to subtract contributing offsets, reflecting the overall difference between the current access and normal access patterns. The larger the offset, the greater the difference between the current access features and the normal access feature baseline.

[0042] A normal access characteristic baseline is a reference standard established based on the access characteristic data of a target user under historical normal conditions. This baseline records the normal value range of various access characteristics for the target user in regular access scenarios, such as a list of frequently used device fingerprints, frequently used IP ranges, distribution of regular access time periods, and regular page access sequence patterns. The normal access characteristic baseline serves as a reference standard for determining whether the current access characteristic data has shifted, and it is also the data basis for subsequently calculating the expected distribution of feature values. This baseline is a personalized baseline for the target user, not a universal standard for all users.

[0043] In one implementation, the system inputs the collected access feature data into a target anomaly detection model. This target anomaly detection model is trained based on historical access feature data and is used to compare the current access feature data with a normal access feature baseline, and output an initial feature offset.

[0044] The model compares the current access feature data with the baseline of normal access features to determine the offset of each access feature, and outputs a quantified initial feature offset. This initial feature offset is a comprehensive value used to reflect the degree to which the current access deviates from the normal access pattern.

[0045] It should be noted that different target anomaly detection models are selected under different account security states, but all models use the normal access feature baseline as a reference. For the same target user and the same access feature data, inputting different target anomaly detection models under different account security states may result in different initial feature offsets. By selecting the appropriate target anomaly detection model based on the account security state, the anomaly detection baseline can be matched with the actual access scenario.

[0046] S104. Determine the status-associated access features corresponding to the account security status from the access feature data, and obtain the actual feature values ​​corresponding to the status-associated access features.

[0047] State-related access features are those that change as the account's security status changes. For example, the IP address changes when logging in from a different location; the device fingerprint changes when logging in from a new device. These changes are caused by normal switching of access scenarios, not by abnormal account behavior. The purpose of identifying state-related access features is to eliminate the portions of subsequent steps that can be attributed to normal scenario changes.

[0048] The actual feature value refers to the specific value of the state-associated access feature in the current access. For example, the actual feature value of the device fingerprint feature is the unique identifier of the terminal device used for the current access, and the actual feature value of the IP address feature is the IP address of the current access.

[0049] In one implementation, the system determines access features associated with the target user's current account security status, i.e., status-related access features. Specifically, the system pre-configures a correspondence between account security statuses and status-related access features, with different status-related access features corresponding to different account security statuses. The system determines the corresponding status-related access features from the access feature data based on the currently identified account security status. After determining the status-related access features, the system obtains the actual values ​​of these features in the current access from the collected access feature data, using them as actual feature values.

[0050] S105. Determine the expected distribution of the feature values ​​of the state-related access feature when the target user is in an account security state, and determine the contribution offset of the state-related access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected feature value distribution.

[0051] The expected distribution of feature values ​​refers to the probability distribution of state-related access features when the target user is in the current account security state. This distribution reflects the probability pattern of each state-related access feature value in the current access scenario—values ​​closer to the center of the distribution have a higher probability of occurrence, while values ​​closer to the tail of the distribution have a lower probability of occurrence.

[0052] Deviation refers to the degree of difference between the actual feature value and the expected distribution of the feature value. It measures whether the position of the actual feature value in the expected distribution deviates from the expected center. The closer the actual feature value is to the center of the expected distribution, the lower the deviation; the closer it is to the tail of the expected distribution or beyond its coverage area, the higher the deviation.

[0053] The contribution offset refers to the portion of the initial feature offset that is attributable to the state-associated access feature.

[0054] In one implementation, the system obtains the expected distribution of feature values ​​of each state-related access feature under the current account security state of the target user.

[0055] The system uses the actual feature values ​​of the access features associated with each state as query conditions, and determines the quantile position of the actual value in the corresponding expected distribution of feature values, denoted as q. q reflects the relative position of the actual feature value in the expected distribution: when q is close to 0.5, the actual feature value is located in the central region of the expected distribution, indicating that the actual value of the feature is common and expected in the current scenario; when q is close to 0 or 1, the actual feature value is located in the tail region of the expected distribution or exceeds the coverage range, indicating that the actual value of the feature is rare and unexpected in the current scenario.

[0056] The system calculates the degree of deviation D based on q: D = 2 × |q - 0.5|. The value of D ranges from [0,1]. D = 0 indicates that the actual feature value is located in the center of the expected distribution, with the lowest degree of deviation; D = 1 indicates that the actual feature value is located in the tail of the expected distribution or exceeds the coverage range, with the highest degree of deviation.

[0057] S106. Subtract the contribution offset from the initial feature offset to obtain the remaining feature offset, and determine whether the target user's account is abnormal based on the remaining feature offset.

[0058] The residual feature offset refers to the offset obtained by subtracting the contribution offset from the initial feature offset. Its physical meaning is: the portion of the initial feature offset that cannot be explained by the current account security status; that is, the feature offset that truly reflects whether the account is abnormal after excluding feature offsets caused by changes in normal access scenarios. This offset is directly used to determine whether the target user's account is abnormal—the higher the residual feature offset, the more abnormal components in the current access that cannot be explained by changes in normal scenarios, and the greater the possibility that the account is abnormal.

[0059] In one implementation, the system subtracts the contribution offset from the initial feature offset to obtain the remaining feature offset.

[0060] After obtaining the remaining feature offset, the system determines whether the target user's account is abnormal based on the remaining feature offset. The system compares the remaining feature offset with preset anomaly detection conditions. If the remaining feature offset meets the anomaly detection conditions, the target user's account is determined to be abnormal; otherwise, it is determined to be normal. The anomaly detection conditions can be flexibly configured according to the actual business scenario and security policy, such as configuring them as anomaly detection thresholds or risk ranges.

[0061] The aforementioned contribution offset can be determined by combining one or more sub-contribution offsets, and the method of combining each sub-contribution offset matches the method of superimposing feature offsets in actual business scenarios. For example, when multiple state-related access features each generate independent offset contributions, the system can use summation or weighted summation to comprehensively process the sub-contribution offsets. The specific method of combination can be configured according to the importance of each state-related access feature in anomaly detection or the needs of actual business scenarios. Those skilled in the art should understand that the above-mentioned combination method is only an illustrative example and does not constitute a limitation of the present invention.

[0062] The beneficial effects of the embodiments of the present invention are as follows: First, based on the account security status, determine the expected changes in the status associated with access features and their corresponding contribution offsets, and deduct the contribution offsets from the initial feature offsets, using only the remaining feature offsets as the basis for determining whether the account is abnormal. This effectively distinguishes changes in account security status such as login status from different locations and login status from new devices from real account anomalies, reducing the false alarm rate.

[0063] Second, the entire anomaly identification process is triggered by the user's current access to the business system and can be continuously executed. It no longer relies on the one-time verification during the login stage, effectively making up for the security blind spot of existing technologies that cannot continuously identify abnormal access behavior after the user has passed login verification, and improving the integrity and reliability of account security protection.

[0064] Optionally, candidate anomaly detection models include the first model, the second model, the third model, and the fourth model.

[0065] The first model is trained based on historical access feature data collected from the target user under normal conditions; the second model is trained based on historical access feature data collected from the target user under security information change conditions; the third model is trained based on historical access feature data collected from the target user under remote login conditions; and the fourth model is trained based on historical access feature data collected from the target user under new device login conditions.

[0066] Historical access feature data refers to the collection of access feature data generated by target users during their past visits to the business system, serving as sample data for training candidate anomaly detection models. Unlike real-time collected "access feature data," historical access feature data consists of past access records, used to construct anomaly detection models corresponding to each state. Historical access feature data collected under different states reflects the access characteristic patterns of target users in different access scenarios, enabling models trained based on this data to identify normal access patterns under corresponding states.

[0067] In one implementation, the system pre-collects historical access feature data generated by the target user under different states, and trains corresponding candidate anomaly detection models accordingly.

[0068] Specifically, for the normal state, the system filters out access records that meet the normal state judgment conditions from the target user's historical access records, extracts access feature data from them, constructs a normal state historical access feature dataset, and trains a first model based on the dataset, so that the first model can identify the normal access feature patterns of the target user in the normal state.

[0069] The system filters out access records that meet the judgment conditions for security information change status from the target user's historical access records, extracts access feature data from them, constructs a historical access feature dataset for security information change status, and trains a second model based on this dataset, enabling the second model to identify the normal access feature patterns of the target user under the security information change status.

[0070] The system targets logins from different locations by filtering out access records that meet the criteria for determining logins from the target user's historical access records, extracting access feature data from these records, constructing a historical access feature dataset for logins from different locations, and training a third model based on this dataset. This third model is then able to identify the normal access feature patterns of the target user when logging in from different locations.

[0071] For new device login status, the system filters out access records that meet the new device login status judgment conditions from the target user's historical access records, extracts access feature data from them, constructs a new device login status historical access feature dataset, and trains a fourth model based on this dataset, enabling the fourth model to identify the normal access feature patterns of the target user in the new device login status.

[0072] The judgment conditions corresponding to each of the above states can be configured according to actual business needs and security policies, such as the time window range. After each model is trained, it is used as a candidate anomaly detection model, and is selected as the target anomaly detection model based on the current account security status of the target user.

[0073] By training corresponding candidate anomaly detection models for different account security states, the appropriate target anomaly detection model can be selected based on the current account security state of the target user. This allows for differentiated detection benchmarks for different access scenarios, avoiding the use of a uniform detection standard to measure access behavior in different scenarios. This further improves the accuracy of anomaly detection and reduces false alarms caused by scenario differences.

[0074] Optionally, based on the account security status, a corresponding target anomaly detection model can be selected from the candidate anomaly detection models, including: When the account security status is normal, the first model is used as the target anomaly detection model; when the account security status is in a security information change state, the second model is used as the target anomaly detection model; when the account security status is in a remote login state, the third model is used as the target anomaly detection model; when the account security status is in a new device login state, the fourth model is used as the target anomaly detection model.

[0075] In one implementation, after identifying the current account security status of the target user, the system determines the corresponding target anomaly detection model based on that status.

[0076] The system pre-configures a mapping between account security status and candidate anomaly detection models, which is stored in the system as a configuration table or rule set. When the system identifies an account security status, it queries the mapping to find a matching candidate anomaly detection model and identifies it as the target anomaly detection model.

[0077] Specifically, when the target user's account security status is normal, the system uses the first model as the target anomaly detection model; when the account security status is in a security information change state, the system uses the second model as the target anomaly detection model; when the account security status is in a remote login state, the system uses the third model as the target anomaly detection model; and when the account security status is in a new device login state, the system uses the fourth model as the target anomaly detection model.

[0078] By configuring corresponding target anomaly detection models for normal status, security information change status, remote login status, and new device login status, the benchmark for anomaly detection is matched with the current access scenario. Each model is trained based on historical access feature data for its corresponding status, enabling accurate identification of normal access feature patterns within that status. When a target user is in different access scenarios, the system uses the model matching that scenario for anomaly detection. This avoids using a uniform detection standard to measure access behavior in different scenarios, ensuring that normal feature shifts caused by scenario changes are not misjudged as anomalies, further improving the accuracy of anomaly detection and reducing the false alarm rate.

[0079] Optionally, the access feature data includes at least one of login environment data and interaction behavior data; the login environment data includes at least one of device fingerprint, IP address and login timestamp; the interaction behavior data includes at least one of page access sequence, page dwell time and operation interval time.

[0080] Login environment data refers to information related to the login environment automatically recorded by the system during the current access process of the target user. It describes the objective environment such as the device, network, and time of the current access. The collection of this data does not involve specific content input by the user; it only records the environmental parameters at the time of the access.

[0081] Interaction behavior data refers to the information automatically recorded by the system during a user's current visit, describing the interaction characteristics between the user and the business system, such as page navigation and timing. This data collection only involves macroscopic statistical characteristics of the visit behavior and does not involve the specific content of the user's interaction with the system, nor does it constitute monitoring or tracking of user operations.

[0082] Device fingerprints are information used to uniquely identify the terminal device currently being accessed by a target user. Device fingerprints are generated based on device characteristics such as hardware information and operating system version, and are used in subsequent steps to determine whether the currently accessed device is a frequently used device by the target user. Device fingerprints do not involve personal user identity information and do not have the ability to track individual users.

[0083] A login timestamp refers to the moment when a target user initiates their current access. It is used in subsequent steps to determine whether the current access time falls within the target user's regular access period. The login timestamp only records the moment the access occurred and is not involved in the continuous tracking of user behavior.

[0084] A page access sequence refers to the order in which a target user visits pages during an online visit. The page access sequence only records the order of page navigation; it does not record the specific actions the user takes on the page, the information they input, or the details of their behavior while on the page, and therefore does not constitute monitoring of user actions.

[0085] Page dwell time refers to the amount of time a target user spends on each page during a visit, describing the user's activity rhythm. Page dwell time only records time intervals and does not involve the specific content of the user's actions on the page.

[0086] Operation interval duration refers to the time interval between two consecutive operations performed by a target user during access, used to describe the user's operation rhythm. Operation interval duration only records the time interval and does not involve the specific content of the user's operation.

[0087] By setting login environment data and interaction behavior data as specific content of access feature data, a data foundation is provided for subsequent determination of state-related access features, determination of contribution offset, and deduction of contribution offset from the initial feature offset, ensuring the accuracy and reliability of the deduction operation. Login environment data describes the objective environment of the current access, such as the device, network, and time, while interaction behavior data describes the user's interaction characteristics during the access process, such as page navigation and timing. The combination of these two data points comprehensively reflects the characteristics of the current access from both environmental and interaction dimensions, ensuring the accurate execution of subsequent operations.

[0088] Optionally, determine the status-associated access features corresponding to the account security status from the access feature data, including: When the account security status is normal, the status-related access feature is set to empty; when the account security status is security information change status, the page access sequence is used as the status-related access feature; when the account security status is remote login status, at least one of IP address, login timestamp, and device fingerprint is used as the status-related access feature; when the account security status is new device login status, the device fingerprint is used as the status-related access feature.

[0089] In one implementation, after identifying the current account security status of the target user, the system determines the corresponding status-related access features based on that status.

[0090] The system pre-configures a mapping relationship between account security status and state-related access characteristics. When the target user's current account security status is normal, the system sets the state-related access characteristics to empty. Under normal status, all access characteristics are within the normal range, and there are no expected characteristic changes due to status changes. Therefore, there is no need to mark any characteristics as state-related access characteristics, meaning there are no objects to deduct contribution offsets.

[0091] When an account's security status is in a "security information changed" state, the system uses the page access sequence as a state-related access feature. After the security information changes, the user may need to complete the security verification process again, causing their page access sequence to change compared to the sequence in the normal access mode. Therefore, the page access sequence is a predictable change in this state and should be marked as a state-related access feature and deducted in subsequent steps.

[0092] When an account's security status indicates a login from a different location, the system will use at least one of the following as a state-related access characteristic: IP address, login timestamp, and device fingerprint. In a login from a different location scenario, the user's IP address may change, the login time may deviate from the usual time period due to time zone differences, and the device may differ from the user's usual device. These characteristics are all expected changes in this scenario and should be marked as state-related access characteristics and deducted in subsequent steps. The system can select one or more of the above characteristics as state-related access characteristics based on the actual access situation.

[0093] When the account security status is "new device login," the system uses the device fingerprint as a state-related access feature. In the new device login scenario, the user's terminal device is different from their usual device, and the change in device fingerprint is an expected change in this scenario. Therefore, it should be marked as a state-related access feature and deducted in subsequent steps.

[0094] By configuring corresponding status-related access features for different account security states, the objects for determining contribution offsets can be accurately matched for different access scenarios. This ensures that only features that can be expected to change in the current scenario are included in the calculation of contribution offsets, avoiding the inclusion of non-attributable features or the omission of attributable features. This ensures that the operation of deducting contribution offsets is adapted to the current access scenario, improving the accuracy and relevance of deducting contribution offsets.

[0095] Example 2 Figure 2 This is a flowchart of an initial feature offset determination method provided in Embodiment 2 of the present invention. It further optimizes and extends the implementation method of "inputting access feature data into the target anomaly detection model for anomaly detection to obtain the initial feature offset" in the above technical solution, and can be combined with the various optional implementation methods described above. For example... Figure 2 As shown, the method includes: S201. The access feature data is detected by the point anomaly detection channel of the target anomaly detection model to obtain the first feature offset.

[0096] The point anomaly detection channel refers to the detection channel in the target anomaly detection model used to detect whether a single data point in the access feature data deviates from the normal access feature baseline. This channel processes each access feature data point as an independent data point, and determines whether an anomaly exists by analyzing whether the position of this data point in the feature space deviates from the distribution area of ​​normal data points.

[0097] The first feature offset is used to characterize the degree of feature offset of the accessed feature data relative to the normal accessed feature baseline in the data point dimension.

[0098] The data point dimension refers to the detection dimension that measures the degree of deviation of access feature data relative to the normal access feature baseline from the perspective of a single access sample. Under this dimension, each access is regarded as an independent data point, and the system determines whether the position of the data point in the feature space deviates from the distribution area of ​​normal data points, so as to measure the degree of abnormality of the current access at the single sample level.

[0099] Feature offset refers to the quantified offset of access feature data relative to the baseline of normal access features in a certain detection dimension. This degree is used to measure the magnitude of the difference between the current access and the normal access pattern in a specific detection dimension, and it serves as the basis for subsequent deduction of contributing offsets and anomaly detection.

[0100] In one implementation, the point anomaly detection channel is based on the Isolation Forest algorithm. During the training phase, the system constructs an Isolation Forest model based on the historical access feature data of the target user in the corresponding state.

[0101] During the detection phase, the system inputs the access feature data of the current visit into the isolated forest model, causing the sample to traverse each isolated tree and record the path length from the root node to a leaf node in each tree. The path lengths across all trees are then averaged to obtain the average path length. Abnormal data points are more easily isolated in the feature space, resulting in shorter average path lengths; normal data points have longer average path lengths. The system maps the average path length to a first feature offset, which characterizes the degree of feature shift of the access feature data relative to the normal access feature baseline in the data point dimension. The shorter the average path length, the higher the first feature offset, indicating a greater degree of feature shift in the data point dimension for the current visit.

[0102] It should be noted that the above implementation method is only an illustrative example, and those skilled in the art can also use other point anomaly detection algorithms to implement this channel, which does not constitute a limitation on the present invention.

[0103] S202. The access feature data is detected by the sequence anomaly detection channel of the target anomaly detection model to obtain the second feature offset.

[0104] The sequence anomaly detection channel refers to the detection channel in the target anomaly detection model used to detect whether the operation sequence in the access feature data deviates from the normal access feature baseline. This channel treats the operation sequence in the current access process as a whole, and determines whether anomalies exist by analyzing whether the sequence features such as the transition pattern and order regularity of the operation sequence deviate from the normal sequence pattern. Unlike the point anomaly detection channel, which focuses on a single data point, the sequence anomaly detection channel focuses on the sequential relationship and transition regularity between operations.

[0105] The second feature offset is used to characterize the degree of feature offset of the accessed feature data relative to the normal accessed feature baseline in the data sequence dimension.

[0106] The data sequence dimension refers to the detection dimension that measures the degree of deviation of access feature data from the normal access feature baseline from the perspective of operation sequence. Under this dimension, the user's operations during an access process are regarded as an ordered sequence. The system judges whether the transition pattern, order rule, etc. of the operation sequence deviate from the normal sequence pattern in order to measure the degree of abnormality of the current access at the sequence level.

[0107] In one implementation, the sequence anomaly detection channel is based on a long short-term memory network.

[0108] During the training phase, the system constructs a long short-term memory network sequence model based on the target user's historical access feature data in corresponding states. This model establishes the pattern characteristics of normal access sequences by learning the temporal dependencies and transition rules between operations in the historical access sequences.

[0109] During the detection phase, the system uses the access feature data generated during the current access process as the input sequence and feeds it into a trained Long Short-Term Memory (LSTM) network model. The LTM network model processes the current operation sequence step by step and outputs a probability value for the current operation sequence based on the learned normal sequence pattern. This probability value reflects the degree to which the current operation sequence conforms to the normal sequence pattern—the higher the probability value, the more the current operation sequence conforms to the normal sequence pattern, and the lower the degree of sequence anomalousness; the lower the probability value, the more the current operation sequence deviates from the normal sequence pattern, and the higher the degree of sequence anomalousness.

[0110] The system determines the second feature offset based on the probability value output by the Long Short-Term Memory network model. The second feature offset characterizes the degree of feature shift of the accessed feature data relative to the baseline of normal access features in the data sequence dimension—the lower the probability value, the higher the second feature offset, indicating a greater degree of feature shift in the data sequence dimension for the current access.

[0111] It should be noted that the above implementation based on Long Short-Term Memory (LSTM) networks is only one possible example. In some states, the sequence anomaly detection channel can also be implemented using Markov transition matrices, determining whether the current operation sequence conforms to the typical access sequence of that state based on the state transition probabilities. Those skilled in the art can choose appropriate sequence anomaly detection algorithms to implement this channel according to actual application scenarios, which does not constitute a limitation of the present invention.

[0112] S203. The access feature data is detected by the statistical distribution deviation detection channel of the target anomaly detection model to obtain the third feature offset.

[0113] The statistical distribution deviation detection channel refers to the detection channel in the target anomaly detection model used to detect whether the overall statistical distribution of access feature data deviates from the normal access feature baseline. Unlike the point anomaly detection channel, which focuses on anomalies in individual data points, and the sequence anomaly detection channel, which focuses on anomalies in the sequence of operations, the statistical distribution deviation detection channel focuses on the distribution characteristics of access feature data at the overall statistical level. It determines whether anomalies exist by analyzing whether the position of the currently accessed feature vector in the multidimensional feature space deviates from the distribution center of normal data points.

[0114] The third feature offset is used to characterize the degree of feature offset of the accessed feature data relative to the normal accessed feature baseline in the statistical distribution dimension.

[0115] The statistical distribution dimension refers to the detection dimension that measures the degree of deviation of access feature data from the baseline of normal access features from the perspective of overall statistical distribution. Under this dimension, multiple access features of the current access are regarded as a whole feature vector. The system determines whether there is an anomaly in the current access by calculating the distance of this feature vector to the center of the distribution of normal data points. This dimension focuses on the position of the feature vector at the overall statistical level, rather than the anomaly of a single feature or sequence.

[0116] In one implementation, the statistical distribution deviation detection channel is based on Mahalanobis distance.

[0117] During the training phase, the system establishes a statistical distribution model of the normal access feature baseline based on the historical access feature data of the target user under corresponding states. Specifically, the system calculates the mean vector μ and covariance matrix Σ for each access feature, where the mean vector μ reflects the average value of each access feature under normal states, and the covariance matrix Σ reflects the correlation between the access features.

[0118] During the detection phase, the system uses the access feature data of the current access as a multidimensional sample x, and calculates the Mahalanobis distance from this sample to the center of the normal access feature baseline distribution μ: ; Where x is a multidimensional vector composed of the access feature data of the current visit, μ is the mean vector of the normal access feature baseline, and Σ is the covariance matrix. Mahalanobis distance takes into account the correlation between each access feature through the covariance matrix, and can accurately measure the degree of deviation of the current access feature vector from the normal access feature baseline at the overall statistical distribution level. The larger the Mahalanobis distance, the farther the current visit deviates from the normal access feature baseline in the statistical distribution dimension.

[0119] The system normalizes the Mahalanobis distance, mapping it to the interval of 0 to 1, and uses this as the third feature offset. The third feature offset characterizes the degree of feature shift of the accessed feature data relative to the normal access feature baseline in the statistical distribution dimension—the larger the normalized Mahalanobis distance, the higher the third feature offset, indicating a greater degree of feature shift in the statistical distribution dimension of the current access.

[0120] It should be noted that the above implementation method based on Mahalanobis distance is only an optional example. Those skilled in the art can also use other statistical distribution deviation detection algorithms to implement this channel, which does not constitute a limitation on the present invention.

[0121] S204. Use the first feature offset, the second feature offset, and / or the third feature offset as the initial feature offset.

[0122] In one implementation, the system detects access feature data through a point anomaly detection channel to obtain a first feature offset; detects access feature data through a sequence anomaly detection channel to obtain a second feature offset; and detects access feature data through a statistical distribution deviation detection channel to obtain a third feature offset.

[0123] The system uses the first feature offset, the second feature offset, and / or the third feature offset as initial feature offsets, respectively. The phrase "as" means that the feature offsets output by each detection channel are all components of the initial feature offsets, existing independently and collectively forming the initial feature offsets.

[0124] It should be noted that the feature offsets used by the system to constitute the initial feature offset can differ depending on the account's security status. For example, the system can select one or more feature offsets corresponding to the current account security status from the first, second, and third feature offsets to participate in the formation of the initial feature offset. Feature offsets not selected will not participate in the formation of the initial feature offset. The specific feature offsets selected can be configured according to actual business needs and security policies.

[0125] Anomaly detection is performed separately through point anomaly detection channel, sequence anomaly detection channel, and statistical distribution deviation detection channel. Each channel obtains its own independent feature offset, which is then used as the initial feature offset. This preserves the independent detection information of each detection channel and avoids information loss caused by premature merging of multi-dimensional anomaly information. It provides a more comprehensive basis for determining the contribution offset and calculating the remaining feature offset, thereby improving the accuracy of anomaly identification.

[0126] Example 3 Figure 3 This is a flowchart of a contribution offset determination method provided in Embodiment 3 of the present invention. It further optimizes and extends the implementation method of "determining the contribution offset of the state-related access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected distribution of the feature value" in the above technical solution, and can be combined with the various optional implementation methods described above. For example... Figure 3 As shown, the method includes: S301. Based on the degree of deviation and the splitting frequency of the state-related access feature in the point anomaly detection channel, determine the first contribution offset of the state-related access feature to the first feature offset.

[0127] Split frequency refers to the number of times a state-related access feature is selected as a split node in each isolated tree splitting path of the point anomaly detection channel. A sample passes through multiple split nodes on its path from the root node to a leaf node in each isolated tree, with each split node corresponding to a feature used for splitting. The split frequency of each feature is obtained by counting the number of split nodes the sample passes through in all isolated trees. The more times a feature is selected as a split node, the greater its contribution to anomaly detection.

[0128] The first contribution offset refers to the contribution of the state-associated access feature to the first feature offset, that is, the portion of the first feature offset that can be attributed to the state-associated access feature offset. The first contribution offset is determined by the degree of deviation and the split frequency. The degree of deviation reflects whether the actual feature value meets expectations, and the split frequency reflects the influence of the feature on anomaly judgment in the point anomaly detection channel. The two are combined to determine the offset that should be deducted from the first feature offset.

[0129] In one implementation, the system inputs the access feature data of the current visit into the isolated forest model to obtain a first feature offset. During this process, the system simultaneously records each split node along the path the sample takes from the root node to a leaf node, and the feature corresponding to each split node. The system counts the number of times the state-associated access feature is selected as a split node in all isolated trees, obtaining the split frequency.

[0130] Specifically, the system determines the first contribution offset by combining the degree of deviation and the split frequency as follows: The system multiplies the degree of deviation by the normalized value of the split frequency to obtain the first contribution offset. The normalized value of the split frequency is obtained by dividing the split frequency of each state-related access feature by the total number of split nodes traversed by the sample across all trees. A higher degree of deviation indicates that the actual value of the state-related access feature is less consistent with expectations in the current scenario; a higher split frequency indicates that the feature has a greater impact on anomaly detection in the point anomaly detection channel. Multiplying the two ensures that the first contribution offset reflects both the interpretability of the feature at the business scenario level (degree of deviation) and the actual contribution of the feature at the model level (split frequency), accurately quantifying the offset that should be deducted from the first feature offset.

[0131] It should be noted that the above combination method is merely illustrative, and those skilled in the art can also use other methods to combine the degree of deviation and the split frequency (e.g., weighted summation), which does not constitute a limitation of the present invention. During the traversal of the isolated tree, the sample can record the path length for calculating the first feature offset, or simultaneously count the number of times each feature is selected as a split node to determine the split frequency. These are different pieces of information obtained in the same detection process and do not conflict.

[0132] S302. Based on the degree of deviation and the gradient sensitivity of the state-related access feature to the sequence anomaly detection channel, determine the second contribution offset of the state-related access feature to the second feature offset.

[0133] Gradient sensitivity refers to the degree of influence of state-related access features on the output of the sequence anomaly detection channel, and is used to measure the magnitude of the impact of changes in this feature on the sequence anomaly detection results. When the sequence anomaly detection channel is implemented based on a long short-term memory network, gradient sensitivity is obtained by calculating the gradient of the model output with respect to the input feature—the larger the gradient value, the greater the impact of changes in this feature on the model output, and the greater the contribution of this feature to the second feature offset.

[0134] The second contribution offset refers to the portion of the second feature offset that is attributable to the state-associated access feature. The second contribution offset is determined by both the degree of deviation and gradient sensitivity—the degree of deviation reflects whether the state-associated access feature meets expectations, and gradient sensitivity reflects the influence of this feature on anomaly detection in the sequence anomaly detection channel. Combining these two factors determines the offset that should be subtracted from the second feature offset.

[0135] In one implementation, the system determines the gradient sensitivity of state association access features to the sequence anomaly detection channel based on the detection results of the long short-term memory network model.

[0136] During the detection phase, the system inputs the access feature data generated during the current access process into a pre-trained Long Short-Term Memory (LSTM) network model. The LTM model processes the current operation sequence step-by-step, outputting a probability value for the current operation sequence based on the learned normal sequence pattern. Furthermore, the system calculates the gradient of the model output with respect to each input feature. The gradient value reflects the degree to which changes in each input feature affect the model's output—a larger gradient value indicates that the feature is more sensitive to the model's output and contributes more to the second feature's offset. The gradient value corresponding to the state-related access feature is the gradient sensitivity of that feature.

[0137] Specifically, the system determines the second contribution offset by combining the degree of deviation and gradient sensitivity as follows: The system multiplies the degree of deviation by the normalized value of the gradient sensitivity to obtain the second contribution offset. The normalized value of the gradient sensitivity is obtained by dividing each feature gradient value by the sum of all feature gradient values. The degree of deviation reflects whether the state-related access feature meets expectations in the current access scenario, while the gradient sensitivity reflects the actual impact of the feature on anomaly detection in the sequence anomaly detection channel. Multiplying the two ensures that the second contribution offset reflects both the interpretability of the feature at the business scenario level (degree of deviation) and the actual contribution of the feature at the model level (gradient sensitivity), accurately quantifying the offset that should be deducted from the second feature offset.

[0138] It should be noted that the above combination method is merely illustrative, and those skilled in the art can also use other methods to combine deviation degree and gradient sensitivity (e.g., weighted summation), which does not constitute a limitation of the present invention. In some states, when the sequence anomaly detection channel is implemented using a Markov transition matrix, gradient sensitivity can be replaced by sensitivity analysis of state transition probabilities. Those skilled in the art can choose an appropriate gradient sensitivity determination method according to the actual application scenario, which does not constitute a limitation of the present invention.

[0139] S303. Based on the degree of deviation, determine the third contribution offset of the state-related access feature to the third feature offset.

[0140] The third contribution offset refers to the portion of the third feature offset that can be attributed to the state-related access feature. It is determined by separating it from the third feature offset based on the degree of deviation and is used to deduct it from the third feature offset in subsequent steps.

[0141] In one implementation, the system determines a third contribution offset based on a determined degree of deviation.

[0142] Specifically, the degree of deviation reflects the difference between the actual feature value of the state-associated access feature and the expected distribution of feature values. The third contribution offset is directly determined by the degree of deviation—the higher the degree of deviation, the less the actual value of the state-associated access feature conforms to the expectations under the current access scenario, and the larger the corresponding third contribution offset. The system multiplies the degree of deviation by a preset mapping coefficient to obtain the third contribution offset. The mapping coefficient can be configured according to actual business needs and security policies.

[0143] It should be noted that the mapping method between the degree of deviation and the third contribution offset is only an illustrative example. Those skilled in the art can also use nonlinear mapping or other mapping methods to convert the degree of deviation into the third contribution offset, which does not constitute a limitation on the present invention.

[0144] By targeting the detection characteristics of the point anomaly detection channel, sequence anomaly detection channel, and statistical distribution deviation detection channel respectively, and combining the degree of deviation with the internal contribution metrics of each channel (splitting frequency, gradient sensitivity), the contribution offset of each channel is determined. This ensures the accuracy and reliability of the contribution offset calculation for each channel, so that the deduction of contribution offset for each channel can accurately reflect the actual contribution of the state association access feature in that channel.

[0145] Optionally, the contributing offset is subtracted from the initial feature offset to obtain the remaining feature offset, including: Subtract the first contribution offset from the first feature offset to obtain the first residual feature offset; subtract the second contribution offset from the second feature offset to obtain the second residual feature offset; subtract the third contribution offset from the third feature offset to obtain the third residual feature offset; and sum the first, second, and third residual feature offsets by weight to obtain the residual feature offset.

[0146] In one implementation, the system performs subtraction operations on the three detection channels respectively: subtracting the first contribution offset from the first feature offset to obtain the first remaining feature offset; subtracting the second contribution offset from the second feature offset to obtain the second remaining feature offset; and subtracting the third contribution offset from the third feature offset to obtain the third remaining feature offset.

[0147] The system performs a weighted sum of the first, second, and third residual feature offsets to obtain the residual feature offset: ; in, This is the offset of the first remaining feature. This is the offset of the second remaining feature. This is the offset of the third remaining feature. , , These are the corresponding weighting coefficients. The weighting coefficients reflect the importance of each detection channel in the overall anomaly detection and can be configured according to the actual business scenario and security strategy.

[0148] This refers to the remaining feature offset, which is the residual offset that cannot be explained by the current account security status after deducting the contribution offsets attributable to the access features associated with each state from the initial feature offset. It is used to determine whether the target user's account is abnormal.

[0149] By using this weighted summation method, the system merges the remaining results of the three channels after deducting the contribution offset into a comprehensive remaining feature offset, which is used to determine whether the target user's account is abnormal.

[0150] It should be noted that the above weighted summation method is only an illustrative example. Those skilled in the art can also use other fusion methods to combine the three remaining feature offsets into a single remaining feature offset, which does not constitute a limitation on the present invention.

[0151] By subtracting the corresponding contribution offsets from the first, second, and third feature offsets respectively, we obtain the first, second, and third residual feature offsets. Then, we weight and sum the three to obtain the residual feature offset. This allows the importance of each detection channel in the overall anomaly determination to be reflected differently, avoiding information distortion caused by treating each channel equally. It ensures that the calculation of the residual feature offset can comprehensively reflect the detection results of the three channels, providing an accurate basis for account anomaly determination.

[0152] Example 4 Figure 4 This is a flowchart of a comprehensive feature offset determination method provided in Embodiment 4 of the present invention. It further optimizes and expands the implementation method of "determining whether the target user's account is abnormal based on the remaining feature offset" in the above technical solution, and can be combined with the various optional implementation methods described above. For example... Figure 4 As shown, the method includes: S401. Determine at least one dynamic rule and the corresponding rule feature offset based on the account security status.

[0153] Dynamic rules refer to rules that dynamically adjust judgment conditions and offsets based on the target user's current account security status. The judgment conditions and corresponding rule feature offsets for the same dynamic rule can differ under different account security statuses. For example, under normal circumstances, a new IP address login might correspond to a higher rule feature offset; while under circumstances of login from a different location, a new IP address login is an expected change under that status, and its corresponding rule feature offset is lower or even zero. Through dynamic rules, the system can further correct the remaining feature offsets at the rule level, improving the accuracy of anomaly detection.

[0154] Rule feature offset refers to the offset value corresponding to a dynamic rule when the target user's current access matches that rule. The rule feature offset is added to the remaining feature offsets in the calculation of the comprehensive feature offset to correct the anomaly detection result. Under different account security states, the same dynamic rule can correspond to different rule feature offsets, thus achieving adaptive adjustment at the rule level.

[0155] In one implementation, the system determines at least one dynamic rule that is in effect under the current account security status of the target user, as well as the rule feature offset corresponding to each dynamic rule, based on the target user's current account security status.

[0156] Taking remote login as an example, the system can configure the following dynamic rules: When the currently accessed IP address does not belong to the commonly used IP range, the "New IP Login" rule is triggered, and the rule feature offset corresponding to this rule in remote login is the first value; when the currently accessed login timestamp does not belong to the commonly used access time period, the "Access During Non-Working Hours" rule is triggered, and the rule feature offset corresponding to this rule in remote login is the second value; when the degree to which the currently accessed page access sequence deviates from the normal sequence pattern exceeds a preset threshold, the "Sequence Anomaly" rule is triggered, and the rule feature offset corresponding to this rule in remote login is the third value. The system sums the rule feature offsets corresponding to each dynamic rule to obtain the total rule feature offset.

[0157] It should be noted that the above configuration of dynamic rules is only an illustrative example. Those skilled in the art can configure different dynamic rules and their corresponding rule feature offsets according to actual business needs and security policies, which does not constitute a limitation on the present invention.

[0158] S402. Match the access feature data with each dynamic rule, and determine the matching dynamic rule from each dynamic rule.

[0159] In one implementation, the system matches the target user's access feature data with each dynamic rule one by one, based on the determined dynamic rules.

[0160] Specifically, the system retrieves various information from the access feature data of the current access. For each dynamic rule, the system determines whether the access feature data of the current access meets the judgment conditions of that dynamic rule.

[0161] For example, regarding IP address-related rules, the system compares the currently accessed IP address with commonly used IP ranges in the normal access feature baseline. If it does not belong to a commonly used IP range, the rule is matched. For example, regarding login time-related rules, the system compares the currently accessed login timestamp with commonly used time periods in the normal access feature baseline. If it is not in a commonly used time period, the rule is matched. For example, regarding device fingerprint-related rules, the system compares the currently accessed device fingerprint with the list of commonly used devices in the normal access feature baseline. If it does not appear in the list of commonly used devices, the rule is matched. For example, regarding sequence anomaly-related rules, the system compares the anomaly level of the currently accessed page access sequence with a preset threshold. If it exceeds the preset threshold, the rule is matched.

[0162] The system identifies the matching dynamic rules from among the various dynamic rules and obtains the rule feature offset corresponding to the matching rule. Unmatched dynamic rules are not included in the calculation of the total rule feature offset for subsequent rules.

[0163] S403. Determine the total number of rule feature offsets based on the rule feature offsets corresponding to the hit dynamic rules, and add the remaining feature offsets to the total number of rule feature offsets to obtain the comprehensive feature offset.

[0164] In one implementation, the system obtains the rule feature offset corresponding to each hit rule based on the determined dynamic rules, and accumulates the rule feature offsets corresponding to each hit rule to obtain the total rule feature offset.

[0165] For example, if the system determines that the hit dynamic rules include rule A, rule B and rule C, then the system obtains the rule feature offset corresponding to rule A, the rule feature offset corresponding to rule B and the rule feature offset corresponding to rule C, and adds the three together to obtain the total rule feature offset.

[0166] The system adds the remaining feature offset to the total rule feature offset to obtain the comprehensive feature offset. The comprehensive feature offset is a comprehensive judgment criterion that integrates the remaining feature offset at the model level and the correction information at the rule level, and is used to determine whether the target user's account is abnormal.

[0167] S404. Determine whether the target user's account is abnormal based on the comprehensive feature offset.

[0168] In one implementation, the system compares the comprehensive feature offset with a preset threshold range, and determines whether the target user's account is abnormal based on the threshold range in which the comprehensive feature offset is located.

[0169] Specifically, the system is pre-configured with a first threshold and a second threshold, where the first threshold is lower than the second threshold. The system compares the comprehensive feature offset with the first threshold and the second threshold: If the overall feature offset is lower than the first threshold, the system determines that the current access is low-risk, allows the access, and records the attribution information of the current access, including the account security status corresponding to the current access, the access characteristics associated with the status, the actual feature value, the degree of deviation, the contribution offset, the remaining feature offset, and the overall feature offset.

[0170] If the overall feature offset is between the first and second thresholds, the system determines the current access to be of medium risk and triggers two-factor authentication. If the two-factor authentication is successful, the system adds the access feature data of this access to the normal fluctuation sample pool for subsequent updates to the normal access feature baseline and optimization of the expected distribution of feature values; if the two-factor authentication fails, the system marks the current access as abnormal.

[0171] If the overall feature offset is higher than the second threshold, the system determines that the current access is high-risk, blocks the current access, freezes the target user's session, and pushes it to the security management center to mark the current access as abnormal.

[0172] It should be noted that the first threshold, the second threshold, and the corresponding handling strategies for each risk level are merely illustrative examples. Those skilled in the art can configure different thresholds and handling strategies according to actual business needs and security policies, which does not constitute a limitation on the present invention.

[0173] Based on the account security status, dynamic rules and their corresponding rule feature offsets are determined. The rule feature offsets corresponding to the hit dynamic rules are accumulated to obtain the total rule feature offset, and then added to the remaining feature offsets to obtain the comprehensive feature offset. This allows the anomaly detection to integrate the remaining feature offsets at the model level and the correction information at the rule level, achieving adaptive matching between dynamic rules and access scenarios, and further improving the accuracy of account anomaly identification.

[0174] As an optional embodiment, the system may also include a model optimization process.

[0175] In one optional implementation, after triggering two-factor authentication, if the authentication passes and the overall feature offset is within a preset medium-risk range (e.g., between a first threshold and a second threshold), it indicates that although the current access triggered a medium-risk assessment, it was confirmed as an operation by the user after two-factor authentication, and the system classifies it as a false alarm. The system analyzes the feature dimension that contributes the most to the current access: if this dimension belongs to the state-related access feature under the current account security status, the system increases the weight of this feature in the subsequent contribution offset determination process, so that similar fluctuations will be attributed to changes in account security status rather than anomalies in the future; if this dimension does not belong to the state-related access feature under the current account security status, the system adds the access feature data of the current access to the normal fluctuation sample pool for subsequent periodic retraining of the model in the point anomaly detection channel and / or sequence anomaly detection channel.

[0176] In another optional implementation, when the target user's account security status changes (e.g., switching from a remote login status to a normal status), the system aggregates the access feature data collected during the status change period and updates the anomaly detection model used in the corresponding status, making subsequent detections more closely match the target user's actual access feature patterns in that status. For example, when the remote login status ends, the system uses the access feature data collected during that remote login period to update the third model corresponding to the remote login status.

[0177] Through the above model optimization process, the system can continuously optimize model parameters based on false alarm analysis and state changes, enabling the anomaly detection model to adapt to changes in the access characteristics of target users and continuously improve the accuracy of anomaly identification.

[0178] Example 5 Figure 5 This is a schematic diagram of the structure of an account anomaly identification device provided in Embodiment 5 of the present invention. It is applicable to scenarios involving real-time analysis of user access behavior in a business system to identify whether an account has been stolen or misused, or other abnormal risks. Figure 5 As shown, the device includes: The access response module 51 is used to respond to the target user's current access to the business system, obtain the target user's account environment information, and identify the target user's current account security status based on the account environment information. The feature data acquisition module 52 is used to select the corresponding target anomaly detection model from the candidate anomaly detection models according to the account security status, and acquire the access feature data generated by the target user during the current access process. The feature offset determination module 53 is used to input the access feature data into the target anomaly detection model for anomaly detection to obtain an initial feature offset; wherein, the initial feature offset is used to characterize the feature offset of the access feature data relative to the normal access feature baseline of the target user. The feature value acquisition module 54 is used to determine the state-related access feature corresponding to the account security status from the access feature data, and to acquire the actual feature value corresponding to the state-related access feature; wherein, the state-related access feature is a feature that is expected to change with the change of the account security status. The offset calculation module 55 is used to determine the expected distribution of the feature values ​​of the state-related access feature when the target user is in the account security state, and to determine the contribution offset of the state-related access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected distribution of the feature value. The account anomaly judgment module 56 is used to subtract the contribution offset from the initial feature offset to obtain the remaining feature offset, and to judge whether the target user's account is abnormal based on the remaining feature offset.

[0179] Optionally, the feature offset determination module 53 is specifically used for: The access feature data is detected by the point anomaly detection channel of the target anomaly detection model to obtain a first feature offset; wherein, the first feature offset is used to characterize the degree of feature offset of the access feature data relative to the normal access feature baseline in the data point dimension. The access feature data is detected by the sequence anomaly detection channel of the target anomaly detection model to obtain a second feature offset; wherein, the second feature offset is used to characterize the degree of feature offset of the access feature data relative to the normal access feature baseline in the data sequence dimension; The access feature data is detected by the statistical distribution deviation detection channel of the target anomaly detection model to obtain a third feature offset; wherein, the third feature offset is used to characterize the degree of feature offset of the access feature data relative to the normal access feature baseline in the statistical distribution dimension. The first feature offset, the second feature offset, and / or the third feature offset are used as the initial feature offset.

[0180] Optionally, the offset calculation module 55 is specifically used for: Based on the degree of deviation and the splitting frequency of the state-related access feature in the point anomaly detection channel, the first contribution offset of the state-related access feature to the first feature offset is determined. Based on the degree of deviation and the gradient sensitivity of the state-associated access feature to the sequence anomaly detection channel, a second contribution offset of the state-associated access feature to the second feature offset is determined. Based on the degree of deviation, determine the third contribution offset of the state-associated access feature to the third feature offset.

[0181] Optionally, the account anomaly detection module 56 is specifically used for: Subtract the first contribution offset from the first feature offset to obtain the first remaining feature offset; Subtract the second contribution offset from the second feature offset to obtain the second residual feature offset; The third residual feature offset is obtained by subtracting the third contribution offset from the third feature offset. The remaining feature offset is obtained by weighted summing of the first remaining feature offset, the second remaining feature offset, and the third remaining feature offset.

[0182] Optionally, the account security status includes normal status, security information change status, login status from another location, and login status from a new device; the candidate anomaly detection model includes a first model, a second model, a third model, and a fourth model; The first model is trained based on historical access feature data collected from the target user under normal conditions; the second model is trained based on historical access feature data collected from the target user under security information change conditions; the third model is trained based on historical access feature data collected from the target user under remote login conditions; and the fourth model is trained based on historical access feature data collected from the target user under new device login conditions. Optionally, the feature data acquisition module 52 is specifically used for: When the account security status is normal, the first model is used as the target anomaly detection model; When the account security status is the security information change status, the second model is used as the target anomaly detection model; When the account security status is the remote login status, the third model will be used as the target anomaly detection model. When the account security status is the new device login status, the fourth model is used as the target anomaly detection model.

[0183] Optionally, the account anomaly detection module 56 is specifically used for: Based on the account security status, at least one dynamic rule and the rule feature offset corresponding to each dynamic rule are determined; The access feature data is matched with each of the dynamic rules, and the matching dynamic rule is determined from each of the dynamic rules; The total number of rule feature offsets is determined based on the rule feature offsets corresponding to the hit dynamic rules, and the remaining feature offsets are added to the total number of rule feature offsets to obtain the comprehensive feature offset. The target user's account is determined to be abnormal based on the comprehensive feature offset.

[0184] Optionally, the access feature data includes at least one of login environment data and interaction behavior data; The login environment data includes at least one of device fingerprint, IP address, and login timestamp; The interactive behavior data includes at least one of page access sequence, page dwell time, and operation interval duration.

[0185] Optionally, the feature value acquisition module 54 is specifically used for: When the account security status is normal, the status-associated access feature is set to empty; When the account security status is in the security information change state, the page access sequence will be used as the status-associated access feature; When the account security status is a remote login status, at least one of the IP address, the login timestamp, and the device fingerprint will be used as the status-associated access feature; When the account security status is new device login status, the device fingerprint is used as the status-associated access feature.

[0186] The account anomaly identification device provided in this embodiment of the invention can execute the account anomaly identification method provided in this embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0187] According to embodiments of this disclosure, embodiments of the present invention also provide an electronic device, a readable storage medium, and a computer program product.

[0188] Figure 6 A schematic diagram of an electronic device 60 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0189] like Figure 6 As shown, the electronic device 60 includes at least one processor 61 and a memory, such as a read-only memory (ROM) 62 and a random access memory (RAM) 63, communicatively connected to the at least one processor 61. The memory stores computer programs executable by the at least one processor. The processor 61 can perform various appropriate actions and processes based on the computer program stored in the ROM 62 or loaded from storage unit 68 into the RAM 63. The RAM 63 can also store various programs and data required for the operation of the electronic device 60. The processor 61, ROM 62, and RAM 63 are interconnected via a bus 64. An input / output (I / O) interface 65 is also connected to the bus 64.

[0190] Multiple components in electronic device 60 are connected to I / O interface 65, including: input unit 66, such as keyboard, mouse, etc.; output unit 67, such as various types of monitors, speakers, etc.; storage unit 68, such as disk, optical disk, etc.; and communication unit 69, such as network card, modem, wireless transceiver, etc. Communication unit 69 allows electronic device 60 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0191] Processor 61 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 61 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 61 performs the various methods and processes described above, such as account anomaly detection methods.

[0192] In some embodiments, the account anomaly detection method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 68. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 60 via ROM 62 and / or communication unit 69. When the computer program is loaded into RAM 63 and executed by processor 61, one or more steps of the account anomaly detection method described above may be performed. Alternatively, in other embodiments, processor 61 may be configured to perform the account anomaly detection method by any other suitable means (e.g., by means of firmware).

[0193] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0194] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0195] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0196] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0197] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0198] A computing system can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product within the cloud computing service system to address the shortcomings of traditional physical hosts and virtual private servers, such as high management difficulty and weak business scalability.

[0199] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0200] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for identifying account abnormality, characterized in that, include: In response to a target user's current access to the business system, the system obtains the target user's account environment information and identifies the target user's current account security status based on the account environment information. Based on the account security status, select the corresponding target anomaly detection model from the candidate anomaly detection models, and obtain the access feature data generated by the target user during the current access process; The access feature data is input into the target anomaly detection model for anomaly detection to obtain an initial feature offset; wherein, the initial feature offset is used to characterize the feature offset of the access feature data relative to the normal access feature baseline of the target user; Determine the state-related access feature corresponding to the account security status from the access feature data, and obtain the actual feature value corresponding to the state-related access feature; wherein, the state-related access feature is a feature that is expected to change with the change of the account security status; Determine the expected distribution of the feature values ​​of the state-associated access feature when the target user is in the account security state, and determine the contribution offset of the state-associated access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected distribution of the feature value; The contribution offset is subtracted from the initial feature offset to obtain the remaining feature offset, and the target user's account is determined to be abnormal based on the remaining feature offset.

2. The method of claim 1, wherein, The step of inputting the access feature data into the target anomaly detection model for anomaly detection to obtain an initial feature offset includes: The access feature data is detected by the point anomaly detection channel of the target anomaly detection model to obtain a first feature offset; wherein, the first feature offset is used to characterize the degree of feature offset of the access feature data relative to the normal access feature baseline in the data point dimension. The access feature data is detected by the sequence anomaly detection channel of the target anomaly detection model to obtain a second feature offset; wherein, the second feature offset is used to characterize the degree of feature offset of the access feature data relative to the normal access feature baseline in the data sequence dimension; The access feature data is detected by the statistical distribution deviation detection channel of the target anomaly detection model to obtain a third feature offset; wherein, the third feature offset is used to characterize the degree of feature offset of the access feature data relative to the normal access feature baseline in the statistical distribution dimension. The first feature offset, the second feature offset, and / or the third feature offset are used as the initial feature offset.

3. The method of claim 2, wherein, The step of determining the contribution offset of the state-related access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected distribution of the feature value includes: Based on the degree of deviation and the splitting frequency of the state-related access feature in the point anomaly detection channel, the first contribution offset of the state-related access feature to the first feature offset is determined. Based on the degree of deviation and the gradient sensitivity of the state-associated access feature to the sequence anomaly detection channel, a second contribution offset of the state-associated access feature to the second feature offset is determined. Based on the degree of deviation, determine the third contribution offset of the state-associated access feature to the third feature offset.

4. The method of claim 3, wherein, The step of subtracting the contribution offset from the initial feature offset to obtain the remaining feature offset includes: Subtract the first contribution offset from the first feature offset to obtain the first remaining feature offset; Subtract the second contribution offset from the second feature offset to obtain the second residual feature offset; The third residual feature offset is obtained by subtracting the third contribution offset from the third feature offset. The remaining feature offset is obtained by weighted summing of the first remaining feature offset, the second remaining feature offset, and the third remaining feature offset.

5. The method of claim 1, wherein, The account security status includes normal status, security information change status, login from a different location, and login from a new device; the candidate anomaly detection models include a first model, a second model, a third model, and a fourth model; The first model is trained based on historical access feature data collected from the target user under normal conditions; the second model is trained based on historical access feature data collected from the target user under security information change conditions; the third model is trained based on historical access feature data collected from the target user under remote login conditions; and the fourth model is trained based on historical access feature data collected from the target user under new device login conditions. The step of selecting the corresponding target anomaly detection model from the candidate anomaly detection models based on the account security status includes: When the account security status is normal, the first model is used as the target anomaly detection model; When the account security status is the security information change status, the second model is used as the target anomaly detection model; When the account security status is the remote login status, the third model will be used as the target anomaly detection model. When the account security status is the new device login status, the fourth model is used as the target anomaly detection model.

6. The method of claim 1, wherein, The step of determining whether the target user's account is abnormal based on the remaining feature offset includes: Based on the account security status, at least one dynamic rule and the rule feature offset corresponding to each dynamic rule are determined; The access feature data is matched with each of the dynamic rules, and the matching dynamic rule is determined from each of the dynamic rules; The total number of rule feature offsets is determined based on the rule feature offsets corresponding to the hit dynamic rules, and the remaining feature offsets are added to the total number of rule feature offsets to obtain the comprehensive feature offset. The target user's account is determined to be abnormal based on the comprehensive feature offset.

7. The method of claim 1, wherein, The access feature data includes at least one of login environment data and interaction behavior data; The login environment data includes at least one of device fingerprint, IP address, and login timestamp; The interactive behavior data includes at least one of page access sequence, page dwell time, and operation interval duration.

8. The method of claim 7, wherein, Determining the state-associated access feature corresponding to the account security status from the access feature data includes: When the account security status is normal, the status-associated access feature is set to empty; When the account security status is in the security information change state, the page access sequence will be used as the status-associated access feature; When the account security status is a remote login status, at least one of the IP address, the login timestamp, and the device fingerprint will be used as the status-associated access feature; When the account security status is new device login status, the device fingerprint is used as the status-associated access feature.

9. An account abnormality recognition apparatus characterized by comprising: include: The access response module is used to respond to the target user's current access to the business system, obtain the target user's account environment information, and identify the target user's current account security status based on the account environment information. The feature data acquisition module is used to select the corresponding target anomaly detection model from the candidate anomaly detection models according to the account security status, and acquire the access feature data generated by the target user during the current access process. The feature offset determination module is used to input the access feature data into the target anomaly detection model for anomaly detection to obtain an initial feature offset; wherein, the initial feature offset is used to characterize the feature offset of the access feature data relative to the normal access feature baseline of the target user. The feature value acquisition module is used to determine the state-related access feature corresponding to the account security status from the access feature data, and to acquire the actual feature value corresponding to the state-related access feature; wherein, the state-related access feature is a feature that is expected to change with the change of the account security status. The offset calculation module is used to determine the expected distribution of the feature values ​​of the state-related access feature when the target user is in the account security state, and to determine the contribution offset of the state-related access feature to the initial feature offset based on the degree of deviation of the actual feature value from the expected distribution of the feature value. The account anomaly detection module is used to subtract the contribution offset from the initial feature offset to obtain the remaining feature offset, and to determine whether the target user's account is abnormal based on the remaining feature offset.

10. An electronic device, comprising: The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

11. A computer readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to perform the method of any one of claims 1-8.

12. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-8.