A security incident notification method and electronic device

By screening and integrating risk characteristic tags of security incidents and dynamically adjusting assessment strategies in conjunction with user behavior data, the problems of misjudgment and silent handling in risk assessment in traditional security systems have been solved, thereby improving the accuracy of assessment and adaptability to user needs.

CN123117782APending Publication Date: 2026-11-17BEIJING QIREN TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611272628.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional security systems are prone to inflated risk scores, misjudgments, or abnormally high alarm levels when multiple high-risk behavior tags are triggered simultaneously. Furthermore, they lack effective capture and utilization of users' proactive behaviors, making it difficult to dynamically optimize risk assessment strategies and adapt to changes in actual user needs.

Method used

By analyzing the raw characteristic data of security incidents through risk assessment strategies, the highest risk characteristic labels of each dimension are selected, non-linear fusion processing is performed to determine the total risk score, and user behavior data is collected under silent processing level to dynamically adjust the risk assessment strategy based on user behavior.

Benefits of technology

It improves the accuracy of risk assessment, reduces the probability of false alarms and false negatives, enhances the system's adaptability to real user needs, and avoids lag in response caused by fixed rules and over-response to global adjustments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117782A_ABST
    Figure CN123117782A_ABST
Patent Text Reader

Abstract

This application discloses a security incident notification method and electronic device, relating to the field of security monitoring technology. It analyzes the original characteristic data of security incidents based on a risk assessment strategy to determine multiple feature tags and their risk scores corresponding to each risk assessment dimension. The method selects the target feature tag with the highest risk score from each dimension, avoiding artificially high scores caused by the accumulation of multiple tags within the same dimension. The scores of each target tag are fused to obtain a total risk score, making the total score more closely reflect the actual threat level and reducing false alarms and false negatives. The risk level of the incident is determined based on the total risk score. When the incident is at the silent handling level, user behavior data is collected within a preset time window. Based on this behavior data and the incident's classification characteristics, the risk assessment strategy is adjusted so that subsequent new incidents with the same classification characteristics are reclassified according to the adjusted strategy, ensuring accuracy while improving the system's adaptability to real user needs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of security monitoring technology, and in particular to a security incident notification method and electronic device. Background Technology

[0002] Traditional solutions for risk scoring of security incidents typically use a linear summation of scores from multiple trigger tags to determine the total score. However, in complex scenarios where a single target simultaneously triggers multiple high-risk behavioral tags—for example, when the same event is simultaneously identified as multiple synonymous or overlapping high-risk semantic tags—this summation method can easily lead to an artificially inflated total risk score, resulting in misjudgments or abnormally high alarm levels, impacting system reliability and user experience.

[0003] Furthermore, traditional security systems often rely on preset static scoring mechanisms to determine the risk level of events and decide whether to push notifications. Events deemed low-risk are typically handled silently without proactive notifications. However, this static scoring mechanism overlooks the fact that users may still show active interest in certain silently handled events even without receiving notifications. Current solutions lack effective capture and utilization of such user behavior, making it difficult to dynamically optimize risk assessment strategies based on users' actual concerns. They rely heavily on fixed thresholds or manually adjusted rules, failing to adapt to changes in actual user needs and easily leading to missed detections of potentially high-risk events.

[0004] Therefore, improving the accuracy of risk assessment is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a security incident notification method and electronic device that can improve the accuracy of risk assessment.

[0006] This application provides a method for notifying security incidents, including: Based on the risk assessment strategy, the original characteristic data of security incidents are analyzed to determine multiple feature labels corresponding to each risk assessment dimension; among them, each feature label has its corresponding risk score. The target feature label with the highest risk score is selected from multiple feature labels corresponding to each risk assessment dimension; each risk assessment dimension has one corresponding target feature label. The risk scores of each target feature label are fused to obtain the total risk score of the security event, and the risk level of the security event is determined based on the total risk score; wherein, the risk level includes at least one silent handling level; When the risk level of a security incident is classified as a silent handling level, user behavior data of the security incident is collected within a preset time window. Based on user behavior data under the silent processing level and the classification feature information corresponding to security events, the risk assessment strategy corresponding to security events is adjusted, and the risk level of new security events with the same classification feature information is determined according to the adjusted risk assessment strategy.

[0007] On the one hand, based on user behavior data under the silent processing level and the classification characteristics of security incidents, the risk assessment strategy for security incidents is adjusted, including: Based on the semantic tags, spatiotemporal features, and device feature information corresponding to multiple silent events, multiple silent events are clustered to determine at least one event category; among them, silent events are security events with a risk level of silent handling level. Based on the user behavior data of multiple security events corresponding to the same event category, determine the behavioral statistics value of the event category; If the behavioral statistics meet the preset conditions, adjust the risk assessment strategy corresponding to the event category.

[0008] On the one hand, based on the user behavior data of multiple security events corresponding to the same event category, the behavioral statistics of the event category are determined, including: Obtain the action type corresponding to user behavior data; wherein, the action type includes at least one of watching live broadcast, watching video playback, downloading video, and sharing events; Based on the preset correspondence between action types and weights, the corresponding action weights are determined; among them, the action weight for downloading a recording is higher than that for sharing an event, the action weight for sharing an event is higher than that for viewing a recording playback, and the action weight for viewing a recording playback is higher than that for viewing a live stream. Based on the action type and behavior weight corresponding to each security event, the user behavior data of multiple security events under the same event category are weighted and statistically analyzed to obtain the behavior statistics value of the event category; among which, the behavior statistics value includes the playback frequency and / or the return visit rate.

[0009] On the one hand, when the behavioral statistics meet preset conditions, the risk assessment strategy corresponding to the event category is adjusted, including: Determine whether multiple security events in an event category that meet preset conditions comply with global adjustment requirements; wherein, global adjustment requirements include multiple security events in an event category that meet preset conditions being distributed across multiple devices of multiple users; If the event category meets the global adjustment requirements, increase the scoring parameters in the risk assessment strategy corresponding to the event category, and apply the increased scoring parameters to the risk assessment of the event category for all devices of the same model. If the event category does not meet the global adjustment requirements, a preset weighting coefficient is set for the scoring parameters corresponding to the event category, and the weighting coefficient is applied to the risk assessment of the event category by a single target device; whereby the single target device is the single device corresponding to the security event that meets the preset conditions in the event category.

[0010] On the one hand, based on the risk assessment strategy, the raw characteristic data of security incidents are analyzed to determine multiple feature labels corresponding to each risk assessment dimension, including: Extract feature labels corresponding to multiple risk assessment dimensions from the original feature data; among them, the risk assessment dimensions include any combination of basic dimension, spatial dimension, duration dimension, facial identity dimension and scenario dimension; Based on the preset score mapping rules, the risk score corresponding to each feature label is determined.

[0011] On the one hand, feature labels corresponding to multiple risk assessment dimensions are extracted from the original feature data, including: If a preset identity category is detected in a security incident, the preset identity category will be used as a feature label under the facial identity dimension. When multiple identity categories are detected in a security incident, the identity category with the highest priority is used as the feature label under the face identity dimension, according to the priority order of identity categories. If multiple non-whitelisted identity categories are detected in a security incident, and the occupational confidence scores of all non-whitelisted identity categories meet the confidence threshold, the identity category with the lowest risk score is selected as the feature label under the facial identity dimension from among the multiple non-whitelisted identity categories.

[0012] On the one hand, the risk scores of each target feature label are fused to obtain the total risk score of the security incident, including: The risk scores of each target feature label are merged using a weighted maximum value method to obtain the total risk score of the security incident; Alternatively, the risk score of each target feature label can be compared with a preset upper limit threshold; if there is a risk score that exceeds the upper limit threshold, the highest risk score that exceeds the upper limit threshold can be determined as the total risk score. Alternatively, the risk scores of each target feature label can be processed according to the set logical combination rules to determine the total risk score of the security incident.

[0013] On the one hand, the risk scores of each target feature label are processed according to the set logical combination rules to determine the total risk score of the security incident, including: If the risk score of the target feature label meets the preset blocking conditions, the preset lowest value will be used as the total risk score. When the risk scores of multiple target feature labels simultaneously meet the preset multidimensional parallel conditions, the upper risk limit is used as the total risk score; where the upper risk limit is a value higher than the risk scores of each target feature label.

[0014] On the one hand, it also includes: Determine whether a security event with a risk level higher than the silent handling level occurs within the time period corresponding to the preset duration; In the absence of security incidents with a risk level higher than the silent handling level, extract structured data of multiple security incidents with a risk level of silent handling within a time period. Generate summary information for structured data and push the summary information through a low-priority channel; the summary information is used to describe the event situation within a time period.

[0015] This application also provides an electronic device, including at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, and when the computer program is executed by the at least one processor, it causes the at least one processor to implement: Based on the risk assessment strategy, the original characteristic data of security incidents are analyzed to determine multiple feature labels corresponding to each risk assessment dimension; among them, each feature label has its corresponding risk score. The target feature label with the highest risk score is selected from multiple feature labels corresponding to each risk assessment dimension; each risk assessment dimension has one corresponding target feature label. The risk scores of each target feature label are fused to obtain the total risk score of the security event, and the risk level of the security event is determined based on the total risk score; wherein, the risk level includes at least one silent handling level; When the risk level of a security incident is classified as a silent handling level, user behavior data of the security incident is collected within a preset time window. Based on user behavior data under the silent processing level and the classification feature information corresponding to security events, the risk assessment strategy corresponding to security events is adjusted, and the risk level of new security events with the same classification feature information is determined according to the adjusted risk assessment strategy.

[0016] As can be seen from the above technical solution, the original characteristic data of security events are analyzed according to the risk assessment strategy to determine multiple feature labels corresponding to each risk assessment dimension; each feature label has its corresponding risk score. The target feature label with the highest risk score is selected from the multiple feature labels corresponding to each risk assessment dimension, effectively avoiding the linear accumulation and artificially inflated risk scores caused by multiple synonymous or overlapping labels triggering simultaneously within the same dimension. This limits the contribution of each dimension to the total risk score to a reasonable upper limit, thereby improving the accuracy of subsequent total risk score calculation and the stability of level determination. Based on this, the risk scores of each target feature label are fused to obtain the total risk score of the security event. This fusion process makes the total risk score more closely reflect the actual threat level of the event, reducing the probability of false alarms and missed alarms. The risk level of the security event is determined based on the total risk score; the risk level includes at least one silent handling level. When the risk level of the security event is at the silent handling level, user behavior data of the security event is collected within a preset time window. Based on user behavior data under the silent processing level and the classification characteristics of security events, the risk assessment strategy for security events is adjusted. This allows new security events with the same classification characteristics to have their risk levels redefined according to the adjusted risk assessment strategy. This reduces the lag in responding to changes in user needs due to long-term reliance on fixed rules, and avoids over-responding to individual abnormal scenarios with global adjustments. While ensuring the accuracy of risk assessment, it also improves the system's adaptability to real user needs. Attached Figure Description

[0017] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a security event notification method provided in this application embodiment; Figure 2 A flowchart illustrating a method for adjusting a risk assessment strategy, as provided in this application embodiment; Figure 3 A flowchart illustrating a method for extracting feature labels from a facial identity dimension, provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a security event notification system provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0020] The terms "comprising" and "having," and any variations thereof, in the specification and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may include steps or units not listed.

[0021] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1: Figure 1 A flowchart of a security event notification method provided in an embodiment of this application is shown. The method includes the following steps.

[0023] S101: Analyze the original characteristic data of security incidents based on the risk assessment strategy to determine multiple feature labels corresponding to each risk assessment dimension.

[0024] Raw characteristic data of security incidents refers to the raw information collected by security monitoring equipment when an incident is detected.

[0025] Security monitoring equipment can include cameras, doorbells, sensors, etc.

[0026] Raw feature data may include video frame images, target detection results, motion trajectories, timestamps, device locations, etc.

[0027] A risk assessment strategy is a pre-defined set of rules used to assess the degree of risk of an event. The risk assessment strategy defines which dimensions of information to extract from the raw feature data, how to map this information to feature labels, and the risk score corresponding to each feature label.

[0028] In this step, feature labels corresponding to multiple risk assessment dimensions are extracted from the original feature data. Risk assessment dimensions are different perspectives on the sources of security incident risks, and can include any combination of basic dimensions, spatial dimensions, duration dimensions, facial identity dimensions, and scene dimensions.

[0029] Basic dimensions reflect the fundamental attributes of an event, such as whether the target that triggers the event is a moving target, and the type of target, such as a person, animal, or vehicle.

[0030] The spatial dimension involves the target's location information in the monitoring screen, such as whether the target has entered a preset restricted or sensitive area.

[0031] The duration dimension reflects the duration characteristics of an event, such as how long a target stays in the monitored area; a longer stay may mean a higher risk.

[0032] The facial identity dimension is the result of identifying the identity of the person appearing in the event, such as identifying them as a family member, an acquaintance, a stranger, or a person with a specific profession (such as a courier or a food delivery worker).

[0033] The scenario dimension describes the specific behavioral semantics corresponding to the event, such as climbing over a wall, stealing a package, ringing a doorbell, or loitering at the door.

[0034] The specific method for extracting feature labels from raw feature data depends on the capabilities of the security equipment. For security equipment with intelligent analysis capabilities, target detection, face recognition, and behavior recognition algorithms can be run directly on the device to output structured feature labels. For security equipment with limited computing power, raw video frames or keyframes can be uploaded to a cloud server, where a recognition model runs and the recognition results are returned as feature labels. This application does not limit the specific execution location of feature label extraction.

[0035] After determining the feature labels for each risk assessment dimension, the risk score corresponding to each feature label is determined according to the preset score mapping rules.

[0036] The score mapping rule is a pre-configured lookup table or mapping function that maps each specific feature label to a numerical risk score. For example, the "family" label is mapped to the lowest risk score, the "stranger" label to a higher risk score, and the "climbing over the wall" label to a high risk score. The risk scores corresponding to different feature labels reflect the degree to which that label contributes to the overall risk in the corresponding dimension.

[0037] Each risk assessment dimension has its own corresponding risk score for different feature labels. Risk scores can be expressed on a 100-point scale. For example, in the basic dimension, changes in lighting and shadows (5 points), swaying leaves (10 points), and detecting a moving object (15 points). In the spatial dimension, distant public streets (10 points), regular activity areas in front of doors (25 points), and designated key warning / restricted areas (85 points). In the duration dimension, fleeting appearances (<5 seconds) (5 points), regular passage (5-30 seconds) (15 points), and prolonged abnormal loitering (>2 minutes) (55 points). In the facial identity dimension, family members (0 points, with veto power over noise reduction), acquaintances / frequent customers (20 points), whitelisted professionals such as deliverymen (30 points), and strangers (65 points). In the scene dimension, pets' regular activities (10 points), regular package delivery (25 points), and violent damage / theft of packages (95 points).

[0038] S102: Select the target feature label with the highest risk score from multiple feature labels corresponding to each risk assessment dimension.

[0039] Multiple feature labels may be extracted under each risk assessment dimension. For example, the scenario dimension may simultaneously detect two labels: regular package delivery and violent vandalism; the facial recognition dimension may simultaneously detect two occupational labels: courier and food delivery worker. Directly summing the scores of multiple labels within the same dimension would lead to an overestimation of the risk contribution of that dimension, because multiple labels may simply be different perspectives on the same risk fact, rather than multiple independent sources of risk.

[0040] To effectively address the issue of overemphasis on risk contribution, this step selects the feature label with the highest risk score from multiple feature labels corresponding to each risk assessment dimension, and uses it as the target feature label for that dimension.

[0041] Each risk assessment dimension ultimately retains only one target feature label for subsequent total risk score calculation. This approach ensures that the contribution of each dimension to the total risk score is limited to the score level of a single label, eliminating the possibility of inflated scores caused by the superposition of multiple labels within the same dimension.

[0042] Taking the scenario dimension as an example, if a security incident triggers both the regular package delivery (25 points) and vandalism (95 points) tags simultaneously, the system only takes vandalism as the target feature tag for that dimension, while the regular package delivery is discarded and not included in subsequent calculations. Similarly, if the facial identity dimension simultaneously identifies a courier (30 points) and a stranger (65 points), the system takes the stranger with the higher risk score as the target feature tag for that dimension.

[0043] S103: The risk scores of each target feature label are fused to obtain the total risk score of the security event, and the risk level of the security event is determined based on the total risk score.

[0044] Each dimension provides a target feature label and its corresponding risk score. This step merges these risk scores from different dimensions to obtain the total risk score for the security incident.

[0045] The fusion process is not a simple linear summation, but a non-linear fusion method is used to avoid the total score getting out of control due to the simple accumulation of scores from different dimensions.

[0046] There are several nonlinear fusion methods. The first method is the weighted maximum value method. By fusing the risk scores of each target feature label using the weighted maximum value method, the total risk score of the security event can be obtained. In specific implementations, a dimension weight is assigned to each risk assessment dimension. The risk score of each target feature label is multiplied by its corresponding dimension weight, and the maximum value is taken as the total risk score. This method is suitable for scenarios where the importance of different dimensions varies, adjusting the upper limit of each dimension's contribution through weights while retaining the principle of extreme value dominance.

[0047] The second fusion method is the threshold judgment method. The risk score of each target feature label is compared with a preset upper limit threshold. If there is a risk score that exceeds the upper limit threshold, the highest risk score that exceeds the upper limit threshold is determined as the total risk score.

[0048] The second method is suitable for scenarios that require rapid identification of extreme risks in a single dimension. As long as the risk score of any one dimension exceeds the preset upper limit threshold, it is sufficient to determine the danger level of the entire event, without having to consider the scores of other dimensions.

[0049] For example, the system presets an absolute upper limit threshold of 80 points. After extracting the extreme values ​​of the five dimensions, the system first determines whether any extreme value exceeds the upper limit threshold. Assuming the extreme value of the scenario dimension is "theft" at 95 points, the threshold penetration principle is triggered, directly anchoring the final total risk score to this highest extreme value of 95 points, ignoring the downward influence of other lower-scoring dimensions, ensuring that high-risk behaviors are never missed. If none of the extreme values ​​exceed the upper limit threshold, then maximum smoothing or weighted averaging can be used to determine the total risk score.

[0050] The third fusion method is the logical combination rule method. The risk scores of each target feature label are processed according to the set logical combination rules to determine the total risk score of the security event.

[0051] Logical combination rules can include two sub-cases.

[0052] Sub-scenario 1: If the risk score of the target feature label meets the preset blocking conditions, the preset minimum value will be used as the total risk score.

[0053] Blocking conditions refer to specific feature labels that inherently possess a veto power; that is, the presence of such a label should lower the overall risk score of an event. For example, when the facial recognition dimension identifies a "family member" label, regardless of any high-risk labels triggered by other dimensions (such as the scene dimension), the overall risk score of the event is set to the lowest value. This is because family members have the highest credibility in security scenarios, and their behavior should not trigger alarms. This design avoids misjudging normal family activities (such as a family member climbing over a low wall in their own yard) as high-risk events.

[0054] Sub-scenario 2: When the risk scores of multiple target feature labels simultaneously meet the preset multidimensional parallel conditions, the upper limit of risk will be used as the total risk score.

[0055] The risk ceiling can be a value higher than the risk score of each target feature label.

[0056] The multidimensional parallel condition refers to the situation where multiple dimensions simultaneously reach a high level of risk. Even if a single dimension does not reach an extreme threshold, the risk level of the event should be adjusted upwards when multiple dimensions are at a high level at the same time.

[0057] For example, IF (identity dimension extreme value == stranger [65 points]) AND (spatial dimension extreme value == key alert area [85 points]) AND (duration dimension extreme value > 2 minutes [55 points]). When the extreme values ​​of these three dimensions simultaneously meet the above logical conditions, the individual scores of each dimension may not exceed their respective upper limit thresholds. However, the fact that all three dimensions are at a high level simultaneously means that the overall risk of the event is significantly higher than the risk of any single dimension. At this time, the logic combination engine will determine that a compound risk has been generated, and thus output a high risk upper limit value, such as assigning 90 points. By raising the total risk score to an upper limit value higher than the scores of each single dimension, the actual threat level of the superposition of multi-dimensional risks is reflected.

[0058] The risk level of a security incident is determined based on its total risk score. The risk level includes at least one silent handling level. In this application, the risk levels can be divided into three levels: L1, L2, and L3. L1 is the highest risk level (extremely dangerous), L2 is the medium risk level (important), and L3 is the lowest risk level (low risk). L3 is the silent handling level; events classified as L3 will not be actively pushed to the user by the system by default, and will only be handled silently. Using a percentage system as an example, L3 can be set to < 40 points, L2 to 40-79 points, and L1 to >= 80 points.

[0059] S104: When the risk level of a security incident is at the silent handling level, collect user behavior data of the security incident within a preset time window.

[0060] Once a security event is classified as an L3 silent event, the system will not immediately discard the event. Instead, it will continuously monitor the user's subsequent behavior related to the event within a preset time window. The time window can be set to within 15 minutes of the event's occurrence. This duration is sufficient to cover the typical user behavior cycle of viewing event-related recordings, while avoiding interference from irrelevant behaviors introduced by an excessively long acquisition window.

[0061] User behavior data includes whether a user actively opens the application to view the live stream or replay of the event. The collected user behavior data may include: a unique event identifier (Event ID), an event generation timestamp, a user identifier (User ID), a user streaming timestamp, and the type of action the user performed (viewing live stream, viewing replay, downloading recording, sharing event, etc.).

[0062] Among these, the behavior of users actively opening the app and viewing the corresponding live stream or replay content after a silent event occurs is defined as an "implicit positive feedback signal." This signal is called implicit because it doesn't require users to perform additional actions such as liking, rating, or complaining; the user's viewing behavior is a natural and unconscious way of expressing concern, without placing any extra operational burden on the user. It is positive feedback because it indicates that the system's initial risk assessment of the event may not be consistent with the user's actual level of concern; the user actually cares about this type of event and therefore actively spends time viewing it.

[0063] S105: Based on user behavior data under the silent processing level and the classification feature information corresponding to security events, adjust the risk assessment strategy corresponding to the security events, and determine the risk level of new security events with the same classification feature information based on the adjusted risk assessment strategy.

[0064] After collecting sufficient user behavior data, the system dynamically adjusts its risk assessment strategy based on this data. The adjustments are based on two aspects: first, the user behavior data itself, which reflects the user's true level of concern about silent events; and second, the classification characteristics of the security events, which are used to apply the adjustments to specific categories of events rather than generalizing them globally, so as to avoid the adjustment strategy being too coarse and affecting the accuracy of the judgment of other event types.

[0065] For example, if a certain type of L3 event, such as a pet lingering at the door, shows frequent user activity within 15 minutes after being initially ignored, it indicates that users actually have a high level of interest in this type of event, suggesting that the system's previous risk assessment for this type of event was too low. Based on this, the system adjusts the scoring parameters for this type of event in its risk assessment strategy. Subsequently, when a new security event with the same classification characteristics occurs again—that is, another pet lingering at the door type—the system recalculates its total risk score according to the adjusted risk assessment strategy. The event's score will be higher than before, potentially jumping from L3 to L2, thus qualifying it for proactive push notifications.

[0066] Adjusting the risk assessment strategy essentially involves building a self-adjusting mechanism based on implicit user behavioral feedback. The system doesn't need to wait for users to submit feedback or complaints, nor does it require manual intervention to adjust rule parameters. Instead, it automatically identifies silently processed but potentially valuable event categories by capturing behavioral signals naturally generated by users in their daily use, and drives the scoring strategy towards directions that better reflect users' true concerns.

[0067] By collecting implicit behavioral data on silent events and adjusting risk assessment strategies accordingly, the system can automatically identify event categories that are misjudged as low priority but actually have high attention value. This enables dynamic optimization of the scoring strategy, reduces reliance on fixed thresholds and manual adjustments, and improves the system's adaptability to users' real needs.

[0068] As can be seen from the above technical solution, the original characteristic data of security events are analyzed according to the risk assessment strategy to determine multiple feature labels corresponding to each risk assessment dimension; each feature label has its corresponding risk score. The target feature label with the highest risk score is selected from the multiple feature labels corresponding to each risk assessment dimension, effectively avoiding the linear accumulation and artificially inflated risk scores caused by multiple synonymous or overlapping labels triggering simultaneously within the same dimension. This limits the contribution of each dimension to the total risk score to a reasonable upper limit, thereby improving the accuracy of subsequent total risk score calculation and the stability of level determination. Based on this, the risk scores of each target feature label are fused to obtain the total risk score of the security event. This fusion process makes the total risk score more closely reflect the actual threat level of the event, reducing the probability of false alarms and missed alarms. The risk level of the security event is determined based on the total risk score; the risk level includes at least one silent handling level. When the risk level of the security event is at the silent handling level, user behavior data of the security event is collected within a preset time window. Based on user behavior data under the silent processing level and the classification characteristics of security events, the risk assessment strategy for security events is adjusted. This allows new security events with the same classification characteristics to have their risk levels redefined according to the adjusted risk assessment strategy. This reduces the lag in responding to changes in user needs due to long-term reliance on fixed rules, and avoids over-responding to individual abnormal scenarios with global adjustments. While ensuring the accuracy of risk assessment, it also improves the system's adaptability to real user needs.

[0069] Example 2: Figure 2 The flowchart of a method for adjusting a risk assessment strategy provided in this application embodiment includes the following operation steps.

[0070] S201: Based on the semantic tags, spatiotemporal features, and device feature information corresponding to multiple silent events, cluster the multiple silent events to determine at least one event category.

[0071] Silent events refer to security events with a risk level of L3 (silent handling level). User behavior data from a single silent event may lack statistical significance due to insufficient sample size; therefore, multiple silent events need to be grouped into event categories according to certain rules.

[0072] The features used for clustering can include semantic labels, spatiotemporal features, and device feature information.

[0073] Semantic tags refer to the semantic descriptive information carried by an event, such as a pet lingering at the door, a deliveryman making a brief stop, leaves rustling, or changes in light and shadow. Semantic tags are typically output by scene-dimensional recognition models and are natural language-level descriptions of the event content.

[0074] Spatiotemporal characteristics can include the time period in which the event occurs (such as morning rush hour, evening rush hour, late at night) and geographical location (such as the entrance of an apartment building, the gate of a villa, or the entrance to a garage).

[0075] Device characteristic information may include the device model that triggered the event (such as doorbell, outdoor camera, etc.) and device capabilities (such as whether it has facial recognition function, whether it has night vision function, etc.).

[0076] Clustering can be done using rule-based grouping or density-based clustering algorithms.

[0077] Rule-based grouping predefines a semantic tag classification system, grouping events with similar semantics into the same category, such as classifying "pets passing by" and "pets loitering" into the "pet activity" category.

[0078] Density-based clustering automatically identifies densely distributed event groups in the feature space as event categories by calculating the similarity between event feature vectors. This application does not limit the specific choice of clustering algorithm, as long as it can aggregate a large number of silent events into statistically significant categories based on feature similarity.

[0079] Through clustering, previously isolated silent events are organized into statistically significant event categories. Each event category represents a set of events with similar semantics, spatiotemporal distribution, and device context, providing a reasonable aggregation granularity for subsequent behavioral statistics.

[0080] S202: Determine the behavioral statistics value of the event category based on the user behavior data of multiple security events corresponding to the same event category.

[0081] After determining the event category, this step aggregates user behavior data for all security events within the same event category and calculates the behavioral statistics for that event category.

[0082] First, obtain the action types corresponding to user behavior data.

[0083] Action type refers to the specific operation performed by the user when viewing an event, which may include at least one of the following: viewing live stream, viewing video playback, downloading video, and sharing the event.

[0084] Different action types reflect varying degrees of user attention. Viewing a live stream refers to a user checking the monitoring footage in real-time after an event occurs; this typically happens when the user happens to open the application, and the intensity of their attention signal is relatively low. Viewing video playback refers to a user actively replaying historical recordings, indicating a strong desire to view the event, as playback requires the user to actively access the history page to search and play the recording after the event. Downloading recordings refers to a user saving the recording file locally, indicating a high level of attention and a willingness to retain the event. Sharing an event refers to a user sharing event information with other users (such as family members), indicating that the user believes the event has informational value or needs to be known by others.

[0085] Then, based on the preset correspondence between action types and weights, the behavior weights corresponding to each action type are determined.

[0086] The weighting rules for different action types can include: the weight of the action corresponding to downloading a recording is higher than the weight of the action corresponding to sharing an event; the weight of the action corresponding to sharing an event is higher than the weight of the action corresponding to viewing a recording playback; and the weight of the action corresponding to viewing a recording playback is higher than the weight of the action corresponding to viewing a live stream.

[0087] The weighting is based on the user's level of engagement and attention reflected in each action type. Downloading a recording requires the most user effort (clicking multiple menus, selecting storage location, etc.) and indicates that the user wants to save the recording long-term, therefore it has the highest weight. Sharing an event involves the user transmitting event information to others, indicating that the user believes the event has social sharing value, so it has the next highest weight. Viewing a recording requires the user to actively access the history interface to search, which requires more effort than real-time viewing, so it has the next highest weight. Viewing a live stream is likely a passive action that the user happens to open the application, with the lowest effort and the lowest weight.

[0088] Finally, based on the action type and behavior weight corresponding to each security event, the user behavior data of multiple security events under the same event category are weighted and statistically analyzed to obtain the behavior statistics value of the event category.

[0089] Behavioral statistics may include replay frequency and / or return visit rate.

[0090] Replay frequency refers to the total number of times a certain type of event is replayed by users within a unit of time (e.g., daily). Replay rate refers to the proportion of times a certain type of event is replayed out of the total number of times that type of event occurs. Replay rate can eliminate the influence of the frequency of event occurrence itself and more accurately reflect the proportion of user attention to this type of event.

[0091] For example, if a type A event occurs 100 times a day and is replayed 20 times, the replay rate is 20%; a type B event occurs 10 times a day and is replayed 5 times, the replay rate is 50%. Although type A events are replayed more frequently, type B events have a higher replay rate, indicating that users pay more attention to type B events, and type B events are more likely to be subject to systemic underestimation of risk.

[0092] S203: If the behavioral statistics value meets the preset conditions, adjust the risk assessment strategy corresponding to the event category.

[0093] The preset condition is a pre-set threshold used to determine whether a certain event category should trigger a policy adjustment.

[0094] In this embodiment, the preset condition can be set as a replay rate exceeding 25% for three consecutive days. That is, when the replay rate of a certain event category remains above 25% for three consecutive days, the system determines that there is a systemic risk of false negatives in that category, meaning that the actual attention given to this type of event is significantly higher than the risk level assigned by the system.

[0095] When the behavior statistics value meets the preset conditions, the system can further determine whether the adjustment range of the event category is global or local.

[0096] In the specific implementation, it is determined whether multiple security events in the event category that meet the preset conditions comply with the global adjustment requirements.

[0097] Global adjustment requirements can include multiple security events within an event category that meet preset conditions, distributed across multiple devices belonging to multiple users. When multiple security events are distributed across multiple devices belonging to multiple users, it indicates that the high playback rate of this type of event is not solely due to the specific preferences of a single user or the unique installation location of a single device, but rather a common phenomenon widely observed across different users and devices. Only when this condition is met can it be concluded that the underestimation of risk for this type of event is a widespread problem, worthy of adjustment on a global scale.

[0098] If the event category meets the global adjustment requirements, increase the scoring parameters in the risk assessment strategy corresponding to that event category, and apply the increased scoring parameters to the risk assessment of all devices of the same model for that event category.

[0099] A global adjustment means modifying the baseline scoring parameters of the strategy to apply to all users and all devices of the same model. For example, the base score for a pet lingering at the door might be raised from the original L3 threshold (e.g., 40 points) to the L2 threshold (e.g., 68 points). After the adjustment, any user's device of the same model that detects a pet lingering at the door will calculate its total risk score according to the increased scoring parameters. This event will then have the opportunity to jump from L3 to L2 level and receive a push notification.

[0100] If the event category does not meet the global adjustment requirements, it indicates that the high playback rate is limited to a single device or a single user. Set a preset weighting coefficient for the scoring parameters corresponding to the event category, and apply the weighting coefficient to the risk assessment of the event category for a single target device.

[0101] A single target device is a single device corresponding to a security event that meets preset conditions within an event category.

[0102] Local adjustments do not change the global baseline scoring parameters. Instead, they add a device-specific weighting coefficient to the global score, causing that device to receive a higher risk score when judging such events, while other devices remain unaffected. Local adjustments are applicable to personalized attention differences caused by specific device installation locations (e.g., a user's pet particularly likes to play in front of a camera). These differences are not universal and therefore only take effect on that specific device.

[0103] The method for adjusting risk assessment strategies proposed in this embodiment effectively solves the problem of sparse and unreliable judgment of single-event behavioral data by clustering silent events into statistically significant event categories and statistically analyzing user behavioral feedback. By assigning differentiated behavioral weights to different action types, the behavioral statistics can more accurately reflect the user's true level of attention. By distinguishing between global and local adjustment modes, strategy optimization can respond to common problems without affecting global stability due to individual abnormal data, achieving precise and controllable strategy iteration.

[0104] Example 3: In this application, feature labels corresponding to multiple risk assessment dimensions are extracted from the original feature data. For the facial identity dimension, the feature label extraction can adopt a three-level adjudication mechanism. Figure 3 A flowchart illustrating a method for extracting feature labels from a facial identity dimension, provided in this application embodiment, is included. S301: When a preset identity category is detected in a security event, the preset identity category is used as the feature label under the face identity dimension.

[0105] The preset identity category can be a user's pre-marked identity as a family member in a whitelist. Family members typically have the highest administrative privileges on security devices and enjoy the highest alarm exemption rights in system judgment. When a family member is detected in an event, regardless of the presence of other individuals, the system will identify the facial recognition feature as a family member, thus avoiding false alarms about family members' daily activities.

[0106] S302: When multiple identity categories are detected in a security event, the identity category with the highest priority shall be used as the feature label under the face identity dimension according to the priority order of identity categories.

[0107] The priority order for identity categories is: family members over acquaintances, and acquaintances over strangers. Acquaintances refer to known individuals that users actively enter and label as acquaintances in the system, including but not limited to regular customers, relatives, friends, and contracted domestic service personnel, but acquaintances do not have device control or viewing permissions. When family members are missing, but acquaintances and strangers coexist, the system selects acquaintances as the feature label for the facial identity dimension, assigning them a risk score between that of family members and strangers.

[0108] S303: If multiple non-whitelisted identity categories are detected in a security incident, and the professional confidence scores of multiple non-whitelisted identity categories all meet the confidence threshold, the identity category with the lowest risk score shall be selected as the feature label under the face identity dimension from among the multiple non-whitelisted identity categories.

[0109] Non-whitelisted identity categories refer to individuals who have not been added to the user's whitelist. The occupational recognition model determines the occupational category of detected individuals, such as delivery drivers or food delivery workers. The occupational recognition model employs a multimodal feature fusion architecture. Input features include: the color histogram and local semantic segmentation features of the target's outer clothing (for identifying uniforms of specific occupations), OCR text recognition and logo topology matching of the target's chest or back area (for identifying company logos), detection of specific carried objects (such as cube-shaped cardboard boxes or insulated food delivery boxes), and macroscopic temporal behavioral features (such as slowing down and stopping while riding a two-wheeled electric vehicle, walking to the door while holding an object of a specific size, etc.).

[0110] When the occupational confidence scores of multiple non-whitelisted individuals all exceed the set confidence threshold, the system selects the occupation with the lowest risk score as the feature label for the facial identity dimension of that event. The occupation with the lowest risk score is chosen because it acts as a noise reducer; individuals identified as belonging to a specific occupation usually have a legitimate reason for being in the monitored area, indicating lower risk. Therefore, using the lowest score maximizes the utilization of occupational information to reduce the false alarm rate.

[0111] In this application, when a pre-defined identity category (such as family member) exists in a security incident, that identity is directly used as the feature label for that dimension. Leveraging the highest credibility and alarm exemption rights of the family member identity in security scenarios, normal family activities are not misjudged as high-risk events due to other concurrent labels, effectively avoiding false alarms caused by identity recognition interference. When multiple identity categories exist in an incident, the highest priority identity is selected as the label according to identity priority order. This ensures that in complex mixed-person scenarios, the system prioritizes the most credible identity, closest to the user's pre-defined perception, to guide the risk scoring for that dimension, preventing low-priority identities from inflating the overall risk score and thus maintaining the rationality of the scoring. When multiple non-whitelisted identities exist in an event, and the confidence levels of each profession meet the confidence threshold, the one with the lowest risk score is selected as the label. This fully utilizes the reasonable discrepancies contained in the profession identification results. Selecting the lowest score when multiple professions coexist maximizes the noise reduction effect of profession features. Simultaneously, the confidence threshold condition ensures that this noise reduction operation only takes effect when the identification results are reliable, avoiding interference from low-confidence profession labels on the scoring. This application employs a three-layer progressive adjudication logic. The facial identity dimension can adaptively select the most suitable and reliable identity label to participate in the subsequent total risk score calculation based on the actual identity composition in the event. This ensures the stability of the identity dimension output and effectively reduces scoring deviations caused by redundant or conflicting identity labels, thereby improving the overall accuracy of security event risk scoring and user trust in the alarm system.

[0112] Example 4: This embodiment provides a solution for proactively conveying system operation information to users during periods when no high-risk events occur, further improving the user experience during quiet periods.

[0113] First, determine whether a security event with a risk level higher than the silent handling level has occurred within the time period corresponding to the preset duration.

[0114] In this embodiment, the preset duration can be 3 hours, meaning the system performs a check every 3 hours.

[0115] The detection window is limited to 9:00 AM to 9:00 PM daily. This time window is chosen based on users' daily routines. After 9:00 AM, users have usually started their daily activities, and before 9:00 PM, they have not yet entered their nighttime rest state. Pushing information within this time window will not disturb users' rest.

[0116] The time window supports time zone adaptation. The system dynamically anchors to the local time from 9:00 to 21:00 by obtaining the system time zone of the user's mobile terminal or the geographical location time zone reported by the device during network configuration, so as to adapt to the needs of users in different regions.

[0117] Users can also customize the start and end times of this time window in the application's settings to meet the needs of people with special work schedules, such as night shift workers.

[0118] If no security incidents of a higher risk level than the silent handling level have occurred, that is, if no L1 or L2 events have occurred in the past 3 hours, it indicates that the period is a silent period. The system determines that it is appropriate to convey information about the normal operation of the system to the user to alleviate the user's anxiety caused by the lack of notification for a long time.

[0119] Therefore, in the absence of security incidents with a risk level higher than the silent handling level, structured data of multiple security incidents with a risk level of silent handling are extracted within this time period.

[0120] Structured data refers to standardized records of L3 events that are processed silently. It can include structured fields that can be directly processed by computer programs, such as event occurrence time, event type (semantic label), detected target type, and target quantity.

[0121] After extracting structured data, a summary of the structured data is generated and pushed through a low-priority channel. The summary describes the events that occurred within that time period.

[0122] The system performs semantic value-based differential filtering on the L3 event stream. The purpose of differential filtering is to select representative events as summary material while filtering out purely environmental noise.

[0123] In the specific implementation, L3 events containing specific entity targets (such as people passing by from a distance, acquaintances staying for a very short time, and pets' routine activities) are given high weight and are used as priority materials for extraction; a filtering threshold is set for purely environmental physical disturbances (such as drastic changes in light and shadow, swaying leaves, flying insects, and other pixel-level noise). If the number of such noise events is too high, the system will collapse or discard them and not include them as the focus of the summary.

[0124] The system inputs the L3 event structured data, which has been filtered by difference, into the natural language generation model, which outputs a natural language description as a summary.

[0125] The generation of summary information follows the principle of positive expression, using positive or neutral terms such as "the system is running normally and X instances of human and pet activity were detected" to describe the event. Words such as "nothing happened" or "no abnormalities" that may cause users to have negative associations are prohibited, such as leading users to think that the system is broken and therefore did not report anything.

[0126] The summary information can also include prompts such as "details can be viewed again" to guide users to view the detailed records themselves.

[0127] Summary information is pushed via a low-priority channel, which is distinct from the high-priority or regular-priority channels used for L1 and L2 events. The low-priority channel pushes information silently via the system, in conjunction with in-app timeline cards. When pushed on mobile devices, the text card is only displayed in the notification center or lock screen, without triggering sound, vibration, or screen activation.

[0128] If the user has enabled the operating system's Do Not Disturb mode, the message will be silently stored in the notification center according to the system-level interception protocol, without forcibly interrupting the user's focus or sleep. After clicking the summary card, the user will be directly redirected to the corresponding L3 event slice video aggregation page via a link, allowing the user to view details. Additionally, the system provides a global switch in the settings page, such as naming it System Work Daily / Summary, which users can disable at any time, giving them complete control.

[0129] The system synchronously records complete log information, which may include timestamp slice logs (recording the time period and triggering conditions for each summary generation), the template ID used for summary generation, triggering condition logs (recording evidence that there were no L1 / L2 events within that time period), and subsequent user clicks or ignore actions. This log data can be used for subsequent performance evaluation and strategy optimization.

[0130] This embodiment proactively generates and pushes natural language summaries based on L3 event streams during periods without high-risk events. Without adding unnecessary notifications, it signals to users that the system is operating normally, effectively alleviating user anxiety and misjudgments of device malfunction caused by prolonged lack of notifications. The summary content is dynamically generated based on actual L3 events, effectively avoiding the mechanical feel of fixed templates. Furthermore, the proactive approach and clear guidance reinforce users' positive perception of system reliability. The design, compatible with low-priority channels and Do Not Disturb mode, ensures the non-intrusiveness of this mechanism, without affecting users' focus or rest during the current period. A complete logging system provides data support for effectiveness evaluation and subsequent optimization.

[0131] Example 5: Figure 4 A schematic diagram of a security event notification system provided in this application embodiment includes an analysis unit 41, a filtering unit 42, a fusion unit 43, a data acquisition unit 44, and an adjustment unit 45; Analysis unit 41 is used to analyze the original characteristic data of security events according to the risk assessment strategy and determine multiple feature labels corresponding to each risk assessment dimension; among them, each feature label has its corresponding risk score. The filtering unit 42 is used to filter out the target feature label with the highest risk score from multiple feature labels corresponding to each risk assessment dimension; wherein, each risk assessment dimension has a corresponding target feature label; The fusion unit 43 is used to fuse the risk scores of each target feature label to obtain the total risk score of the security event, and to determine the risk level of the security event based on the total risk score; wherein, the risk level includes at least one silent processing level; The data collection unit 44 is used to collect user behavior data of security events within a preset time window when the risk level of the security event is the silent handling level. The adjustment unit 45 is used to adjust the risk assessment strategy corresponding to the security event based on the user behavior data under the silent processing level and the classification feature information corresponding to the security event, and to redetermine the risk level of the security event based on the adjusted risk assessment strategy.

[0132] In some embodiments, the adjustment unit includes a clustering subunit, a determination subunit, and an adjustment subunit; The clustering subunit is used to cluster multiple silent events according to their semantic tags, spatiotemporal features, and device feature information to determine at least one event category; wherein, the silent event is a security event with a risk level of silent handling level. The sub-unit is used to determine the behavioral statistics of the event category based on the user behavior data of multiple security events corresponding to the same event category. The adjustment sub-unit is used to adjust the risk assessment strategy corresponding to the event category when the behavioral statistics value meets the preset conditions.

[0133] In some embodiments, a subunit is determined for acquiring the action type corresponding to the user behavior data; wherein the action type includes at least one of watching live stream, watching video playback, downloading video, and sharing events; Based on the preset correspondence between action types and weights, the corresponding action weights are determined; among them, the action weight for downloading a recording is higher than that for sharing an event, the action weight for sharing an event is higher than that for viewing a recording playback, and the action weight for viewing a recording playback is higher than that for viewing a live stream. Based on the action type and behavior weight corresponding to each security event, the user behavior data of multiple security events under the same event category are weighted and statistically analyzed to obtain the behavior statistics value of the event category; among which, the behavior statistics value includes the playback frequency and / or the return visit rate.

[0134] In some embodiments, the adjustment subunit is used to determine whether multiple security events in an event category that meet preset conditions meet global adjustment requirements; wherein, the global adjustment requirements include multiple security events in an event category that meet preset conditions being distributed across multiple devices of multiple users; If the event category meets the global adjustment requirements, increase the scoring parameters in the risk assessment strategy corresponding to the event category, and apply the increased scoring parameters to the risk assessment of the event category for all devices of the same model. If the event category does not meet the global adjustment requirements, a preset weighting coefficient is set for the scoring parameters corresponding to the event category, and the weighting coefficient is applied to the risk assessment of the event category by a single target device; whereby the single target device is the single device corresponding to the security event that meets the preset conditions in the event category.

[0135] In some embodiments, the analysis unit includes an extraction subunit and a score determination subunit; Extraction sub-units are used to extract feature labels corresponding to multiple risk assessment dimensions from the original feature data; among them, risk assessment dimensions include any combination of basic dimension, spatial dimension, duration dimension, facial identity dimension and scene dimension; The score determination subunit is used to determine the risk score corresponding to each feature label according to the preset score mapping rules.

[0136] In some embodiments, the extraction subunit is used to use the preset identity category as a feature label under the face identity dimension when a preset identity category is detected in a security event; When multiple identity categories are detected in a security incident, the identity category with the highest priority is used as the feature label under the face identity dimension, according to the priority order of identity categories. If multiple non-whitelisted identity categories are detected in a security incident, and the occupational confidence scores of all non-whitelisted identity categories meet the confidence threshold, the identity category with the lowest risk score is selected as the feature label under the facial identity dimension from among the multiple non-whitelisted identity categories.

[0137] In some embodiments, the fusion unit is used to fuse the risk scores of each target feature label according to a weighted maximum value method to obtain the total risk score of the security event; Alternatively, the risk score of each target feature label can be compared with a preset upper limit threshold; if there is a risk score that exceeds the upper limit threshold, the highest risk score that exceeds the upper limit threshold can be determined as the total risk score. Alternatively, the risk scores of each target feature label can be processed according to the set logical combination rules to determine the total risk score of the security incident.

[0138] In some embodiments, the fusion unit is configured to use the preset minimum value as the total risk score when the risk score of the target feature label meets the preset blocking condition. When the risk scores of multiple target feature labels simultaneously meet the preset multidimensional parallel conditions, the upper risk limit is used as the total risk score; where the upper risk limit is a value higher than the risk scores of each target feature label.

[0139] In some embodiments, the system further includes a judgment unit, an extraction unit, a generation unit, and a push unit; The judgment unit is used to determine whether a security event with a risk level higher than the silent handling level has occurred within the time period corresponding to the preset duration. The extraction unit is used to extract structured data of multiple security events with a risk level of silent handling within a time period, provided that no security events with a risk level higher than the silent handling level have occurred. The generation unit is used to generate summary information for structured data; the summary information is used to describe the events within a time period. The push unit is used to push summary information through a low-priority channel.

[0140] For a description of the features in the corresponding embodiment of the security event notification system, please refer to the relevant description of the corresponding embodiment of the security event notification method, which will not be repeated here.

[0141] As can be seen from the above technical solution, the analysis unit analyzes the original feature data of security events according to the risk assessment strategy, and determines multiple feature tags corresponding to each risk assessment dimension; each feature tag has its corresponding risk score. The filtering unit selects the target feature tag with the highest risk score from the multiple feature tags corresponding to each risk assessment dimension, effectively avoiding the linear accumulation and artificially high risk scores caused by multiple synonymous or overlapping tags triggering simultaneously within the same dimension. This limits the contribution of each dimension to the total risk score to a reasonable upper limit, thereby improving the accuracy of subsequent total risk score calculation and the stability of level determination. On this basis, the fusion unit fuses the risk scores of each target feature tag to obtain the total risk score of the security event. Through fusion processing, the total risk score is made more consistent with the actual threat level of the event, reducing the probability of false alarms and false negatives. The risk level of the security event is determined based on the total risk score; the risk level includes at least one silent processing level. When the risk level of the security event is at the silent processing level, the collection unit collects user behavior data of the security event within a preset time window. The adjustment unit adjusts the risk assessment strategy corresponding to security events based on user behavior data under the silent processing level and the classification feature information of the security events. This allows new security events with the same classification feature information to have their risk levels re-determined based on the adjusted risk assessment strategy. This reduces the situation where the system lags behind changes in user needs due to long-term reliance on fixed rules, and avoids over-responding to individual abnormal scenarios in the global adjustment. While ensuring the accuracy of risk assessment, it improves the system's adaptability to real user needs.

[0142] This application also provides an electronic device, including at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, and when the computer program is executed by the at least one processor, it causes the at least one processor to implement: Based on the risk assessment strategy, the original characteristic data of security incidents are analyzed to determine multiple feature labels corresponding to each risk assessment dimension; among them, each feature label has its corresponding risk score. The target feature label with the highest risk score is selected from multiple feature labels corresponding to each risk assessment dimension; each risk assessment dimension has one corresponding target feature label. The risk scores of each target feature label are fused to obtain the total risk score of the security event, and the risk level of the security event is determined based on the total risk score; wherein, the risk level includes at least one silent handling level; When the risk level of a security incident is classified as a silent handling level, user behavior data of the security incident is collected within a preset time window. Based on user behavior data under the silent processing level and the classification feature information corresponding to security events, the risk assessment strategy corresponding to security events is adjusted, and the risk level of new security events with the same classification feature information is determined according to the adjusted risk assessment strategy.

[0143] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0144] The foregoing has provided a detailed description of a security incident notification method, system, and electronic device provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only intended to help understand the method and core ideas of this application. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for notifying security incidents, characterized in that, include: Based on the risk assessment strategy, the original characteristic data of security incidents are analyzed to determine multiple feature labels corresponding to each risk assessment dimension; each feature label has its corresponding risk score. The target feature label with the highest risk score is selected from multiple feature labels corresponding to each risk assessment dimension; each risk assessment dimension has one corresponding target feature label. The risk scores of each target feature tag are fused to obtain the total risk score of the security event, and the risk level of the security event is determined based on the total risk score; wherein, the risk level includes at least one silent processing level; When the risk level of the security incident is at the silent handling level, user behavior data of the security incident is collected within a preset time window; Based on the user behavior data under the silent processing level and the classification feature information corresponding to the security event, the risk assessment strategy corresponding to the security event is adjusted, and the risk level of a new security event with the same classification feature information is determined according to the adjusted risk assessment strategy.

2. The security event notification method according to claim 1, characterized in that, Based on user behavior data under the aforementioned silent processing level and the classification feature information corresponding to the security event, adjust the risk assessment strategy corresponding to the security event, including: Based on the semantic tags, spatiotemporal features, and device feature information corresponding to multiple silent events, the multiple silent events are clustered to determine at least one event category; wherein, the silent events are security events with a risk level of silent handling level; Based on the user behavior data of multiple security events corresponding to the same event category, determine the behavioral statistics value of the event category; If the behavioral statistics meet preset conditions, the risk assessment strategy corresponding to the event category is adjusted.

3. The security event notification method according to claim 2, characterized in that, Based on user behavior data from multiple security events corresponding to the same event category, determine the behavioral statistics for the event category, including: Obtain the action type corresponding to the user behavior data; wherein, the action type includes at least one of watching live stream, watching video playback, downloading video, and sharing events; Based on the preset correspondence between action types and weights, the behavior weight corresponding to each action type is determined; among them, the behavior weight corresponding to downloading a recording is higher than the behavior weight corresponding to sharing an event, the behavior weight corresponding to sharing an event is higher than the behavior weight corresponding to viewing a recording playback, and the behavior weight corresponding to viewing a recording playback is higher than the behavior weight corresponding to viewing a live stream. Based on the action type and behavior weight corresponding to each security event, the user behavior data of multiple security events under the same event category are weighted and statistically analyzed to obtain the behavior statistics value of the event category; wherein, the behavior statistics value includes replay frequency and / or return visit rate.

4. The security event notification method according to claim 2, characterized in that, When the behavioral statistics value meets preset conditions, the risk assessment strategy corresponding to the event category is adjusted, including: Determine whether multiple security events in the event category that meet the preset conditions meet the global adjustment requirements; wherein, the global adjustment requirements include multiple security events in the event category that meet the preset conditions being distributed across multiple devices of multiple users; If the event category meets the global adjustment requirements, the scoring parameters in the risk assessment strategy corresponding to the event category are increased, and the increased scoring parameters are applied to the risk assessment of the event category for all devices of the same model. If the event category does not meet the global adjustment requirements, a preset weighting coefficient is set for the scoring parameters corresponding to the event category, and the weighting coefficient is applied to the risk assessment of the event category by a single target device; wherein, the single target device is the single device corresponding to the security event in the event category that meets the preset conditions.

5. The security event notification method according to claim 1, characterized in that, Based on the risk assessment strategy, the raw characteristic data of security incidents are analyzed to determine multiple feature labels corresponding to each risk assessment dimension, including: Extract feature labels corresponding to each of the multiple risk assessment dimensions from the original feature data; wherein, the risk assessment dimensions include any combination of basic dimension, spatial dimension, duration dimension, facial identity dimension and scene dimension; Based on the preset score mapping rules, the risk score corresponding to each feature label is determined.

6. The security event notification method according to claim 5, characterized in that, From the original feature data, feature labels corresponding to each of the multiple risk assessment dimensions are extracted, including: If a preset identity category is detected in the security event, the preset identity category will be used as a feature label under the face identity dimension; If multiple identity categories are detected in the security event, the identity category with the highest priority is used as the feature label under the face identity dimension according to the priority order of identity categories. If multiple non-whitelisted identity categories are detected in the security event, and the professional confidence scores of multiple non-whitelisted identity categories all meet the confidence threshold, the identity category with the lowest risk score is selected from the multiple non-whitelisted identity categories as the feature label under the face identity dimension.

7. The security event notification method according to claim 1, characterized in that, The risk scores of each of the target feature tags are fused to obtain the total risk score of the security event, including: The risk scores of each target feature label are fused using a weighted maximum value method to obtain the total risk score of the security event; Alternatively, the risk score of each target feature label can be compared with a preset upper limit threshold; if there is a risk score that exceeds the upper limit threshold, the highest risk score that exceeds the upper limit threshold can be determined as the total risk score. Alternatively, the risk scores of each target feature label can be processed according to a set logical combination rule to determine the total risk score of the security event.

8. The security event notification method according to claim 7, characterized in that, The risk scores of each target feature label are processed according to a set logical combination rule to determine the total risk score of the security event, including: If the risk score of the target feature label meets the preset blocking conditions, the preset lowest value will be used as the total risk score. When the risk scores of multiple target feature labels simultaneously meet the preset multidimensional parallel conditions, the upper risk limit value is taken as the total risk score; wherein, the upper risk limit value is a value higher than the risk scores of each target feature label.

9. The security event notification method according to claim 1, characterized in that, Also includes: Determine whether a security event with a risk level higher than the silent handling level occurs within the time period corresponding to the preset duration; In the absence of security events with a risk level higher than the silent handling level, extract structured data of multiple security events with a risk level of the silent handling level within the time period; A summary of the structured data is generated and pushed through a low-priority channel; wherein the summary is used to describe the events within the time period.

10. An electronic device, characterized in that, It includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, which, when executed by the at least one processor, causes the at least one processor to implement: Based on the risk assessment strategy, the original characteristic data of security incidents are analyzed to determine multiple feature labels corresponding to each risk assessment dimension; each feature label has its corresponding risk score. The target feature label with the highest risk score is selected from multiple feature labels corresponding to each risk assessment dimension; each risk assessment dimension has one corresponding target feature label. The risk scores of each target feature tag are fused to obtain the total risk score of the security event, and the risk level of the security event is determined based on the total risk score; wherein, the risk level includes at least one silent processing level; When the risk level of the security incident is at the silent handling level, user behavior data of the security incident is collected within a preset time window; Based on the user behavior data under the silent processing level and the classification feature information corresponding to the security event, the risk assessment strategy corresponding to the security event is adjusted, and the risk level of a new security event with the same classification feature information is determined according to the adjusted risk assessment strategy.