A method for testing causality effect robustness based on adversarial perturbation

Abstract

The application discloses a kind of causality effect robustness inspection method based on counter disturbance, it is related to geological disaster risk analysis technical field, including: based on causal inference method obtains the causal factor of geological disaster prediction model, associated factor, and the causality effect index of corresponding factor;Factor disturbance space with physical rationality constraint is constructed, and the value range of factor disturbance amount is defined;Zero-order optimization counter attack method is used to generate factor-level counter disturbance samples in factor disturbance space;The change amount of model output before and after factor applies counter disturbance is calculated with the causality effect change amount, and factor-level counter robustness index is obtained;Causal robustness index is constructed, and the reliability evaluation and classification of causal factor are completed.The application solves the problems that the existing technology lacks robustness verification of causal effect, the causal factor may be affected by counter disturbance and fail, lacks causal robustness evaluation system of factor level and cannot identify pseudo-causality in causal inference result.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence causal inference methods and prediction models and geological disaster risk analysis technology, specifically to a robustness test method for causal effects based on adversarial perturbations. Background Technology

[0002] Geological hazards are characterized by complex proliferative environments, diverse causative factors, and strong coupling of triggering mechanisms. In recent years, geological hazard susceptibility / hazard models based on machine learning and statistical analysis have been widely used for the prediction and risk identification of hazards such as landslides, collapses, and debris flows. These models typically rely on multi-source observation data and use data mining methods to identify key factors influencing hazard occurrence. However, traditional data-driven models mainly reveal the statistical correlation between factors and hazard outcomes, making it difficult to accurately distinguish between causal factors, correlated factors, and spurious factors, resulting in insufficient explanation of hazard mechanisms and limited model reliability.

[0003] To address the confusion between correlation and causality, causal inference methods (such as propensity score matching, causal forests, and dual robust estimation) have been gradually introduced into the field of geological disaster analysis. These methods are used to quantify the causal effects of factors on disaster occurrence and identify causal factors with actual disaster-causing effects. However, existing causal inference methods generally rely on the assumption of data distribution stability and are highly sensitive to factors such as noise disturbances, variable bias, and model fragility. When input factors are disturbed, causal effects may change significantly, leading to unstable or even invalid causal conclusions, making it difficult to meet the reliability requirements of causal interpretation in engineering applications.

[0004] On the other hand, adversarial perturbation techniques (such as zero-order optimization adversarial attacks) are widely used in the robustness research of artificial intelligence models, which construct small perturbations in the input space to assess the sensitivity and vulnerability of the model. However, existing adversarial perturbation methods are mainly used for the security assessment of classification models, and have not yet been used to test the stability of causal inference results, nor have they formed a robustness evaluation system for the causal factor level.

[0005] Therefore, there is an urgent need for a method that can combine causal inference with anti-perturbation techniques to systematically test the stability of the causal effects of causal and correlation factors, identify possible spurious causal relationships or unstable factors in the causal inference results, and thus improve the credibility and engineering applicability of geological disaster susceptibility / hazard models and other causal explanation models. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a robustness testing method for causal effects based on adversarial perturbations. This method solves the problems of existing technologies, such as the lack of robustness verification of causal effects, the possibility that causal factors may fail due to adversarial perturbations, the lack of a factor-level causal robustness evaluation system, and the inability to identify spurious or unstable factors in causal inference results.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a robustness testing method for causal effects based on adversarial perturbations, comprising the following steps: S1: Obtain the causal factors, correlation factors, and corresponding causal effect indicators of the geological disaster prediction model based on the causal inference method; S2: Based on causal factors, related factors, and causal effect indicators of corresponding factors, construct a factor perturbation space with physical rationality constraints, and define the range of values ​​for factor perturbation quantities. S3: Employ a zero-order optimization adversarial attack method to generate factor-level adversarial perturbation samples within the factor perturbation space; S4: Based on the factor-level adversarial perturbation samples, calculate the changes in model output and causal effects before and after the factor is subjected to adversarial perturbation, and obtain the factor-level adversarial robustness index. S5: Based on the robustness index of causal effects and factor-level adversarial forces, construct a causal robustness index to complete the reliability assessment and classification of causal factors.

[0008] Furthermore, the causal inference methods include propensity score matching, dual robust estimation, causal forest, structural equation modeling, and Granger causality test; The causal effect indicators include average treatment effect, conditional average treatment effect, and individual treatment effect.

[0009] Furthermore, the factor perturbation space with physical rationality constraints is:

[0010] in, For the first The perturbation of each factor The set of perturbations to satisfy physical rationality constraints, For norm, This represents the upper limit of the disturbance amplitude.

[0011] Furthermore, the zero-order optimization adversarial attack method is a type of black-box zero-order attack method, and the generated perturbation satisfies:

[0012] in, For the first Black-box gradient estimation of each factor For disaster prediction models, The original input factor vector, For small perturbation step size, For the first Unit vectors in each factor direction; The objective of generating factor-level adversarial perturbation samples is:

[0013] in, Let be the perturbation variable to be solved. Let be the perturbation vector applied to the input factor. Output the change threshold for the model.

[0014] Furthermore, the factor-level adversarial robustness index is:

[0015] in, As a factor-level robustness indicator, For mathematical expectation operators, For indicator functions; The change in the causal effect is:

[0016] in, For the first The change in the causal effect of each factor For the first The causal effect of each factor.

[0017] Furthermore, the causal robustness index is:

[0018] in, This is the causal robustness index.

[0019] Furthermore, the reliability assessment and classification rules for the causal factors include: Disaster-causing factors: and Significant; False causal factors: Significant but ; Correlation factors: Insignificant and ; Unstable factors: ; When causal effect significance level At that time, it was believed Significant; When causal effect significance level At that time, it was believed Not significant; in, The threshold for the causal robustness index. and The robustness index is the grading threshold. The preset threshold for causal effect change. This is the threshold for the significance of causal effects.

[0020] Furthermore, it also includes a multi-factor combined perturbation test step: A combined adversarial perturbation is applied to multiple factors, and the combined adversarial perturbation satisfies:

[0021] in, For the subset of factors participating in the combined perturbation, This represents the total number of factors.

[0022] The beneficial effects of this invention are: The robustness testing method for causal effects based on adversarial perturbation proposed in this invention can effectively overcome the shortcomings of existing causal inference methods in stability verification, and has the following significant advantages: (1) To achieve robust verification of causal inference results and improve the credibility of causal conclusions. Existing causal inference methods (such as PSM, DML, causal forest, etc.) usually assume that the data distribution is stable, but in actual geological disaster scenarios, input factors are often affected by noise, bias, monitoring errors, etc. This invention constructs a factor perturbation space and applies zero-order optimization to counteract the perturbation, which can test the changes of causal effects under perturbation conditions, thereby evaluating the stability and reliability of causal factors and significantly improving the credibility of causal inference results.

[0023] (2) Identify spurious causal factors and unstable factors to avoid erroneous explanations of disaster mechanisms. By calculating the change in causal effect and the robustness index, this invention can identify spurious causal factors with significant causal effects but low robustness and unstable factors with causal effects that are highly sensitive to disturbances. These factors may be misjudged as "important factors" in traditional causal inference, but they can be accurately identified under the robustness test of this invention, thereby avoiding erroneous explanations of disaster mechanisms and unreasonable engineering decisions.

[0024] (3) A causal robustness index is proposed to achieve a comprehensive credibility evaluation of causal factors. The causal robustness index constructed in this invention can simultaneously reflect the intensity of causal effects and stability under resistance to disturbances, thereby achieving a comprehensive credibility evaluation of causal factors. This index can be used for screening key disaster-causing factors, ranking and verifying causal inference results, and enhancing the credibility of model interpretation results, providing a more reliable causal explanation basis for geological disaster prediction models.

[0025] (4) Supports causal robustness testing of black-box models and has a wide range of applications. This invention adopts a zero-order optimization adversarial attack method, which does not rely on model gradient information and can be directly applied to random forests, XGBoost, LightGBM, neural networks and other black-box prediction models. Therefore, it has good versatility and can be applied to various geological disaster susceptibility / hazard models and other prediction models that require causal explanation.

[0026] (5) Supports single-factor and multi-factor combined perturbations, enabling analysis of the robustness of factor interactions. This invention not only supports single-factor perturbations but also multi-factor combined perturbations, enabling the testing of the influence of inter-factor coupling relationships on causal effects, thereby identifying interaction-sensitive factors, combined pseudo-causal factors, and key disaster-causing factors under complex coupling conditions, thus enhancing the depth and explanatory power of causal robustness analysis.

[0027] (6) Improve the engineering applicability and reliability of geological disaster prediction models. Through robustness testing of causal factors, this invention can provide more reliable key factor identification results, support disaster mechanism explanation and monitoring deployment optimization, improve the credibility of geological disaster prediction models in actual engineering, reduce the risk of misjudgment caused by pseudo-causality or unstable factors, and thus significantly enhance the engineering application value of geological disaster risk assessment and early warning systems. Attached Figure Description

[0028] Figure 1 This is a flowchart of a robustness test method for causal effects based on adversarial perturbations.

[0029] Figure 2 A schematic diagram of the zero-order optimization process for generating adversarial perturbations.

[0030] Figure 3 This is a schematic diagram illustrating the calculation of changes before and after a causal effect disturbance.

[0031] Figure 4 This is a schematic diagram illustrating the construction of the causal robustness index. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1As shown, a robustness test method for causal effects based on adversarial perturbations includes the following steps: S1: Obtain the causal factors, correlation factors, and corresponding causal effect indicators of the geological disaster prediction model based on the causal inference method; The causal inference methods include propensity score matching (PSM), dual robust estimation (DML), causal forest, structural equation modeling (SEM), and Granger causality test (time series scenario). The causal effect indicators include average treatment effect (ATE), conditional average treatment effect (CATE), and individual treatment effect (ITE), which are used to identify causal factors, correlation factors, and spurious correlation factors, providing basic input for subsequent robustness tests.

[0034] S2: Based on causal factors, related factors, and causal effect indicators of corresponding factors, construct a factor perturbation space with physical rationality constraints, and define the range of values ​​for factor perturbation quantities. To simulate the changes in input factors under noise, bias, or external disturbances, this invention constructs a factor perturbation space. A perturbation amount is defined for each factor. ,satisfy:

[0035] in, For the first The perturbation of each factor The set of perturbations to satisfy physical rationality constraints, The norm is used to constrain the magnitude of the disturbance. This represents the upper limit of the disturbance amplitude.

[0036] Physical rationality constraints include: Topographic factors such as slope, aspect, and undulation must be kept within a reasonable range; Hydrological factors such as rainfall and pore water pressure must not be negative; Discrete factors such as lithology and stratigraphic type must maintain category validity; Factors such as earthquake PGA and focal depth must meet seismic constraints.

[0037] This step ensures that counter-disturbances do not compromise the physical realism of the geological disaster scenario.

[0038] S3: Employ a zero-order optimization adversarial attack method to generate factor-level adversarial perturbation samples within the factor perturbation space; Zero-order optimization methods include: ZOO Attack, NES Attack, SPSA Attack, Bandit Attack, and other black-box zero-order attack methods.

[0039] The zero-order optimization adversarial attack method is one of the black-box zero-order attack methods, and the generated perturbation satisfies:

[0040] in, For the first Black-box gradient estimation of each factor For disaster prediction models, This is the original input factor vector (containing the original values ​​of all factors). For small perturbation step size, For the first Unit vectors in each factor direction; The objective of generating factor-level adversarial perturbation samples is:

[0041] in, The disturbance variable (optimization variable) to be solved is called the variable to be solved. Let be the perturbation vector applied to the input factor. Output the change threshold for the model.

[0042] This step is used to find the minimum perturbation that causes a significant change in the model output, thereby assessing the model's sensitivity to factor perturbations.

[0043] S4: Based on the factor-level adversarial perturbation samples, calculate the changes in model output and causal effects before and after the factor is subjected to adversarial perturbation, and obtain the factor-level adversarial robustness index. The factor-level robustness index is:

[0044] in, As a factor-level robustness indicator, It is the mathematical expectation operator (averaging over the sample or perturbation distribution). This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The change in the causal effect is:

[0045] in, For the first The change in the causal effect of each factor For the first The causal effect of each factor.

[0046] This step is used to quantify the sensitivity of causal effects to perturbations, reflecting the stability of causal factors.

[0047] S5: Based on the robustness index of causal effects and factor-level adversarial forces, construct a causal robustness index to complete the reliability assessment and classification of causal factors.

[0048] The causal robustness index is:

[0049] in, It is a causal robustness index used to comprehensively assess the credibility of causal factors.

[0050] The reliability assessment and classification rules for the causal factors include: Disaster-causing factors: and Significant; False causal factors: Significant but ; Correlation factors: Insignificant and ; Unstable factors: ; When causal effect significance level At that time, it was believed Significant; When causal effect significance level At that time, it was believed Not significant; in, The threshold for the causal robustness index can be set based on model performance, engineering experience, or statistical quantiles (such as the 75th percentile). and The robustness index grading threshold can be set according to the sample distribution (e.g., 33rd and 66th percentiles). The preset threshold for causal effect variation can be set based on the standard deviation of the causal effect, confidence interval, or engineering experience. This is the threshold for the significance of causal effects, which can be 0.05, 0.01, or set according to engineering requirements.

[0051] To assess the impact of factor interactions on causal effects, this invention applies combined adversarial perturbations to multiple factors to examine the influence of factor coupling relationships on the stability of causal effects. The combined adversarial perturbations satisfy:

[0052] in, For the subset of factors participating in the combined perturbation, This represents the total number of factors.

[0053] This invention is applicable to the causal effect verification of geological hazard susceptibility / hazard models, including landslides, collapses, debris flows, secondary earthquake disasters, and other geological hazard scenarios. It can also be extended to models in medicine, finance, meteorology, and other fields that require causal explanations.

[0054] In one embodiment of the present invention, combined with Figures 2 to 4 The application of the method of the present invention in landslide susceptibility analysis is explained.

[0055] The study area selected a landslide-prone zone in a mountainous region and constructed a 10 m × 10 m raster dataset with a total of 10,000 samples, including 1,200 samples showing landslide occurrences and 8,800 samples showing no landslide occurrences. Nine commonly used landslide influencing factors were selected as input variables, including: slope (°), aspect (classification), elevation (m), curvature, distance from river (m), distance from road (m), maximum annual daily rainfall (mm), lithology (8 classes), and NDVI. A landslide susceptibility prediction model was constructed using XGBoost. The output is the probability of a landslide occurring.

[0056] This embodiment uses two methods, PSM + DML and causal forest, to estimate the causal effects of key factors: 1) Slope (>25°) as a treatment variable ATE≈0.18 (A steeper slope increases the probability of landslides by an average of 18%) 2) The maximum annual daily rainfall (>120 mm) was used as a treatment variable. CATE (soft rock zone) ≈ 0.22 CATE (hard rock zone) ≈ 0.09 3) Distance from the river (<100 m) was used as a treatment variable. ATE≈0.05 (weakly significant) Based on this preliminary identification: Causal factors: slope, rainfall Weak causal / correlation factors: distance from river, distance from road This embodiment sets the perturbation rules as shown in Table 1: Table 1 Disturbance Rules

[0057] upper limit of disturbance amplitude The significance of each factor engineering is determined.

[0058] Figure 2 This illustrates the process of generating zero-order optimization adversarial perturbations. The figure shows how, by applying small positive and negative perturbations to the input factors, the black-box gradient is estimated and the minimum perturbation is searched to make the model output exceed a threshold. The process embodies the general mechanism of zero-order optimization methods such as ZOOAttack, NES Attack, and SPSA Attack.

[0059] This embodiment uses ZOO Attack, with the following settings: step size. Output change threshold .

[0060] For a sample with a high slope (original landslide probability 0.72): Apply disturbance to the slope The output becomes 0.51, and the change is 0.21 > 0.15 → Attack successful; Disturbance to rainfall The output becomes 0.69, and the change is 0.03 < 0.15 → Attack failed.

[0061] Success rate of attacks based on 10,000 samples: Slope: 62% Rainfall: 25% Distance from river: 55% Figure 3 This figure illustrates the changes in causal effects before and after the factor perturbation. Changes, including the original causal effect Causal effects after disturbance Differences This is used to illustrate the calculation method for sensitivity to causal effect perturbations.

[0062] Change in causal effect and factor-level robustness indicators The calculation results are shown in Table 2: Table 2 Calculation Results

[0063] slope The large value indicates that the causal effect is sensitive to disturbances.

[0064] Figure 4 Demonstrates the causal robustness index A schematic diagram illustrating the construction of the diagram. This diagram demonstrates causal effects. With resistance robustness index Combination methods, through A comprehensive credibility evaluation index for causal factors is formed and used for the reliability classification of causal factors.

[0065] Causal robustness index The calculation results are shown in Table 3: Table 3 Causality Robustness Index Calculation results

[0066] According to the classification rules: Rainfall: high( )and Significant ( → Stable disaster-causing factors slope: Significant ( )but Low( → Pseudo-causal factor Distance from the river: Not significant ( )and medium( → General correlation factor This embodiment demonstrates that the present invention can not only identify causal factors, but also further determine their stability and reliability; although slope is a traditionally important factor, it is sensitive to resistance in the model of this region and belongs to an unstable causal factor; rainfall factors have high causal effects and high robustness and are key disaster-driving factors; the present invention can effectively identify pseudo-causal factors and improve the reliability of the explanation of geological disaster mechanisms.

[0067] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the invention.

Claims

1. A robustness testing method for causal effects based on adversarial perturbations, characterized in that, Includes the following steps: S1: Obtain the causal factors, correlation factors, and corresponding causal effect indicators of the geological disaster prediction model based on the causal inference method; S2: Based on causal factors, related factors, and causal effect indicators of corresponding factors, construct a factor perturbation space with physical rationality constraints, and define the range of values ​​for factor perturbation quantities. S3: Employ a zero-order optimization adversarial attack method to generate factor-level adversarial perturbation samples within the factor perturbation space; S4: Based on the factor-level adversarial perturbation samples, calculate the changes in model output and causal effects before and after the factor is subjected to adversarial perturbation, and obtain the factor-level adversarial robustness index. S5: Based on the robustness index of causal effects and factor-level adversarial forces, construct a causal robustness index to complete the reliability assessment and classification of causal factors.

2. The robustness testing method for causal effects based on adversarial perturbation according to claim 1, characterized in that, The causal inference methods include propensity score matching, dual robust estimation, causal forest, structural equation modeling, and Granger causality test; The causal effect indicators include average treatment effect, conditional average treatment effect, and individual treatment effect.

3. The robustness testing method for causal effects based on adversarial perturbation according to claim 1, characterized in that, The factor perturbation space with physical rationality constraints is: in, For the first The perturbation of each factor The set of perturbations that satisfy physical rationality constraints. For norm, This represents the upper limit of the disturbance amplitude.

4. The robustness testing method for causal effects based on adversarial perturbation according to claim 3, characterized in that, The zero-order optimization adversarial attack method is one of the black-box zero-order attack methods, and the generated perturbation satisfies: in, For the first Black-box gradient estimation of each factor For disaster prediction models, The original input factor vector, For small perturbation step size, For the first Unit vectors in each factor direction; The objective of generating factor-level adversarial perturbation samples is: in, Let be the perturbation variable to be solved. Let be the perturbation vector applied to the input factor. Output the change threshold for the model.

5. The robustness testing method for causal effects based on adversarial perturbation according to claim 4, characterized in that, The factor-level robustness index is: in, As a factor-level robustness indicator, For mathematical expectation operators, For indicator functions; The change in the causal effect is: in, For the first The change in the causal effect of each factor For the first The causal effect of each factor.

6. The robustness testing method for causal effects based on adversarial perturbation according to claim 5, characterized in that, The causal robustness index is: in, This is the causal robustness index.

7. The robustness testing method for causal effects based on adversarial perturbation according to claim 6, characterized in that, The reliability assessment and classification rules for the causal factors include: Disaster-causing factors: and Significant; False causal factors: Significant but ; Correlation factors: Insignificant and ; Unstable factors: ; When causal effect significance level At that time, it was believed Significant; When causal effect significance level At that time, it was believed Not significant; in, The threshold for the causal robustness index. and The robustness index is the grading threshold. The preset threshold for causal effect change. This is the threshold for the significance of causal effects.

8. The robustness testing method for causal effects based on adversarial perturbation according to claim 7, characterized in that, It also includes a multi-factor combined perturbation test step: A combined adversarial perturbation is applied to multiple factors, and the combined adversarial perturbation satisfies: in, For the subset of factors participating in the combined perturbation, This represents the total number of factors.

Images