Adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation
Through the Gaussian process-based adaptive software system abnormal behavior detection system, using technologies such as active control signals and parameter deviation random frameworks, the problem of insufficient observation data is solved, the accuracy and timeliness of detection are improved, and the system abnormality rate is significantly reduced. , improves the system's effectiveness.
Patent Information
- Application Number
- CN202410934010.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-12
AI Technical Summary
The model parameter deviation detection method of existing adaptive software systems ignores the problem of insufficient observation data information, resulting in insufficient detection to reveal the potential deviation of model parameter θ.
Adaptive software system abnormal behavior detection system based on Gaussian process is adopted, and through the parameter deviation random framework, bidirectional detector and signal optimizer, the active control signal stimulation system is used to generate more valuable observation data to evaluate and correct model parameter deviations.
It improves the accuracy and timeliness of the parameter deviation detection of adaptive software system model, reduces detection delay, reduces system abnormality rate, and improves the overall effectiveness of the system through active signaling strategies.
Smart Images

Figure CN118939534A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of computer software, and in particular relates to an abnormal behavior detection system for an adaptive software system based on Gaussian process deviation evaluation. Background Art
[0002] Model parameter deviations in adaptive software systems can lead to many reliability issues. In order to detect this parameter deviation quickly and accurately, existing work has proposed a variety of runtime monitoring methods for adaptive software systems. These deviation detection methods determine whether model parameter deviations have occurred by monitoring the execution trajectory of the self-running adaptive software system, but ignore the challenge of insufficient observation data information. This detection challenge may result in such detection methods that rely solely on passive observations being insufficient to reveal potential deviations in the model parameters θ. Summary of the invention
[0003] In view of the deficiencies in the prior art, the present invention provides an abnormal behavior detection system for an adaptive software system based on Gaussian process deviation evaluation. Compared with the existing fuzzy observation method, this technology stimulates the adaptive software system to generate more valuable observation data through appropriate active control signals to improve the detection of model parameter deviations of the adaptive software system.
[0004] To achieve the above object, the present invention adopts the following technical solutions:
[0005] The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation is characterized by comprising:
[0006] A parameter deviation stochastic framework to connect the nominal model and observable behavior of adaptive software systems and to explain the impact of ambiguous observations on the detection of model parameter deviations;
[0007] A bidirectional detector that assesses the confidence of the adaptive software system in whether it suffers from model parameter deviations and initiates active control signal injection when the confidence exceeds expectations;
[0008] Signal optimizer, used to determine the strength of the active control signal.
[0009] To optimize the above technical solutions, the specific measures taken also include:
[0010] Furthermore, the parameter deviation random framework consists of a priori prediction based on Gaussian fuzzy and a posterior regression based on Gaussian fuzzy; the prior prediction describes the temporal changes of the model parameters, and the posterior regression describes the dependency between the model parameters and the observed variables.
[0011] Furthermore, the parameter deviation random framework is expressed as follows:
[0012]
[0013] Where, formulas a and b represent the prior prediction based on Gaussian blur, and formula c represents the posterior regression based on Gaussian blur; θ t Represents the model parameter, which is a random variable that obeys the normal distribution. The subscript represents any time t; f t The transition function of the model parameters is used to describe the deviation of the model parameters caused by internal factors as a random walk process, and the deviation of the model parameters caused by external factors as a change point with low probability. represents normal distribution; Q represents the variance of model parameters; ∈ t is the parameter describing the model θ t The random walk step length of the smooth change process; ω·Θ is the parameter θ describing the model t The gradual change of the mutation process, ω represents the probability of gradual change, Θ represents the amplitude of gradual change; the control parameter u t 、System output y t and environmental input α t is the observed variable, LS represents the time-invariant linear system, is the system output y t The observation error.
[0014] Furthermore, in the Gaussian blur-based posterior regression, a linear time-invariant model γ is used to describe the environmental input α t For system state x t The influence of , thus obtaining a time-invariant linear system LS as follows:
[0015]
[0016] In the formula, θ D ,θ C ,θ O They are the model parameters θ t The delay component, control component and observation component of
[0017] Furthermore, the bidirectional detector evaluates the confidence of the adaptive software system in whether it suffers from model parameter deviation, specifically: on the one hand, when a non-zero control signal can be observed, traditional Bayesian inference is used to determine the current model parameter θ t Whether it deviates from its safety interval; on the other hand, when zero-value control signals affected by fuzzy observations are continuously observed, the results of hypothesis testing and worst-case estimation are combined to decide whether to trigger active control signal injection.
[0018] Furthermore, when a non-zero control signal can be observed, the abnormal behavior detection and evaluation using non-fuzzy observation is specifically as follows: first, the model parameter θ is predicted according to the Gaussian process functiont The prior distribution θ t |f t , and then use the observed variables to calculate the model parameters θ t The posterior distribution θ t |f t , and finally based on the model parameters θ t The posterior distribution of is used to calculate the confidence C that no abnormal behavior has occurred dev , the confidence level C dev is the posterior distribution θ t |f t Falling into its safety zone Probability within.
[0019] Furthermore, when the zero-value control signal affected by the fuzzy observation is continuously observed, the parameter deviation of the fuzzy observation is evaluated, specifically: the bidirectional detector calculates the model parameter θ during the fuzzy observation t In the control parameter u t The confidence level C is evaluated by the probability that the value is still in its safe interval when it remains zero. vague ; Bidirectional detector according to model parameters θ t The transition function f t Calculate the required confidence level C vague .
[0020] Furthermore, the bidirectional detector uses C rw and C cp They represent the confidence of the occurrence of random walk type parameter deviation and change point type parameter deviation during fuzzy observation, and the evaluation process of the two is as follows:
[0021] For confidence level C rw , the bidirectional detector is estimated based on the parameter deviation random framework: for a given model parameter θ at the previous moment t-1 , assuming that the current model parameters θ t The variance has increased in is the variance of the random walk step length;
[0022] For confidence level C cp The bidirectional detector uses a t Hypothesis test method is used to calculate: First, combine x t Gaussian process and observed variables to obtain x t The estimated value of is then evaluated through hypothesis testing. cp .
[0023] Furthermore, the signal optimizer determines the strength of the active control signal, specifically: the signal optimizer first determines the signal strength u AThe constraints are then taken into account, and the detection improvement effect, cost and risk of the active control signal are quantified through the objective function. Finally, the control signal generation is described as a multi-objective optimization problem, and the constraint solver is used to obtain the optimal signal strength.
[0024] Furthermore, the signal strength u A The constraints include dynamic constraints and safety constraints, wherein the dynamic constraints are a time-invariant linear system LS based on the nominal model, as shown in the following formula:
[0025] y t+1 =LS(θ l ,u t , α t )
[0026] In the formula, θ l Indicates the value of a parameter at a specific moment;
[0027] The safety constraint is a safety range of signal strength, as shown in the following formula:
[0028]
[0029] In the formula, u and They represent the lower and upper bounds of the signal strength controlled by the adaptive software system, y and They represent the lower and upper bounds of the system output corresponding to the safe operation of the adaptive software system;
[0030] The signal optimizer uses the effect function Fit e To describe the signal strength u A The quantitative relationship with the improvement of detection accuracy is obtained through the cost function Fit c To describe the action cost of the active control signal, through the hazard function Fit r To describe the degree of danger of behavioral response under active control signal, as follows:
[0031]
[0032] In the formula, Fit(u A ) represents the total revenue function, r e Represents the profit coefficient of improved detection accuracy, r c represents the action cost coefficient, r r represents the profit coefficient of action safety, T represents the duration of the active signal, y and They represent the lower and upper bounds of the system output corresponding to the safe operation of the adaptive software system.
[0033] The beneficial effect of the present invention is that the present invention uses Gaussian processes to evaluate the deviation between the behavior of the adaptive software system and its expected model behavior, thereby improving the accuracy and timeliness of the detection of abnormal behavior of the system. When the observed data is not sufficient to support the judgment of the current system operation status, active signals with optimized strength will be injected by selecting the appropriate time. These signals can stimulate the adaptive software system to produce clear feedback behavior, thereby improving the timeliness and accuracy of the abnormal behavior detection of the system. Compared with traditional abnormal behavior detection technology, the present invention not only finds a good balance between timeliness and accuracy, reduces detection delays, improves detection accuracy, achieves efficient monitoring at a lower cost, and significantly reduces the system abnormality rate, but also effectively improves the overall utility of the system through its carefully designed active signal strategy. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 This is a basic architecture diagram of the adaptive software system of the present invention.
[0035] Figure 2 This is the parameter deviation evaluation process of the fuzzy observation of the present invention.
[0036] Figure 3 This is an example of how existing detection methods cannot accurately and timely detect abnormal system behavior.
[0037] Figure 4 This is an example of how the detection method proposed in the present invention can accurately and timely detect abnormal behavior of the system.
[0038] Figure 5 Comparison of the effectiveness of different detection methods for abnormal behavior detection.
[0039] Figure 6 Comparison of the practicality of different detection methods for abnormal behavior detection. DETAILED DESCRIPTION
[0040] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.
[0041] The adaptive software system abnormal behavior detection technology based on Gaussian process deviation evaluation, referred to as AMM, is designed based on the specificity of the model deviation problem faced by control-SASs. For these systems, model deviation reduces the reliability of the system and requires a mechanism to recover from unsafe execution. Software engineers usually rely on approximate nominal models to design control-SASs. Since software systems are difficult to describe from first principles, it is almost impossible to derive an accurate nominal model for control-SAS design. The gap between the approximate model and the actual behavior of the system leads to system anomalies and triggers unsafe execution.
[0042] However, for traditional control systems, system anomalies actually help optimize the control strategy of the system. This is because these systems follow physical laws that allow accurate nominal models to be identified. Control engineers use accurate nominal models to describe not only the behavior of the system, but also the possible deviations. At runtime, the detection of system anomalies determines which nominal model (or model parameters) is more appropriate for the current system's operating behavior and invokes a monitoring-switching loop to support the optimized control strategy.
[0043] Secondly, the effective monitoring mechanism of AMM is designed according to the specific types of abnormal system behaviors suffered by control-SASs. The abnormal system behaviors faced by control-SASs are divided into different categories. Different techniques are proposed to detect different types of abnormal system behaviors without precise description. When estimating the abnormal behavior of random walk system under non-fuzzy observations, standard Gaussian process functions are used to approximate the abnormal behavior of random walk system. The standard Gaussian process is applicable because in many cases, the normal distribution can represent the non-dramatic uncertainty suffered by adaptive systems. Hypothesis testing is introduced when estimating the abnormal behavior of change point system using non-fuzzy observations. This is because such abnormal system behavior can cause the control-SAS to exhibit significant abnormal behavior, and for black-box software systems, hypothesis testing is very effective in detecting such behavior.
[0044] When estimating abnormal behavior of a system with vague observations, the monitor compromises accuracy to ensure timeliness. Specifically, false alarms reported by the monitor can be tolerated, but missed detections cannot be accepted. This is because active signals can compensate for false alarms by providing non-fuzzy estimates, but cannot recover undetected abnormal behavior of the system. Two pessimistic techniques are used to estimate abnormal behavior of the system under vague observations, which make decisions based on worst-case assumptions. This makes both techniques too crude and impractical for general control systems, but they are still effective for triggering active control signals.
[0045] Finally, the inputs required by the AMM are simplified based on the specific background of the software engineer. In addition to the variables used to develop the control-SAS, the AMM only requires the variance Q of the model parameters θ and the variance R of the system output y as inputs. Moreover, the values of these two variables can be obtained during the system identification process with minimal statistical knowledge. Unlike many control system-oriented methods (such as the window size and threshold of the sliding window method), the AMM does not require engineers to fine-tune the parameters.
[0046] AMM focuses on adaptive software systems, whose architecture is as follows Figure 1As shown, it consists of a controlled system (controlled software) that carries the software business logic, a management system that carries the software adaptive logic, and an external operating environment.
[0047] The fuzzy observation phenomenon targeted by AMM is mainly caused by the rounding operation performed by the controlled system in the adaptive software system when receiving the adjustment instructions (control signals) issued by the management system. This problem will cause the adjustment instructions to be empty for a long time during the operation of the adaptive software system, resulting in the problem that the external detection method cannot accurately and timely detect the abnormal behavior of the system.
[0048] Based on the above content, the present invention proposes an adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation, namely, an AMM system, comprising:
[0049] A parameter deviation stochastic framework that connects the nominal model of an adaptive software system with its observable behavior to explain the impact of ambiguous observations on the detection of model parameter deviations;
[0050] A bidirectional detector that assesses the confidence of an adaptive software system in whether it suffers from model parameter deviations and initiates active control signal injection when the confidence exceeds expectations;
[0051] A signal optimizer is used to determine the strength of the active control signal.
[0052] In this embodiment, it is proposed to use the parameter deviation random framework of formula (1) to describe the abnormal behavior of the system in the adaptive software system caused by the deviation between the nominal model introduced in the system design process and the actual status of the current operating environment of the system.
[0053]
[0054] In the formula, θ t represents the model parameters, the subscript t represents any time t; f t Transition functions representing model parameters; represents normal distribution; Q represents the variance of model parameters; ∈ t is the parameter describing the model θ t The random walk step length of the smooth change process; ω·Θ is the parameter θ describing the model t The gradual change of the mutation process, ω represents the probability of gradual change, Θ represents the amplitude of gradual change; the control parameter u t 、System output y t and environmental input α t is the observed variable, LS represents the time-invariant linear system, is the system output y t The observation error is y tThe variance of the distribution.
[0055] The framework consists of two parts: Gaussian blur-based prior prediction (a) / (b) and Gaussian blur-based posterior regression (c). The former describes the nominal model parameters θ t time series changes, which describes the model parameters θ t and observable variables (including control parameter values u t 、System output value y t and the environmental input value α t ) between them.
[0056] Among them, in the prior prediction based on Gaussian fuzzy, the framework assumes that the value of the controlled software nominal model parameter at any time is θ t is a random variable that follows a normal distribution, θ t The time series change of is reflected in the mean of its normal distribution, while the variance remains unchanged; the variance Q can be identified by the system of the controlled system behavior, based on the sampling of its system behavior, the numerical fitting of the model parameters for the sampled values is performed, and the model parameter value θ with the smallest fitting error is set; the framework is through the transition function f t Model the deviation of the model parameters as changes in their means, specifically, f t The parameter deviation caused by internal factors is described as a random walk process, while the parameter deviation caused by external factors is described as a change point with a low probability of occurrence, representing unexpected changes in the controlled software and operating environment.
[0057] Among them, in the Gaussian fuzzy-based posterior regression, the random framework uses a linear time-invariant model (i.e., γ) to describe the environmental input α based on the original assumption that the environmental input of the adaptive software system changes slowly. t For the system internal state x t The influence of , thus obtaining a time-invariant linear system LS, as shown in formula (2).
[0058]
[0059] In the formula, θ D ,θ C ,θ O is the model parameter θ t The three components of θ D is the time delay component, which describes the influence of the system state at the previous moment on the system state at the next moment during the system change process, θ C is the control component, describing the influence of the external control signal on the system state, θ O It is the observation component, which describes the influence of the system state on the current system output.
[0060] In addition, the stochastic framework uses normal distribution to describe possible measurement errors. Finally, the observed data can be used to perform a posteriori calibration on the stochastic model of model parameters. Specifically, it may take a certain propagation time from the change point of the model parameters to the abnormality of the system output. Based on the stochastic framework, multiple observable variables can be combined to estimate the internal variables of the system (such as model parameters θ or state variables x) to predict potential change points.
[0061] In this embodiment, a bidirectional detector is proposed to evaluate the confidence of the adaptive software system in whether it suffers from model parameter deviation. On the one hand, when a non-zero control signal can be observed, traditional Bayesian inference is used to judge the current θ t Whether it has deviated from its safety interval; on the other hand, when a zero-value control signal affected by fuzzy observation is continuously observed, the results of hypothesis testing and worst-case estimation will be combined to decide whether to trigger active control signal injection.
[0062] When a non-zero control signal can be observed, the parameter deviation estimation of the non-fuzzy observation can be used; specifically, the model parameter value θ is first predicted according to the Gaussian process function t The prior distribution of θ is then used to calculate θ using the observed variables t The posterior distribution of t The posterior distribution of is used to calculate the confidence C that no abnormal behavior has occurred dev .
[0063] Among them, the bidirectional detector uses the Gaussian process function GP(0, k t ) to derive θ t The prior distribution θ t |f t , where k t is the kernel function of the standard Gaussian process; based on the Gaussian process function GP, the bidirectional detector can effectively simulate the random walk process of the model parameter θ and the occurrence of change points.
[0064] Among them, the prior distribution θ of the model parameter values based on the prediction t |f t , we can use the actual observed variable value (i.e. y t ,u t-1 and α t ) to further calculate the posterior distribution θ of the model parameter value t |f t .
[0065] The bidirectional detector calculates the confidence C that no abnormal behavior has occurred in each adaptive cycle. dev , that is, the distribution θ t |f t Falling within its safety range θ] is the probability within .
[0066] When a zero-valued control signal is continuously observed that is affected by the ambiguous observation, the parameter deviation estimation of the ambiguous observation can be used; specifically, the bidirectional detector does not need to directly evaluate the confidence C during the ambiguous observation. dev , but rather evaluates the confidence C of the fuzzy observations and the model-free parameter deviations together vague , that is, the bidirectional detector calculates θ t in u t The confidence level C is evaluated by the probability that the value is still in its safe interval when it remains zero. vague ; Bidirectional detector according to model parameters θ t The transition function f t Calculate the required confidence level C vague Since the transition function f of the random frame t Based on the common model parameter deviation patterns, two cases, random walk and change point, are described. The bidirectional detector uses C rw and C cp They represent the confidence of the occurrence of random walk type parameter deviation and change point type parameter deviation during fuzzy observation. The specific evaluation process is as follows: Figure 2 shown.
[0067] Among them, for the confidence level C rw , the bidirectional detector makes a conservative estimate based on the parameter deviation random framework, that is, assuming that each time θ t Specifically, for a given model parameter value θ at the previous moment t-1 , assuming that the current model parameter value θ t The variance has increased in is the variance of the random walk step length.
[0068] Among them, for the confidence level C cp The bidirectional detector uses a t The hypothesis testing method is used for calculation; Figure 2 The confidence evaluation process of model parameter deviation based on change points during ambiguous observations is shown; the bidirectional detector first combines x t Gaussian process and observed variable value (ie, y t and α t ) to find x t The estimated value of is then evaluated through hypothesis testing. cp .
[0069] In this embodiment, it is proposed to use a signal optimizer to determine the strength u of the active control signal A Specifically, the signal optimizer first determines the signal strength u AConstraints, and then through each objective function (i.e. Fit e 、Fit c and Fit t ) respectively quantify the detection improvement effect, cost and risk of active control signals. Finally, the control signal generation is described as a multi-objective optimization problem, and a constraint solver is used to obtain the optimal signal strength.
[0070] Among them, the signal strength u A The constraints include dynamic constraints and safety constraints. The former gives the quantitative relationship between the active control signal and other input and output signals of the controlled software. According to the constructed parameter deviation random framework, the dynamic constraint considered by the signal optimizer is the time-invariant linear system LS based on the nominal model, as shown in formula (3).
[0071] y t+1 =LS(θ l ,u t , α t ) Formula (3)
[0072] The latter gives a safe range of signal strength values, within which the strength value will not cause abnormal behavior of the system. The safe range of signal strength values can be simply expressed as the active control signal u t and its response output y t+1 The value range of is shown in formula (4). u and They represent the lower and upper bounds of the signal strength controlled by the adaptive software system, y and They represent the lower and upper bounds of the system output corresponding to the safe operation of the adaptive software system.
[0073]
[0074] Among them, the objective function (i.e. Fit e 、Fit c and Fit r ) comprehensively considers the various goals that need to be met by actively controlling the signal strength; specifically, the signal optimizer uses the effect function Fit e To describe the signal strength u A The quantitative relationship with the improvement of detection accuracy is to optimize the timeliness and accuracy of detection, and the effect function Fit e The design is shown in formula (5), r e The benefit coefficient for improving detection accuracy is used to meet the needs and preferences of actual systems or users.
[0075] Fit e =r e ·(y t+1 )2 Formula (5)
[0076] The signal optimizer uses the cost function Fit c To describe the action cost of the active control signal, in order to reduce the damage of the active control signal to the optimization of the adaptive software system. For the controlled software, this action cost is mainly the additional consumption of system configuration resources. The cost function Fit c The design is shown in formula (6), r c represents the action cost coefficient, and T represents the duration of the active signal.
[0077] Fit c =r c ·T·(u t+1 ) 2 Formula (6)
[0078] The signal optimizer uses the hazard function Fit r To describe the degree of danger of the behavioral response under the active control signal, so as to reduce the harm of the active control signal to the safe operation of the adaptive software system, the danger function Fit r The design is shown in formula (7), r r A benefit factor that indicates the safety of an action.
[0079]
[0080] Among them, for the signal strength u A The choice of is implemented as a multi-objective optimization problem, followed by constraint solving; specifically, the signal strength u A By solving the optimization problem shown in formula (8), the total benefit function Fit(u A ) Comprehensively considers the detection improvement effect of active control signals (Function Fit e ), cost (function Fit c ) and risk (Function Fit r ).
[0081]
[0082] Next, taking the Adaptive Body Sensor Network as the target adaptive software system, we further illustrate that the existing detection methods cannot accurately and effectively detect abnormal system behavior. Figure 3 The detection results of traditional detection methods under two system abnormal behaviors are given.
[0083] Figure 3-(a) shows an abnormal system behavior caused by the change in the sensor battery capacity. In this example, the low controllability stage begins when the system generates its first control signal u(132) (indicated by mismatch ②). In this stage, the system utility decreases due to the deviation causing the management system control error, hindering its ability to regulate the system output to the set point (the set point is 0.9 in this example). As the mismatch between u(t) and y(t) increases (from mismatch ② to mismatch ④), the gap between the response y(t) and the set point also increases, eventually causing the system to oscillate after mismatch ⑤.
[0084] To detect such abnormal system behavior, existing methods rely on identifying four groups of mismatched u(t) and y(t) pairs. However, given the uncertainty borne by control-SASs, it is not easy to distinguish these mismatched pairs (such as from mismatches ② to ⑤) from matched pairs (such as mismatch ①). Although some mismatched pairs of u(t) and y(t) (such as mismatch ⑤) may be easy to detect, identifying such obvious mismatched pairs is of little significance because oscillations have actually occurred.
[0085] Discretizing the control signal further exacerbates this situation by rounding fractional values of the control signal to zero. Figure 3 (a) depicts that from time 298 to 590, the management system generates a non-zero control signal u in response to the fluctuation y(t). th (t). However, u(t) received by the management system during the same period remains zero. In other words, if a detection method fails to identify the mismatch④, it has to wait until the last set of mismatched pairs appears before it can detect abnormal system behavior.
[0086] Figure 3 -(b) shows a system behavior anomaly caused by a cyber attack. The attack causes the control signal to temporarily reverse (i.e., θ C becomes -θ C ). This means that in actual operation, the system may encounter unexpected disturbances, such as external attacks or communication errors, which may cause unexpected changes in control instructions. Such short-term signal reversals are difficult to detect immediately and accurately, especially if the monitoring mechanism is not sensitive enough or the data analysis method does not fully consider all potential abnormal patterns. Therefore, it is crucial for designers and maintainers of control systems to develop advanced monitoring and diagnostic tools that can quickly respond and accurately identify such system behavior anomalies to ensure system stability and safety.
[0087] In this example, soon after the managed system receives its first non-zero control signal u(500) (indicated by mismatch ②), the system experiences violent oscillations. This means that existing observation-based methods can neither prevent utility loss nor prevent the system from entering an unsafe state. This rapidly deteriorating condition is caused by u(530) responding strongly to environmental fluctuations at time close to 530. The strength of this signal quickly causes the system to enter an abnormal state with a severely deviated θ(t). Since the environmental input α(t) remains stable and small values of u th (t) is rounded, and u(t) remains zero from time 110 to 530. If the control signal is not discretized, even a small control signal may cause the system output to reverse, and the existing methods can easily detect this mismatch.
[0088] Figure 3 It is shown that without active signal support, fuzzy observations make it impossible to detect model deviations before the system enters oscillation, because all observed y(t) and u(t) are consistent. However, after the proposed AMM detection method introduces its active signal, Figure 4 As shown in Figure 2, the mismatch between active signals u(211) and y(212) (indicated by mismatch ②) enables it to detect the introduced deviation in a timely and accurate manner. This shows that the active signal strategy effectively overcomes the challenges brought by fuzzy observations and warns of model deviations in advance, thereby enhancing the system's self-adaptation and fault prevention capabilities.
[0089] Taking three adaptive software systems, Adaptive Body Sensor Network (BSN), Smart Water Treatment System (SWaT) and Rice University Bidding System (RUBiS) as the objects, the proposed AMM method is compared with the Sliding-Window (SWD) method, Auto-Regressive Moving Average (ARMA) method, Parameter Estimation (PED) method and Model-guided Deviation Detector (MoD2) method in detecting abnormal behavior of the target system. The experimental results are shown in Figure 2. Figure 5 shown.
[0090] Figure 5First, the effectiveness of AMM and baseline methods in detecting model deviations is compared. In terms of detection timeliness, AMM has the smallest average delay for BSN and SWaT systems (93.8% smaller on average for BSN and 25.9% smaller on average for SWaT). However, for RUBiS, AMM achieves a higher mean time to detection (MTD) than SWD and ARMA. This is because AMM requires a closed-loop identification procedure when determining its coefficient of variation, namely the variance Q of θ(t) and the variance R of y(t). For RUBiS, this identification uses only two workloads. The lack of workloads hinders sampling execution and obtaining accurate identification values for Q and R. This in turn affects the timeliness of AMM. It is worth noting that the other two estimation methods (PED and MoD2) also require a closed-loop identification process, which is why they have significantly higher MTDs than SWD and ARMA.
[0091] In terms of accuracy, AMM achieves satisfactory false negative (FN) and false positive (FP) rates when detecting abnormal behavior of the system. On average, AMM achieves an FN rate of 7.5% and its comparable FP rate is 1.67%. Except for RUBiS, where AMM has a slightly higher FN rate than PED (by 2.5%) and a slightly higher FP rate than MoD2 (by 0.6%), its FN and FP rates are the lowest compared to other methods. There is a considerable difference in the performance of AMM between BSN / SWaT and RUBiS. Again, the imprecision of AMM coefficient identification contributes to this gap. However, in contrast to the timeliness results, AMM's identification accuracy is significantly better than SWD and ARMA. This reflects the challenge of abnormal behavior of adaptive software systems, which is to strike a balance between timeliness and accuracy. By sacrificing accuracy, detection latency can be significantly reduced, but this also leads to reporting a large number of false positive results (54.4% FP rate for SWD and 68.1% FP rate for ARMA). AMM attempts to find a reasonable balance between the two to minimize detection delays while ensuring high accuracy.
[0092] Among the five methods compared, AMM and MoD2 achieved better performance. In some cases, the difference between them is small (for example, in the case of the subject RU BiS). Both AMM and MoD2 consider the characteristics of control-SASs rather than traditional control systems, which makes them have higher detection accuracy. The main difference between AMM and MoD2 is their timeliness, with AMM reducing detection delay (34.5% on average) by using active signals. Therefore, if the system under test rarely suffers from ambiguous observations (for example, a Web service system like RUBiS), using AMM or MoD2 may not make much difference. If the system under test often suffers from ambiguous observations (for example, cyber-physical systems like BSN and SWaT), then AMM will be a better choice.
[0093] Taking three adaptive software systems, Adaptive Body Sensor Network (BSN), Smart Water Treatment System (SWaT) and Rice University Bidding System (RUBiS) as the objects, the proposed AMM method is compared with the Sliding-Window (SWD) method, Auto-Regressive Moving Average (ARMA) method, Parameter Estimation (PED) method and Model-guided Deviation Detector (MoD2) method in detecting abnormal behavior of the target system. The experimental results are shown in Figure 2. Figure 6 shown.
[0094] Figure 6The abnormality rate and utility ratio of the tested system under different supervision mechanisms are shown. First, we focus on the abnormality rate. The results show that compared with the original unsupervised system, the AMM-based supervision mechanism greatly reduces the abnormality rate of the tested system (an average reduction of 19.3%). In addition, compared with other supervision mechanisms, the AMM mechanism shows better results, with the lowest average abnormality rate of only 2.4%. This result reflects that AMM's ability to strike a balance between timeliness and accuracy exceeds all baseline methods. As for the utility ratio, AMM also performs well. It not only successfully reduces operational anomalies caused by abnormal system behavior, but also significantly improves the overall utility of the system. Through the timely intervention of active signals, AMM helps the system recover faster from the impact of abnormal system behavior, reduces resource waste and performance degradation, and thus improves the utility ratio. This improvement means that under the supervision of AMM, the system can operate more efficiently and stably, ensuring service quality and user experience, and also verifies the effectiveness and practicality of AMM in maintaining the healthy operation of the system.
[0095] The reduced anomaly rate indicates that the supervision mechanism is effective in avoiding problems, while the utility shows the benefits it brings to the system. The results show that AMM improves the utility of control-SASs by 88.2% (range 30.8% to 145.9%) on average compared to the original system without supervision mechanism. In all cases, AMM achieved the highest utility, which provides some insights into the impact of AMM's active control signals on the execution of the tested system. The high utility achieved by AMM shows that its active signals have minimal side effects. In other words, the utility gains brought by these signals far exceed the potential utility losses that may be caused by activating them. In summary, AMM not only finds a good balance between timeliness and accuracy and significantly reduces the system anomaly rate, but also effectively improves the overall utility of the system through its carefully designed active signal strategy, confirming its important value in improving the robustness and performance of control-SASs. This is not only an important contribution to the existing field of adaptive control, but also provides new ideas and practical paths for the self-management and optimization of future intelligent systems.
[0096] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. Adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation, characterized by: include: A parameter deviation stochastic framework to connect the nominal model and observable behavior of adaptive software systems and to explain the impact of ambiguous observations on the detection of model parameter deviations; A bidirectional detector that assesses the confidence of the adaptive software system in whether it suffers from model parameter deviations and initiates active control signal injection when the confidence exceeds expectations; Signal optimizer, used to determine the strength of the active control signal.
2. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 1, characterized in that: The parameter deviation random framework consists of a priori prediction based on Gaussian fuzzy and a posterior regression based on Gaussian fuzzy; the priori prediction describes the temporal changes of model parameters, and the posterior regression describes the dependency between model parameters and observed variables.
3. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 2, characterized in that: The parameter deviation random framework is expressed as follows: Where, formulas a and b represent the prior prediction based on Gaussian blur, and formula c represents the posterior regression based on Gaussian blur; θ t Represents the model parameter, which is a random variable that obeys the normal distribution. The subscript, represents any, time; f t The transition function of the model parameters is used to describe the deviation of the model parameters caused by internal factors as a random walk process, and the deviation of the model parameters caused by external factors as a change point with low probability. represents normal distribution; Q represents the variance of model parameters; ∈ t is the parameter describing the model θ t The random walk step length of the smooth change process; ω·Θ is the parameter θ describing the model t The gradual change of the mutation process, ω represents the probability of gradual change, Θ represents the amplitude of gradual change; the control parameter u t 、System output y t and environmental input α t is the observed variable, LS represents the time-invariant linear system, is the system output y t The observation error.
4. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 3, characterized in that: In the Gaussian blur-based posterior regression, a linear time-invariant model γ is used to describe the environmental input α t For the system state x t The influence of , thus obtaining a time-invariant linear system LS as follows: In the formula, θ D ,θ C ,θ o They are the model parameters θ t The delay component, control component and observation component of 5. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 3, characterized in that: The bidirectional detector evaluates the confidence of the adaptive software system in whether it suffers from model parameter deviation, specifically: on the one hand, when a non-zero control signal can be observed, traditional Bayesian inference is used to determine the current model parameter θ t Whether it deviates from its safety interval; on the other hand, when zero-value control signals affected by fuzzy observations are continuously observed, the results of hypothesis testing and worst-case estimation are combined to decide whether to trigger active control signal injection.
6. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 5, characterized in that: When a non-zero control signal can be observed, the abnormal behavior detection and evaluation using non-fuzzy observation is specifically as follows: first, the model parameter θ is predicted according to the Gaussian process function. t The prior distribution θ t |f t , and then use the observed variables to calculate the model parameters θ t The posterior distribution θ t |f t , and finally based on the model parameters θ t The posterior distribution of is used to calculate the confidence C that no abnormal behavior has occurred dev , the confidence level C dev is the posterior distribution θ t |f t Falling into its safety zone Probability within.
7. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 5, characterized in that: When the zero-value control signal affected by the fuzzy observation is continuously observed, the parameter deviation of the fuzzy observation is evaluated, specifically: the bidirectional detector calculates the model parameter θ during the fuzzy observation t In the control parameter u t The confidence level C is evaluated by the probability that the value is still in its safe interval when it remains zero. vague ; Bidirectional detector according to the model parameters θ t The transition function f t Calculate the required confidence level C vague .
8. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 7, characterized in that: The bidirectional detector uses C rw and C cp They represent the confidence of the occurrence of random walk type parameter deviation and change point type parameter deviation during fuzzy observation, and the evaluation process of the two is as follows: For confidence level C rw , the bidirectional detector is estimated based on the parameter deviation random framework: for a given model parameter θ at the previous moment t-1 , assuming that the current model parameters θ t The variance has increased in is the variance of the random walk step length; For confidence level C cp The bidirectional detector uses a t Hypothesis test method is used to calculate: First, combine x t Gaussian process and observed variables to obtain x t The estimated value of is then evaluated through hypothesis testing. cp .
9. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation as claimed in claim 1, characterized in that: The signal optimizer determines the strength of the active control signal, specifically: the signal optimizer first determines the signal strength u A The constraints are then taken into account, and the detection improvement effect, cost and risk of the active control signal are quantified through the objective function. Finally, the control signal generation is described as a multi-objective optimization problem, and the constraint solver is used to obtain the optimal signal strength.
10. The adaptive software system abnormal behavior detection system based on Gaussian process deviation evaluation according to claim 9, characterized in that: The signal strength u A The constraints include dynamic constraints and safety constraints, wherein the dynamic constraints are a time-invariant linear system LS based on the nominal model, as shown in the following formula: y t+1 =LS(θ l ,u t ,α t ) In the formula, θ l Indicates the value of a parameter at a specific moment; The safety constraint is a safety range of signal strength, as shown in the following formula: In the formula, u and They represent the lower and upper bounds of the signal strength controlled by the adaptive software system, y and They represent the lower and upper bounds of the system output corresponding to the safe operation of the adaptive software system; The signal optimizer uses the effect function Fit e To describe the signal strength u A The quantitative relationship with the improvement of detection accuracy is obtained through the cost function Fit c To describe the action cost of the active control signal, through the hazard function Fit r To describe the degree of danger of behavioral response under active control signal, as follows: In the formula, Fit(u A ) represents the total revenue function, r e Represents the profit coefficient of improved detection accuracy, r c represents the action cost coefficient, r r represents the profit coefficient of action safety, T represents the duration of the active signal, y and They represent the lower and upper bounds of the system output corresponding to the safe operation of the adaptive software system.
