A Real-Time Penetration Estimation Method for AI Attackers Based on Bayesian-Gaussian Mixture Models

By combining a Bayesian Gaussian mixture model with three-dimensional feature vectors and online updates of Beta-Binomial soft counts, the problem of the inability to quantify the penetration of AI attackers in real time in existing technologies is solved, enabling real-time and accurate quantitative assessment of AI attackers and supporting dynamic adjustments to security investment strategies.

CN123116267APending Publication Date: 2026-11-17GUANBAO NETWORK SECURITY TECHNOLOGY (GUANGDONG) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611346907.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing cybersecurity defense systems are unable to effectively identify and quantify the penetration of AI attackers in real time, preventing enterprises from dynamically adjusting their security investment strategies.

Method used

We employ a Bayesian Gaussian mixture model-based approach, combining three-dimensional feature vectors (request rate, vulnerability complexity, and attack speed) to estimate the penetration of AI attackers in real time. Through Gaussian mixture modeling, online updates of Beta-Binomial soft counts, and concept drift adaptation, we achieve continuous quantitative assessment of AI attackers.

Benefits of technology

It enables real-time and accurate quantitative assessment of the penetration of AI attackers, supports dynamic adjustment of security investment strategies, and is efficient, accurate and robust, capable of estimating in seconds and providing robust input.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116267A_ABST
    Figure CN123116267A_ABST
Patent Text Reader

Abstract

This invention discloses a real-time estimation method for AI attacker penetration based on a Bayesian Gaussian mixture model, relating to the fields of cybersecurity and artificial intelligence. The method extracts three-dimensional features—request rate, vulnerability complexity, and attack speed—from network security data streams and performs logarithmic transformation to establish a bicomponent Gaussian mixture model. This model is then iteratively estimated using EM to obtain an offline reference benchmark. In the online phase, the likelihood score of the AI ​​attacker is calculated based on the three-dimensional features and recursively updated using a Beta-Binomial conjugate soft-counting method. Adaptive resetting is achieved through EWMA smoothing and Mahalanobis distance drift detection. The method outputs an AI penetration θ-point estimate and a posterior confidence interval, which also serves as an early warning basis for downstream decision-making. This invention can quantify AI attacker penetration in real time and adaptively adapt behavioral migration, providing core input for adaptive security investment decisions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of cybersecurity and artificial intelligence, specifically to a real-time estimation method for AI attacker penetration based on a Bayesian Gaussian mixture model (GMM), used to identify and quantify the current AI attack threat level in a cybersecurity defense system in real time. Background Technology

[0002] With the rapid development of AI technology, cyberattacks have shifted from traditional manual attack methods to AI agent-driven automated attack methods. AI-automated attacks exhibit significant advantages in cost, speed, and parallel scalability, and these differences are publicly recognized in the industry, constituting the objective technical background for the "attacker heterogeneity" described in this invention.

[0003] Traditional cybersecurity defense systems, based on the assumption of "homogeneity of attackers," are unable to effectively identify and respond to AI-driven attack threats. Existing security systems lack the ability to quantitatively assess the penetration of AI attackers in real time, preventing enterprises from dynamically adjusting their security investment strategies (i.e., security defense resource allocation plans) according to threat levels.

[0004] In the existing technology, there are several schemes that apply statistical learning methods to attack detection. For example, some methods are based on Gaussian mixture models (GMM) for network intrusion anomaly detection, fitting traffic features to a mixture Gaussian distribution to identify outlier attack behavior; some methods are based on Bayesian online learning for probability estimation, recursively updating the posterior distribution by counting observed events; and some methods are based on concept drift detection to build an adaptive intrusion detection framework, using metrics such as Mahalanobis distance to monitor feature distribution shifts and trigger model reset. These schemes are all common technical routes known in the field, and this invention refers to them collectively as "existing statistical learning attack detection schemes". However, the above schemes all have the following shortcomings: (1) They focus more on the binary judgment of "whether an attack exists", rather than on "the proportion of AI attackers in the overall attack (penetration)". (1) Real-time quantization of the continuous quantity "); (2) The joint distribution characteristics of multidimensional features and the distinguishability of human / AI attacker types are not systematically modeled; (3) Hard-label integer counting is used in the online update process, making it difficult to utilize the membership uncertainty information contained in the continuous likelihood score. Against this background, this invention proposes a real-time AI penetration estimation method that combines GMM joint modeling, online Beta-Binomial soft counting update and concept drift adaptation.

[0005] Technical basis for the distinguishing effect of three-dimensional features: This invention selects request rate, vulnerability complexity, and attack speed as three-dimensional features to distinguish human attackers (H) from AI attackers (A), and their distinguishability has clear theoretical and data support. In the applicant's internal simulation benchmark (based on the GMM classifier benchmark dataset labeled with historical attack samples), the typical feature profile of human attackers is labeled as low request rate, medium vulnerability complexity, and slow speed (typical values ​​are approximately request rate 10, CVSS 3.0, and attack interval 7200 seconds), while the typical profile of AI attackers is high request rate, high vulnerability complexity, and fast speed (typical values ​​are approximately request rate 150, CVSS 8.5, and attack interval 30 seconds). The two types show significant separation in the three-dimensional joint feature space. The above typical profile values ​​are derived from the applicant's internal simulation benchmark settings and are only used to illustrate the separation trend of the two types of attackers in the feature space, and do not constitute a limitation on the statistical distribution of the real network environment. It should be noted that the above values ​​are merely illustrative central values ​​under a simulation environment. In real-world environments, the feature distributions of the two types of attackers overlap (for example, slow scanners and fast manual brute-force attacks may have similar request rates). The difference in central values ​​does not necessarily indicate complete linear separability in reality. This is precisely why this invention employs a Gaussian Model (GMM) to model the three-dimensional joint distribution as a whole and uses soft counting to handle continuous membership uncertainty, rather than a simple threshold determination. It should be pointed out that even if the two types are somewhat separable in single-dimensional features, a simple threshold determination still cannot provide a continuous estimate of the mixing ratio (i.e., the proportion of AI attackers in the overall attack). The uncertainty cannot be quantified; this invention uses GMM to model the three-dimensional joint distribution and performs likelihood scoring and online conjugate updates on this basis, aiming to simultaneously obtain The continuous estimation, posterior confidence interval, and drift adaptation capabilities are precisely what simple thresholding schemes lack. Furthermore, based on the annotation and feature analysis of real events in the VERIS community database (VCDB, a public cybersecurity event database), AI candidate attacks (Botnet / Automated / Script / Scanning, etc.) exhibit observable differences in distribution trends relative to human attackers across three dimensions: request rate, vulnerability complexity, and attack speed. These differences are also observed in the labeled subsets, showing a significant tendency towards clustering. It should be noted that VERIS / VCDB itself does not provide raw traffic-level time-series data (such as per-event request rate or attack interval), nor does it directly label the "AI attacker / human attacker" field. The conclusions regarding the differences in the above dimensions are qualitative judgments based on publicly available event descriptions and existing knowledge of automated attack behavior patterns in the field. These judgments support the rationale for selecting three-dimensional features as the distinguishing criterion in this invention, rather than limiting the precise statistics on a specific dataset. It should be noted that although both request rate and attack speed are related to the time scale in terms of units, they describe behavioral attributes in different dimensions. Request rate reflects the scale of parallel enumeration by the attacker, while attack speed reflects the time efficiency of a single attack chain from initiation to success. Furthermore, this invention compresses scale differences by applying a logarithmic transformation to the three-dimensional feature vector in step 2 and uses a two-component GMM to model the three-dimensional joint distribution as a whole, so that the distinction depends on the structural shift of the joint distribution rather than the independence of a single dimension, thereby avoiding the degradation of distinction caused by related features.

[0006] Therefore, there is an urgent need for a technical solution that can identify the penetration level of AI attackers in real time and accurately, providing core input for adaptive security investment decision-making systems.

[0007] The technical limitations of existing commercial security products: Most mainstream commercial network security detection products (such as rule-based IDS / IPS, signature-based antivirus software, and some NDR platforms based on abnormal traffic) typically output isolated event alerts or binary judgment results when detecting AI-driven automated attacks. They cannot provide a macro-level situational indicator such as "the proportion of AI penetration in the current attack," thus failing to support Security Operations Centers (SOCs) in differentiated defense resource scheduling. Although some EDR products have introduced ML classifiers, their output is hard-labeled per event (Human / AI), and the models are mostly statically trained, lacking the ability to adapt to the long-term evolution of attacker behavior. This invention addresses these product-level gaps by providing a situational (rather than event-level) AI threat quantification assessment solution.

[0008] Feature-by-feature comparison between existing technologies and the present invention: To clearly define the differences between the present invention and existing technologies, the following table provides a comparison of key features between existing statistical learning attack detection schemes (GMM anomaly detection, Bayesian online counting, and drift detection resetting, three typical approaches):

[0009] Technical features GMM Anomaly Detection Solution Bayesian online counting scheme Drift detection reset scheme This invention Output Abnormal / Normal Binary Judgment Probability estimation (hard counting) Drift Marker Continuous Quantization Update method Offline / Batch Training Integer counting recursion Model Reset Soft count conjugate update Adaptability none (Recursive update) Yes (Reset) Joint Reset (Model + Prior) Quantification of uncertainty none There is a (posterior distribution). none It has (posterior distribution + confidence interval)

[0010] As can be seen from the table above, the GMM anomaly detection scheme outputs a binary judgment, the Bayesian online counting scheme is a hard counting recursion, and the drift detection reset scheme is a conceptual drift marker. The three schemes solve different technical problems, have different output calibers, and different working mechanisms. Those skilled in the art have no motivation to combine the three schemes according to the specific structure of this invention, which is "GMM offline reference benchmark → Beta soft counting online accumulation → drift trigger joint reset". The closed-loop synergy formed by this invention is not obvious.

[0011] References: The existing statistical learning attack detection schemes described in the background of this invention refer to the following three common technical routes known in the art, and their representative published documents are as follows (the following documents are the reference objects for comparison and distinction in this invention, and the specific implementation of the scheme of this invention does not depend on the specific parameters or conclusions of the documents): (a) Network intrusion anomaly detection based on Gaussian Mixture Model (GMM): R. Blanco, P. Malagón, S. Briongos, JM Moya, “Anomaly Detection Using Gaussian Mixture Probability Model to Implement Intrusion Detection System,” in Proc. of the 14th International Conference on Hybrid Artificial Intelligent Systems (HAIS 2019), Lecture Notes in ComputerScience, vol. 11734, Springer, Cham, 2019, pp. 648-659 (DOI: (a) Probability estimation based on Bayesian online learning: J. Xu, CR Shelton, “Intrusion Detection Using Continuous Time Bayesian Networks,” Journal of Artificial Intelligence Research, vol. 39, 2010, pp. 745-774 (DOI: 10.1613 / jair.3050); (b) Adaptive intrusion detection framework based on concept drift detection: J. Gama, I. Žliobaitė, A. Bifet, M. Pechenizkiy, A. Bouchachia, “A Survey on Concept Drift Adaptation,” ACM Computing Surveys, vol. 46, no. 4, Article 44, 2014 (DOI:10.1145 / 2523813). The above-mentioned documents are all prior art known to those skilled in the art. This invention uses them as references for comparison and distinction, and does not constitute a limitation on the specific content or patent number of the above-mentioned documents. Summary of the Invention

[0012] 1. Purpose of the invention

[0013] The purpose of this invention is to provide a real-time estimation method for AI attacker penetration based on a Bayesian-Gaussian mixture model, overcoming the aforementioned shortcomings of existing technologies and achieving accurate estimation of AI attacker penetration. Real-time, accurate quantitative assessment.

[0014] 2. Technical Solution

[0015] This invention, along with the applicant's other two submissions filed on the same day, "A Dynamic Security Investment Optimization Method Specific to Attacker Types" and "An Adaptive Honeypot Deployment and Self-Recovery Mechanism Based on Stackelberg Game Theory," constitute an adaptive security defense system based on the same inventive concept, corresponding to the three stages of threat perception, investment decision-making, and defense execution, respectively. This invention focuses on the penetration rate of AI attackers. The real-time estimation layer provides core input for investment decisions. This invention employs the following technical solution:

[0016] Step 1: Feature Extraction

[0017] Extracting three-dimensional feature vectors from network traffic, attack logs, and system alert data (such as...) Figure 2 (as shown)

[0018] - Request rate characteristics The number of attack requests per unit of time reflects the attacker's parallel scale.

[0019] - Vulnerability complexity characteristics The average CVSS score of the exploited vulnerabilities reflects the technical depth of the attack.

[0020] - Attack speed characteristic The time interval from the start of an attack to its success reflects the attacker's efficiency.

[0021] The specific extraction process is as follows: (a) Request rate Based on short sliding time window (window length) (a) Calculate the number of attack request packets within a 60-second time window (based on empirical values ​​or dynamically set according to traffic stability). This involves aggregating the original pcap / NetFlow packets by source IP or attack session and counting the packets per unit time. A short window is used to capture the attacker's instantaneous parallel enumeration scale. (b) Vulnerability complexity (c) Attack speed Defined by the time interval from the "attack start" moment (the arrival of the first attack request) to the "attack success" moment (the achievement of the predetermined intrusion goal, such as privilege acquisition / data leakage, and alarm confirmation), a long window at the attack chain level is adopted (and...). (Short window scale decoupling) to characterize the time efficiency of a single attack chain, thereby enhancing and Independence in engineering implementation; for attack events where no clear success indicator is observed (such as scanning and probing, vulnerability verification, etc., which have not entered the actual intrusion stage), the end time of the observation window or the current time is used as the cutoff point, and the right-censoring data processing method in survival analysis is adopted, with the truncation time interval as... The upper limit of the substitution value is used to avoid selection bias caused by only selecting "successful" samples and underestimation of the actual activity of AI attackers. In subsequent likelihood scoring and GMM modeling, a censoring indicator variable is introduced for right-censored samples. ( This indicates that a clear success was observed. (Indicates right censoring), and applies censoring weights to the likelihood score. ,in The censoring penalty coefficient (typically 0.5, but can be dynamically adjusted based on historical censoring rates or cross-validation results) is used to penalize unobserved successful attack events. The contribution of features is appropriately downweighted rather than treated exactly the same as successful samples, thus avoiding bias of censored samples on the GMM reference distribution and likelihood score. The censoring penalty coefficient... The calibration process is as follows: Select a subset of events from the historical validation set that are clearly marked as successful. The subset was manually labeled with AI tags and soft tags before deletion and weight reduction. A comparison is performed with the goal of minimizing the log-likelihood loss. Perform a grid search within the specified range to select the option that minimizes the AI ​​penetration estimation error on the validation set. Value. In real-time deployment scenarios where no labeled data is available, embodiments of the present invention also support label-free adaptive settings based on the censoring rate within the observation window: Let (in For the most recent sliding window The percentage of events with clearly observed success (i.e., the percentage of successful events) within the observation window, and the lower the percentage of confirmed successful attack events within the observation window (i.e., the more prevalent right censoring), the better. The smaller the value, the stronger the deweighting of right censoring events, thus enabling online adaptive adjustment of the censoring penalty intensity without relying on manual annotation.

[0022] The above three-dimensional feature vector is represented as: Logarithmic transformation is applied to the three-dimensional components before modeling. ( This aims to compress scale differences and improve the stability of GMM covariance estimation.

[0023] Step 2: Gaussian Mixture Modeling

[0024] Construct a Gaussian mixture model with two components, corresponding to a human attacker (component 0) and an AI attacker (component 1):

[0025]

[0026] in:

[0027] - Mixing weights for the k-th component

[0028] - Let be the mean vector of the k-th component (in the logarithmic feature space).

[0029] - The covariance matrix of the k-th component

[0030] - The probability density function of a multivariate Gaussian distribution

[0031] Step 3: Parameter Estimation

[0032] The Expectation-Maximization (EM) algorithm is used to iteratively estimate the GMM parameters (its log-likelihood convergence curve is shown in Figure 1). Figure 3 (as shown)

[0033] ① E-step: Calculate the posterior probability (responsibility) based on the current parameters.

[0034]

[0035] ② M-step: Update parameters based on the responsibility value of the E-step

[0036]

[0037]

[0038] Role of the GMM module in the framework: In this invention, the GMM model is positioned as an offline calibration module. Its parameters (category priors) Mean of the reference distribution With covariance Once the offline training phase in steps 2-3 is complete, the training path is fixed. In the online phase, the lightweight scoring path from step 4 is used by default to reduce latency; meanwhile, the GMM posterior probability path is retained. As an optional scoring reference / backup inference channel—this channel uses fixed parameters to analyze the characteristics of the current event. Calculating the posterior probability is part of the inference process and does not modify the model parameters; its output changes with the input features. GMM parameters do not participate in event-by-event updates during the regular online inference phase; they are only updated when concept drift triggers a joint reset, referencing the distribution parameters. and category prior The reset involves replacing the entire feature empirical distribution within the latest sliding window. This reset is an event-driven parameter reconstruction, rather than conventional online continuous training. The method claims of this invention revolve around the sequence of "feature extraction → GMM modeling → likelihood scoring → Beta conjugate update → output." The specific configuration of the online / offline GMM role (default lightweight path or enabling the GMM backup path) does not constitute a limitation on the core steps of this invention. To define the boundary between "offline fixing" and "drift reset," this invention clarifies that the time frequency of the reset operation is much lower than that of event-by-event updates (the reset interval is usually on the order of hours to days, while event-by-event soft-count updates are on the order of event arrival). Furthermore, during the reset, the old parameters are completely replaced with the latest window empirical distribution, and weighted fusion with the old parameters is rejected. This distinguishes it from continuous, gradual online learning updates. Therefore, "offline fixing" means that the GMM reference distribution remains constant between two resets and does not gradually change with individual events, rather than being absolutely immutable.

[0039] Step 4: AI Attack Likelihood Score

[0040] For each newly arriving attack event, its likelihood score from the AI ​​attacker is calculated based on the three-dimensional feature vector. :

[0041] - Request rate score:

[0042] - Vulnerability complexity score:

[0043] - Attack speed score:

[0044] Define "soft labels" based on continuous values ​​[0,1]. Unlike traditional binary hard classification labels (human / AI), soft labels retain the uncertainty information of feature membership, providing fine-grained counting input for subsequent Bayesian conjugate updates.

[0045] Weighted summation of soft label likelihood:

[0046]

[0047] The normalized denominators (200, 10, 3600) and weighting coefficients (0.35, 0.30, 0.35) in the above scoring formula are exemplary empirical values ​​used to illustrate the scoring construction method and do not constitute a limitation on the application scenarios or parameter range of the present invention. In a preferred embodiment, the corresponding thresholds and weights are calibrated according to the actual attack feature distribution (see Example 4A for the specific calibration method). See Example 4A for the specific calibration method of the normalization and weighting coefficients.

[0048] The propagation path of censoring weights: Step 1, the censoring indicator variable With censorship weight ( The event is passed to the online update via the following path: For events that have been deleted and downgraded, their soft tags are first deleted and re-weighted. And make the failure side a pseudo-count. In the Beta conjugate update in step 5, and Replace unweighted and ,Right now , Therefore, no clearly successful right censoring events were observed. The posterior contribution is limited to Within a certain proportion, avoid modeling it with the same weight as successful samples, thereby suppressing bias towards the GMM reference distribution and likelihood score.

[0049] The likelihood score is designed with an offline-trained GMM (AttackTypeClassifier) ​​as the benchmark: the GMM is fitted offline based on manually labeled historical samples (i.e., true binary labels), and the output... As The statistical reference anchor point for the coefficient design; in the online phase, to reduce latency, the above lightweight scoring is used by default, while the GMM posterior probability path is reserved as an optional scoring reference / backup inference channel, and the output of both is consistent (GMM parameters are fixed offline and called as needed in the online phase, with the lightweight scoring path used by default). It should be clarified that the reference benchmark here is the statistical anchor point for the lightweight scoring function coefficient design, not an external source of truth correction during online operation; the online estimation... It is entirely determined by soft count conjugate updates and drift resets, without relying on online GMM a posteriori for correction.

[0050] Switching logic between lightweight scoring path and GMM posterior path: In a preferred embodiment, the lightweight scoring path is the default enabled path, and the GMM posterior probability path is enabled in one of the following situations: (a) the output of the lightweight scoring path... With GMM posterior probability When the absolute value of the deviation exceeds a preset threshold (typically 0.15), a GMM posterior cross-check is triggered, and the current GMM output is used to replace the previous one. (b) Each system (c) When an external system requests it. It should be clarified that: the substitution described in scenario (a) only applies to the currently triggered event; subsequent events will still default to using the lightweight scoring path until the next trigger condition is met; this single substitution aims to eliminate occasional large deviations, does not change the dominant position of the lightweight scoring path, and does not use the GMM posterior path as a continuously running online correction channel. Except for the above scenarios, the lightweight scoring path is always running during the online phase to maintain low latency; the switching conditions are only preferred implementation methods and do not constitute a limitation on the relationship between the default lightweight path and the GMM backup channel. The master-slave roles of the two need to be further clarified: the lightweight scoring path is the target method (dominant channel) in the online phase of this invention, while the GMM posterior probability path is positioned as a periodic self-checking and deviation correction channel—that is, used to periodically verify whether the lightweight scoring has deviated significantly and to make a one-time correction to the current event when the deviation exceeds the limit, rather than a continuous online inference channel parallel to the lightweight scoring path; this master-slave positioning does not affect the availability of the GMM posterior probability path as an optional inference channel, and those skilled in the art can also change to using the GMM posterior probability path as the main channel in deployment scenarios with sufficient computing power or accuracy priority, and this configuration selection does not constitute a limitation on the core steps of this invention.

[0051] Step 5: Bayesian conjugate online update

[0052] Using a Beta-Binomial conjugate structure, the likelihood score is... The conjugate parameter is updated online as a soft count (fractional count). The illustration of the convergence of the estimate with cumulative observations is shown below. Figure 4 (as shown)

[0053]

[0054] in It serves as the posterior of the previous time step (and as the prior of the current time step).

[0055] Theoretical basis: Classical Beta-Binomial conjugate requires observations to be binomial distribution integer counts, while this invention uses continuous values... Consider it as a “soft count” imposed on a single observation – that is, the first observation is treated as a “soft count”. The interpretation of this observation is based on probability. For "success", in terms of probability This is a "failed" Bernoulli observation, and the continuous value is used as a pseudo-count for the linear update of the Beta distribution parameters. Since the Beta distribution is in the parameter... This invention maintains a closed-form update (the updated posterior remains a Beta distribution) under linear recursive updates, and is compatible with classic Beta-Binomial updates in the sense of "maintaining a closed-form posterior"; this soft counting... When degenerating to 0 or 1, it strictly degenerates into a classic Beta-Binomial integer count update. It should be noted that in classic Beta-Binomial, the posterior variance varies with the integer count. by The conclusion of rate contraction is based on the assumption that the observations are independent and identically distributed integer Bernoulli counts; in the soft-counting case of this invention, when the likelihood score When the corresponding true binary labels are approximately unbiased and the successive observations are approximately independent and identically distributed, the effective sample size is... The accumulation of these factors also causes the posterior variance to tend to shrink. Convergence is a description of asymptotic behavior under the above approximation assumptions (rather than a rigorous mathematical guarantee for general non-ideal scoring), and its convergence rate and accuracy depend on... The calibration quality is as follows (see Example 4A). Since the EWMA smoothing in step 6 has suppressed sudden disturbances in the time dimension, and the spurious counting error accumulated by soft counting is gradually diluted by the concentration effect of the conjugate posterior as the number of observations increases, this extension is controllable and reproducible in engineering. Further explanation is needed: the actual attack event flow may exhibit autocorrelation in the time dimension (e.g., clusters of similar attacks arriving within a short period). In this case, successive observations are not strictly independent and identically distributed. The convergence rate is only an asymptotic description under approximately unbiased and approximately iid conditions. In this case, the present invention uses EWMA smoothing in step 6 to suppress sudden temporal disturbances and uses concept drift joint reset to address distributional nonstationarity, making the estimator robust within engineering tolerances. However, the convergence rate may be slowed down due to autocorrelation, but this does not constitute a limitation on the feasibility of the method of the present invention. The above-mentioned "soft counting" statement has been explicitly stated in the claims and constitutes one of the improvements of the present invention to the existing hard counting Bayesian online update method.

[0056] The fundamental technical difference between soft counting and hard counting: Traditional hard-labeled counting directly classifies each observation as 0 or 1 (i.e., "success / failure"). Once a single event is misclassified, its impact on the conjugate parameter ( or The increment of ) is a complete unit, causing a one-time complete deviation in the posterior estimate, and the misjudgment effect is irreversible. The soft counting described in this invention uses continuous values. The probability that this observation belongs to an AI attacker is represented by the increment of the conjugate parameter. (and This means that the impact of a single misjudgment on the posterior parameter is limited to a finite amplitude (maximum not exceeding 1, and usually much less than 1), rather than pushing the posterior to an extreme all at once. With the accumulation of subsequent observations, the impact of occasional misjudgments is gradually diluted by the concentrated effect of the conjugate posterior. This characteristic of "finite impact + cumulative correction" is the essential technical contribution that distinguishes soft counting from the binary decision of hard counting, and it is also the key mechanism by which this invention can maintain estimation robustness under limited labeling quality. It should be emphasized that the "soft counting" mechanism itself constitutes an independent innovation of this invention compared to existing Bayesian online counting schemes (hard counting recursion)—it is not a simple mathematical replacement of the classic Beta-Binomial conjugate, but an improvement proposed for the specific technical problem of "hard labels introducing misjudgments when the boundaries are ambiguous and the impact is irreversible." The resulting gain in estimation smoothness and robustness is an important support for the high level of inventiveness of this invention. It should be noted that although "score counting / pseudo-counting" in Bayesian statistics has been disclosed in offline smoothing or fixed-weight scenarios such as click-through rate prediction and A / B testing, the soft counting described in this invention is fundamentally different: this soft counting is based on online dynamic scoring. It is event-driven, changes in real time with 3D features, and forms a closed-loop system together with 3D soft label design, censoring weight processing and drift joint reset, rather than an isolated offline smoothing or fixed weight design. Its creativity lies in the synergy of the modules in the above closed loop, rather than advocating the mathematical form of "soft counting" alone.

[0057] In the overall technical architecture of this invention, the "soft counting" update method is not an independent innovative proposition, but rather an important component of a closed-loop system in synergy with the following elements: (a) Three-dimensional feature-driven online dynamic scoring. (b) Weight correction for right censoring events; (c) Recursive update of Beta conjugate parameters; (d) Joint reset of reference distribution and prior under concept drift trigger. It is the overall synergy of the above steps, rather than the "soft counting" mechanism itself, that enables this invention to achieve adaptive tracking of attacker behavior migration and smooth suppression of the impact of single misjudgments in the specific technical problem of AI penetration estimation. Even though mathematical forms of "score counting" or "pseudo-counting" exist in the prior art, their application scenarios (such as offline smoothing, fixed-weight A / B testing) are completely different from the technical problem of "continuous quantification of AI penetration in real-time attack streams" solved by this invention. Those skilled in the art have no motivation to apply them to this and form the complete closed-loop solution of this invention. Any penetration estimation method that drives soft counting conjugate updates with online dynamic scoring and combines censoring weights and joint resets falls within the equivalent protection scope of claim 1 of this invention.

[0058] Step 6: Exponentially Weighted Moving Average Smoothing and Concept Drift Adaptation

[0059] To avoid sudden traffic surges Jitter, EWMA smoothing of the posterior mean:

[0060]

[0061] in The smoothing coefficient (typical value 0.1, typical range [0.05, 0.3]) is related to the Beta distribution parameter in step 5. They are independent of each other and will not be confused.

[0062] Simultaneously, concept drift detection is performed on the feature distribution based on Mahalanobis distance: a reference feature distribution (the mean obtained from offline calibration by GMM) is maintained. With covariance Composition, or composed of its weighted mixture), logarithmic characteristics of each new event. Calculate its Mahalanobis distance from the reference distribution. ;when When the threshold is exceeded (typical value 3.0, typical range [2.5, 3.5]), concept drift is determined to have occurred, triggering a reset. During the reset, the old reference feature distribution and prior parameters are reset to a distribution re-estimated based on the latest sliding window (window length W is taken as an empirical value of 200~500 events, or dynamically set according to the stability of attack traffic) (reset rather than weighted fusion with the old distribution) to quickly adapt to the attacker's behavioral shift; this reset is consistent with the offline fixing principle described in step 3: the GMM reference distribution parameters are only completely replaced with the latest sliding window empirical distribution when drift is triggered, which is an abnormal event-driven parameter reconstruction rather than regular online continuous training. If the threshold is continuously exceeded, the drift is re-detected using the reset time as the new reference point after a single reset; if the latch window after the single reset (typical length is...) Within each event If the limit is exceeded again, it indicates that the attack behavior is in a period of rapid change. At this point, continuous resets are triggered, and the sliding window length is shortened. Reduce the value to 50% of the original value to accelerate adaptation, and restore it after the distribution stabilizes. The usual values.

[0063] Step 7: Estimate the output

[0064] Smoothed output Point estimation and its posterior confidence interval based on the Beta distribution (95th percentile). This confidence interval also serves as an early warning basis for downstream decision-making: when the interval width exceeds the early warning threshold or a significant shift occurs in the point estimate, an early warning signal is output. The composition of each module of the method of this invention and its data flow connection relationship are as follows: Figure 1 As shown.

[0065] Self-checking and correction of systematic bias (“estimation-verification-correction” closed loop): It should be noted that the soft-count conjugate update described in step 5 is related to the likelihood score. When systematic bias (rather than random noise) is present, its posterior will still contract robustly with the accumulation of observations, but may converge to a level that deviates from the true value. The erroneous value is the failure mode of "high confidence but systematic deviation"; this mode cannot be detected by the confidence interval width itself (therefore the time zone width has narrowed to within the warning threshold). To address this, the present invention sets up an external consistency check in parallel during the output stage: when the posterior confidence interval has shrunk to within the warning threshold (high confidence) and the point estimation... When the deviation between the likelihood score and independent observations from the business side (including but not limited to the AI ​​attack percentage range released by the threat intelligence platform, the percentage of AI events obtained by manual sampling and labeling, or feedback indicators of downstream defense effectiveness) exceeds a preset consistency threshold, the likelihood score is determined to be true. If a systematic bias exists, a calibration alarm is output and a forced reset is triggered (the reference distribution and conjugate prior are reconstructed using the latest sliding window according to the joint reset process described in step 6) or manual intervention is requested for review. Therefore, this invention addresses low resolution due to insufficient samples with a "confidence interval width warning," addresses distribution-level non-stationarity with a "concept drift joint reset," and addresses systematic bias in scoring with "external consistency verification." These three elements together form a complete closed loop of "estimation-verification-correction," enabling... It is estimated that the system possesses detectable and recoverable capabilities under all three typical failure modes. The aforementioned external consistency verification is an optional enhancement measure, and the selection of its independent observations depends on the deployment environment, and does not constitute a limitation on the core steps of this invention.

[0066] 3. Technical Effects

[0067] Compared with the prior art, the present invention has the following advantages:

[0068] - Real-time performance: Able to complete within seconds It is estimated that this supports the dynamic adjustment of safe investment strategies.

[0069] - Accuracy: Based on joint modeling of 3D feature vectors and GMM, it can effectively distinguish between human attackers and AI attackers.

[0070] - Convergence: in likelihood scoring Under the assumptions that the relative true binary labels are approximately unbiased and that successive observations are approximately independent and identically distributed, the estimator is updated online via Beta-Binomial conjugate and smoothed using EWMA. The posterior confidence interval asymptotically shrinks with increasing effective sample size, and the posterior uncertainty exhibits a monotonically decreasing trend (this convergence property is a qualitative asymptotic description under the above approximation assumptions, not a rigorous mathematical guarantee for general non-ideal scoring; the actual convergence speed depends on...). The calibration quality and degree of deviation do not constitute a universal guarantee; in cases where attacker types are easily distinguishable, an engineering-usable confidence width can be achieved with approximately several hundred observations (this number of observations is an exemplary empirical value, and the actual convergence speed depends on the likelihood score). The degree of deviation from the actual label does not constitute a universal guarantee.

[0071] - Robustness: This invention provides robustness for dynamic and secure investment decisions. Input. Because the estimator cannot know the true value during the online phase. This invention does not rely on comparison with the true value, but uses the posterior confidence interval that can be calculated online as the basis for early warning: when the confidence interval width exceeds the early warning threshold, a low confidence warning signal is output, suggesting that downstream companies postpone high-risk decisions, thereby maintaining output stability under extreme conditions where estimation uncertainty increases, and avoiding the failure of safety investments due to estimation distortion.

[0072] - Scalability: Supports expansion of feature dimensions and increase in the number of GMM components.

[0073] - Verifiable real-time performance: Under standard testing environment (mainstream x86 architecture commercial server CPU, total time for single-event feature extraction + likelihood scoring + conjugate update < 5 ms, see Example 4B for details), the latency requirements for dynamic adjustment of security investment strategies are met.

[0074] - Synergistic Effect Explanation: The three core components of this invention—GMM offline calibration, Beta soft count online update, and drift-triggered joint reset—are not simply superimposed, but rather work together to produce synergistic effects that cannot be achieved by a single technical means: (a) GMM provides an offline reference benchmark, enabling the lightweight scoring path to maintain low latency while remaining consistent with the GMM posterior probability caliber, thus resolving the contradiction of "precision-latency" being difficult to achieve simultaneously; (b) Beta soft count update introduces a conjugate accumulation mechanism on top of the GMM reference benchmark, causing the estimation accuracy to continuously improve with the accumulation of observations (the posterior variance gradually decreases with the increase of the effective sample size), which is a temporal accumulation gain that GMM single classification itself cannot provide; (c) When Mahalanobis distance detects concept drift, the "joint reset" simultaneously reconstructs the GMM reference distribution and the Beta conjugate prior, enabling the estimator to quickly adapt to the attacker's behavior shift while retaining a short-term convergence advantage. In contrast, existing technologies that only reset the classification model (without involving Bayesian prior collaboration) require the re-accumulation of historical information after drift and recover slowly. The synergistic effect of "reference benchmark supporting lightweight path + conjugate accumulation improving accuracy + joint reset taking into account both adaptation and convergence" constitutes the non-obvious technical contribution of this invention compared to existing schemes that use GMM classification, Bayesian recursion or drift detection alone.

[0075] Industrial Application Statement: The GMM offline calibration, soft counter online update and joint reset process described in this invention has been fully disclosed in this specification. Those skilled in the art can carry out industrial deployment and reproduction based on the formulas, parameter ranges and calibration process recorded in this specification, without relying on specific undisclosed data or proprietary code. Attached Figure Description

[0076] Figure 1 This is a schematic diagram showing the composition of each module and the data flow connection relationship of the method of the present invention.

[0077] Figure 2 This is a schematic diagram of the source and transformation process of three-dimensional feature extraction in this invention.

[0078] Figure 3 This is a schematic diagram of the convergence curve of the log-likelihood value changing with the number of iterations during the training process of the GMM model EM algorithm of this invention.

[0079] Figure 4 For the present invention A schematic diagram showing the convergence of the estimated value and the 95% confidence interval as the number of observed events accumulates.

[0080] Figure 5 For the present invention in different A schematic diagram of the estimation error distribution under the true value setting.

[0081] (Note: Figure 3-5This diagram illustrates the convergence and accuracy of the method of the present invention. The number of convergence iterations and the error magnitude are derived from typical operating results under the synthetic benchmark data described in Example 4, and are actual measured data under non-specific real network environments. Verification data under real / semi-real environments are shown in Examples 4 and 6. Detailed Implementation

[0082] Example 1: Offline Training Phase

[0083] 1. Collect labeled attack samples (1000 human attacks and 1000 AI attacks). Samples can be sourced from public cybersecurity databases such as Veris / VCDB, or from the company's historical attack logs.

[0084] 2. Extract the three-dimensional feature vectors according to the procedure in step 1 and apply a logarithmic transformation.

[0085] 3. Train the GMM parameters using the EM algorithm (Ledoit-Wolf covariance shrinkage can be used to improve estimation stability). The initial mean can be referenced from the typical profile of a human attacker (request rate 10, CVSS 3.0, interval 7200 seconds) and the typical profile of an AI attacker (request rate 150, CVSS 8.5, interval 30 seconds).

[0086] 4. Verify the classification accuracy of GMM (target: ≥95%), and output the prior class. With reference distribution parameters As an online reference benchmark; it should be noted that the above typical profile values ​​are calibration values ​​under the applicant's internal simulation benchmark environment. They are schematic diagrams presented to facilitate understanding of the technical principles of this invention. In actual deployment, the mean parameter and covariance matrix of GMM need to be recalibrated based on historical attack samples of the real network environment. This invention does not provide any performance guarantee for its statistical presets.

[0087] Example 2: Online Estimation Stage

[0088] 1. Receive attack event data in real time.

[0089] 2. Extract the three-dimensional feature vectors as in step 1 and apply a logarithmic transformation.

[0090] 3. Calculate the likelihood score based on the lightweight scoring path in step 4. (or GMM posterior probability) (as an offline reference benchmark)

[0091] 4. Employ Beta-Binomial soft-count conjugate update. Posterior distribution (step 5)

[0092] 5. Output Point estimation and confidence interval (step 7); the present invention in different A schematic diagram of the estimation error distribution under the true value setting is shown below. Figure 5 As shown.

[0093] Example 3: Adaptive Investment Linkage

[0094] 1. [The following is a list of items / items] The estimated value and its confidence interval are passed to a dynamic, safe investment decision-making unit (such as the Gordon-Loeb optimizer).

[0095] 2. According to Values ​​and confidence intervals adjust attacker-type-specific vulnerability functions

[0096] 3. Output a dynamically optimal safety investment allocation scheme

[0097] Example 4A: Hyperparameter Calibration Process

[0098] (Note: Embodiments 4A and 4B of this invention are two parallel sub-items of the same embodiment 4, corresponding to hyperparameter calibration and real-time benchmark testing, respectively; the following embodiments are numbered sequentially from embodiment 5 onwards, and the numbering system remains consistent.)

[0099] The normalized denominators (200, 10, 3600) and weighting coefficients (0.35, 0.30, 0.35) in step 4 are exemplary values. In actual deployment, they should be calibrated according to the following process: (a) Statistically analyze the attacks of human and AI attackers separately from historical attack sample sets (such as VERIS / VCDB labeled data or enterprise historical logs). (a) The normalized denominator for each dimension is set to the 90th percentile of the AI ​​attacker's experience distribution for that dimension (making the score of a typical AI event close to 1), where the vulnerability complexity dimension is directly linearly normalized using the CVSS full score of 10; (b) The weighting coefficients are obtained by performing logistic regression or grid search on the labeled samples to make the overall score... Compared with GMM offline reference benchmark The Pearson correlation coefficient is maximized, and the resulting coefficients satisfy a sum of 1 and each lies within the interval [0.2, 0.4]. The calibrated thresholds and coefficients can be updated periodically or triggered by events based on historical distribution (see claim 6). To avoid the overfitting risk caused by using the same labeled set to train the GMM and calibrate the scoring coefficients, the calibration process is performed on an independent test set or using cross-validation. That is, the labeled samples used for GMM training and the validation samples used for scoring coefficient calibration are independent in sample partitioning, so that the correlation between the calibrated scoring coefficients and the offline reference benchmark of the GMM can be reproduced on unseen samples, rather than being a post-fit of the training set.

[0100] Example 4B: Real-time Benchmarking

[0101] In a standard testing environment (mainstream x86 architecture commercial server CPU / 32GB RAM / Python 3.11 compatible runtime environment), 10,000 synthetic attack event streams (labeled with AI percentages) were tested. =0.5) Perform single-event processing latency test:

[0102] - Single-event feature extraction: Sub-millisecond level

[0103] - Likelihood score (lightweight path): Sub-millisecond

[0104] - Beta conjugate update + EWMA smoothing: Submillisecond level

[0105] - The total single-event processing latency is within milliseconds (< 5 ms), meeting the requirements for second-level or even sub-second-level policy refresh.

[0106] Single event processing latency < 5 ms, calculated based on a refresh strategy every 1000 events, overall... It is estimated that the process can be completed within seconds, which is consistent with the real-time requirements described in Section 1 of the technical performance specification.

[0107] This test is based on synthetic benchmark data and is used to verify the computational feasibility of the method. It does not constitute a performance commitment for real network environments.

[0108] Example 5: Parameter Boundary Protection and Cold Start

[0109] Cold start: When the system has no historical prior knowledge, this embodiment of the invention initializes the Beta conjugate prior as a uniform prior, that is... , corresponding to unbiased initial Estimate; accumulated from observations by likelihood score Recursive update.

[0110] Numerical boundary protection: In extreme continuous attack scenarios (such as...) (continuously 1 or continuously 0), Beta distribution parameter It will monotonically increase. To prevent excessively large parameter values ​​from causing update stagnation or floating-point overflow, this embodiment of the invention presets an upper limit for the parameter (e.g., ...). If it exceeds the limit, use a scaling factor. right Constant amplitude compression, maintain estimated value Unchanged. This protection mechanism is an optional engineering measure used to maintain numerical stability during long-term operation.

[0111] Conceptual Drift Reset: When Mahalanobis drift detection is triggered, this embodiment of the invention resets the old reference feature distribution and prior parameters to a distribution re-estimated based on the latest sliding window (window length W is taken as an empirical value of 200~500 events, or dynamically set according to the stability of attack traffic) (resetting rather than weighting and fusing with the old distribution) to quickly adapt to the attacker's behavioral shift. During reference distribution reconstruction, the logarithmic mean and covariance of the events within the sliding window are used to replace the original GMM reference parameters, and the Beta conjugate prior is reset accordingly to the current posterior (specific operation: let...). Take the empirical mean / covariance of the logarithmic features within the window. The method involves summing the cumulative soft counts within the window (i.e., completely replacing the old reference distribution and prior with the new estimate, rather than weighting and fusing it with the old parameters), enabling the estimator to quickly converge to the migrated attacker behavior distribution. This reset operation simultaneously applies to the GMM reference distribution and the Beta conjugate prior parameters, constituting a "joint reset" strategy—a concept of collaborative reset between drift detection and Bayesian priors, not disclosed in existing schemes using drift detection alone, and a key mechanism to ensure the estimator does not fail under long-term evolution of attacker behavior. To mitigate the tailing effect of data mixed with the new reference distribution before and after drift, embodiments of this invention support accurately defining the drift boundary after drift detection using change point detection (such as based on cumulative sum CUSUM or bisection log-likelihood ratio test), reconstructing the reference distribution and prior only with window data after the drift point; or using exponentially decaying weights (rather than equal-weighted sliding windows) to weight samples within the window, gradually forgetting the influence of old samples far from the current time. The above window filtering / decay mechanisms are optional engineering measures and do not constitute a limitation on the core steps of this invention. In scenarios where drift occurs frequently, if historical accumulated information is completely discarded with each reset, the estimator may remain in a low-confidence state for a long time. To this end, the present invention also supports an alternative to adaptive reset: partially retaining historical prior information during reset (for example, using a weighted fusion of historical posterior and new window estimation as the reset target, or adopting a mechanism that adaptively shortens or extends the sliding window length W with the drift frequency) to achieve a balance between rapid adaptation and preservation of historical information. This alternative is an optional engineering measure and does not constitute a limitation on the core steps of the present invention.

[0112] Example 6: Semi-realistic verification based on publicly available VERIS data

[0113] Semi-realistic verification is performed using real cybersecurity event records from the VERIS Community Database (VCDB). Data source constraints must be noted: VERIS / VCDB provides event-level descriptions (e.g., victim, attack type, impact) and CVSS vulnerability scores, but not raw traffic-level time-series data. Therefore, this embodiment only extracts the vulnerability complexity dimension (approximated by the event-related CVSS score) from the three-dimensional features that can be directly parsed from the event description, and uses it as a single-dimensional feature for AI candidate subset verification. The request rate and attack speed time-series dimensions are not extracted in this embodiment; their distinguishing power is verified by the synthetic benchmark calibration in Embodiments 1-2 and internal simulation data (including event timestamps), respectively, avoiding reliance on time-series fields not included in VERIS. An event subset containing AI candidate types such as Botnet / Automated / Script / Scanning is selected as AI attacker annotation, and the rest are used as human / other annotations. Subsequent estimation processes are performed based on the extractable single-dimensional vulnerability complexity features (i.e., only the vulnerability complexity dimension is extracted in step 1, the score of this dimension participates normally in weighting in step 4, and the request rate and attack speed time-series dimensions are updated in soft-count conjugate and...). During estimation, missing values ​​are treated, and the online estimation process in steps 5-7 proceeds as usual. The rule for determining the AI ​​candidate type is as follows: the above-mentioned type keywords appearing in the action / actor field of the VERIS event are used as the inclusion standard; when the same event is labeled with multiple candidate types or contains both human and AI candidate labels, the priority rule of "any AI candidate type matching is judged as an AI candidate" is applied, and the main attack type field of the event is used for disambiguation; for events with missing actor fields or types of "Unknown / Unknown", they are excluded from the AI ​​candidate subset to avoid interference from unlabeled samples in the statistical caliber. Verification shows that: (a) on the AI ​​candidate events labeled with VERIS, the output of this invention The mean is significantly higher than that of the human event subset, proving that the vulnerability complexity dimension has discriminative power in real data; (b) on the event stream constructed according to this single-dimensional feature, The estimated value increases monotonically with the proportion of AI candidate events, and is correlated with the annotation penetration rate. (i.e., the proportion of AI attackers) shows a positive correlation trend, indicating that the estimation results of this invention are consistent with the changing direction of actual attack composition; (c) using VERIS annotation as the binary classification ground truth (AI candidate / non-AI candidate) to evaluate the accuracy of this invention in AI discrimination of a single event, it shows usable classification and discrimination capabilities on the labeled subset. It should be noted that VERIS / VCDB itself does not directly annotate the "AI attacker / human attacker" field. This invention uses automated attack types such as Botnet / Automated / Script / Scanning as a proxy subset of "AI candidate". The above verification is a semi-realistic illustrative result based on this proxy annotation. Its specific discrimination value varies with the selected subset and statistical caliber, and does not constitute a quantitative limitation on the discrimination power of this invention; the feasibility of this invention is supported by the synthetic benchmark calibration of Examples 1-2 and the real data trend of this embodiment.

[0114] Example 7: Exemplary Mapping of Threat Level Three-Tier Classification

[0115] To ensure the feasibility of the "low, medium, and high threat level discrimination requirement" as described in claim 4, the embodiments of the present invention provide... An exemplary mapping between estimated values ​​and threat level classifications is shown in the table below. This mapping is only used to illustrate the engineering implications of the confidence interval width threshold and does not constitute a technical definition of threat level classification. The range of values ​​or the boundaries of the levels are limited; in actual deployment, the threshold values ​​of each level can be adjusted according to the distribution of historical attack characteristics based on the dynamic calibration process described in claim 6.

[0116] Posterior confidence interval width Point estimation interval (example) Threat Level Recommendations for defense resource configuration (example) ≥ 0.1 (low confidence) — Pending confirmation (low resolution) Trigger supplementary observations / manual assessment < 0.1 (high confidence) [0, 0.33) Low Maintaining baseline defense < 0.1 (high confidence) [0.33, 0.67) middle Enhance targeted monitoring and response < 0.1 (high confidence) [0.67, 1] high Initiate enhanced defense resource allocation

[0117] The interval boundaries (0.33, 0.67) shown in the table above are examples of equal divisions that are easy for operations and maintenance to understand in typical attack and defense scenarios; when the confidence interval width is ≥0.1, it indicates that the current effective sample size is insufficient to represent the given information. Upon reaching the three resolution levels, this invention outputs a pending confirmation state instead of forcibly classifying the data, thus preventing downstream decision-making from failing based on low-resolution input. It should be noted that the "pending confirmation state" itself is a valid situational output (indicating insufficient current observation samples and limited estimation resolution), and does not negate the invention's real-time quantification capability for #imgpt164# under sufficient observation conditions. In the high confidence region with a confidence interval width <0.1, this invention immediately outputs low / medium / high three-level classification results, meeting real-time requirements. The pending confirmation state only appears in cold starts with scarce samples or in strong noise windows, and automatically transitions to a high confidence level as observations accumulate, which does not conflict with the invention's objective of "real-time, accurate quantitative assessment."

Claims

1. A real-time estimation method for AI attacker penetration applied to network security defense systems, implemented based on a Bayesian-Gaussian mixture model, characterized in that... Includes the following steps: S1: Extract request rate features, vulnerability complexity features, and attack speed features from network security data streams in real time to form a three-dimensional feature vector; wherein, the request rate feature is calculated based on the number of attack request packets counted within a short sliding time window, the vulnerability complexity feature is obtained by parsing the CVSS base score of the exploited CVE, and the attack speed feature is defined by the time interval from the start of the attack to the time of attack success or truncation, and a censoring indicator variable is introduced for right censoring events for which no clear success indicator is observed; S2: Establish a two-component Gaussian mixture model that includes both human attacker components and AI attacker components; S3: The Gaussian mixture model parameters are iteratively estimated using the expectation-maximization algorithm. The resulting model is used as an offline reference benchmark and provides prior probability support for online likelihood scoring. S4: Calculate the AI ​​attacker likelihood score s_t∈[0,1] based on the three-dimensional feature vector, as a soft label; the likelihood score is obtained as follows: the posterior probability p(AI|x̃_t) output by the Gaussian mixture model trained in step S3 is used as the offline reference benchmark; in the online stage, a lightweight scoring path that performs normalized weighted calculation on each component of the three-dimensional feature vector is used by default. The normalization threshold and weighting coefficient of the lightweight scoring path are pre-calibrated by maximizing the Pearson correlation coefficient between the comprehensive score s_t output by the path and the GMM posterior probability p(AI|x̃_t); the Gaussian mixture model posterior probability path is reserved as a reference and backup inference channel for online scoring; S5: The Beta-Binomial conjugate structure is adopted, and the likelihood score s_t is used as a soft count to update the conjugate parameters online: α_t=α_(t-1)+s_t, β_t=β_(t-1)+(1-s_t); wherein the soft count uses continuous values ​​s_t and (1-s_t) as pseudo-count increment values ​​for the "success" side and "failure" side, respectively, which is different from the hard label counting method that directly judges the observed event as an integer increment of 0 or 1, thereby preserving the uncertainty of feature membership degree through continuous values ​​s_t∈[0,1]. S6: Output the AI ​​penetration point estimate θ and its posterior confidence interval; the θ and confidence interval serve as input to the downstream dynamic security investment decision module, driving the configuration of defense resources specific to the attacker type.

2. The method according to claim 1, characterized in that, The three-dimensional feature vector includes: request rate feature, vulnerability complexity feature (average CVSS score of the exploited vulnerability), and attack speed feature (time interval from attack start to success). The likelihood score is obtained by weighting the request rate score, vulnerability complexity score, and attack speed score, where the sum of the weighting coefficients is 1, and the typical value range of each coefficient is [0.2, 0.4]. Before the weighted summation in step S4, each component in the three-dimensional feature vector is normalized: the request rate feature is truncated and normalized using a preset maximum concurrent request count, the vulnerability complexity feature is linearly normalized using a CVSS full score of 10, and the attack speed feature is inversely normalized using a preset time window length. The preset maximum concurrent request count and preset time window length are statistically calibrated based on the distribution of historical attack features.

3. The method according to claim 1, characterized in that, Also includes: The posterior mean obtained in step S5 is smoothed by an exponentially weighted moving average, with the smoothing coefficient typically ranging from [0.05, 0.3]. Concept drift detection is performed on the feature distribution based on Mahalanobis distance. When the new event feature deviates from the reference distribution by more than a threshold, the GMM reference distribution and the Beta conjugate prior parameters are simultaneously and jointly reset. The reset is an abnormal event-driven synchronous reconstruction of the GMM reference distribution and the Beta conjugate prior, which is different from the conventional online update or single model reset that only applies to the classification model parameters. The typical range of the drift detection threshold is [2.5, 3.5].

4. The method according to claim 3, characterized in that, By using online Beta-Binomial conjugate updates and exponentially weighted moving average smoothing, the posterior confidence interval gradually shrinks as the effective sample size increases. When the width of the 95% confidence interval is less than a preset threshold (typically 0.1), the posterior distribution is determined to be concentrated in the high confidence region. This threshold corresponds to the resolution requirement of θ estimation to distinguish between low, medium, and high threat levels.

5. The method according to claim 1, characterized in that, The AI ​​penetration point estimate θ and its confidence interval are used as input for dynamic security investment decisions and provide uncertainty quantification basis for downstream decisions. When the posterior confidence interval width exceeds the warning threshold, the method outputs a warning signal to maintain output stability and prevent downstream decisions from failing due to input distortion. Furthermore, when the system has no historical prior knowledge, the Beta-Binomial conjugate prior in step S5 is initialized as a uniform prior, i.e., α_0=β_0=1. When the Beta distribution parameters α and β exceed the preset upper limit, α and β are compressed with an equal-amplitude scaling factor to maintain numerical stability, and the estimated value of θ=α / (α+β) remains unchanged after compression.

6. The method according to claim 1, characterized in that, In step S4, the normalization threshold and weighting coefficient used in the normalization weighting calculation of each component of the three-dimensional feature vector in the online default lightweight scoring path are dynamically calibrated and periodically or event-triggered updated based on the distribution of historical attack features.

7. A real-time AI attacker penetration estimation system deployed in a network security defense system, characterized in that, include: The method comprises a data acquisition module, a feature extraction module, a Gaussian mixture model (GMM) modeling module, a Bayesian conjugate online update module, an exponentially weighted moving average smoothing and concept drift detection module, and an estimation output module. The output of the data acquisition module is connected to the input of the feature extraction module. The output of the feature extraction module is connected to the inputs of both the GMM modeling module and the Bayesian conjugate online update module. The output of the GMM modeling module is connected to the input of the Bayesian conjugate online update module. The GMM modeling module provides an offline reference baseline and reference distribution parameters to the Bayesian conjugate online update module. The output of the Bayesian conjugate online update module is connected to the input of the estimation output module via the exponentially weighted moving average smoothing and concept drift detection module. The GMM modeling module is used to offline calibrate the category prior and reference distribution parameters and provide a reference baseline to the Bayesian conjugate online update module. Each module is configured to perform the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1 to 6.

9. The method according to claim 1, characterized in that, In step S1, for attack events for which no clear success indicator is observed, a censoring indicator variable δ_t∈{0,1} and a censoring weight w_t=δ_t+(1-δ_t)·λ are introduced, where λ∈(0,1) is the censoring penalty coefficient. In the Beta-Binomial conjugate update in step S5, the censored weighted soft count s_t^(w)=w_t·s_t and its failure side (1-s_t^(w)) are used to replace the unweighted s_t and (1-s_t) in the conjugate parameter update.

10. The method according to claim 4, characterized in that, In step S6, when the width of the posterior confidence interval is less than a preset threshold, the interval where the θ point is estimated is output as low, medium, or high threat level; when the width of the posterior confidence interval is greater than or equal to the preset threshold, a pending confirmation status is output instead of forced classification.