Evaluating Robustness of Landslide Susceptibility Mapping Models Using ZOO Attack and Machine Learning

ABSTRACT Landslide Susceptibility Mapping (LSM) research has traditionally emphasized predictive accuracy, often overlooking model robustness—a critical yet underexplored dimension given the substantial uncertainties in LSM workflows (e.g., inventory quality, sampling strategies, factor selection). To address the lack of standardized robustness evaluation frameworks, this study applies two complementary approaches: (1) an attack‐based empirical evaluation using the Zeroth‐Order Optimization (ZOO) adversarial attack to generate adaptive perturbations, and (2) a formal robustness verification utilizing two established metrics—verified error and average bound. This work represents the first application of adversarial machine learning techniques for robustness evaluation in geohazard susceptibility modeling. We apply the proposed framework to compare two widely used models—Random Forest (RF) and eXtreme Gradient Boosting (XGBoost)—yielding two key insights. First, predictive accuracy and robustness are distinct properties: although RF achieved higher baseline accuracy (AUC: 0.9372 vs. 0.9000), XGBoost demonstrated markedly greater resilience to ZOO‐generated perturbations (AUC reduction: 0.97% vs. 8.31% for RF), challenging the assumption that higher accuracy implies greater robustness. Second, the consistency between the attack‐based empirical evaluation and the formal verification metrics confirms XGBoost's superior robustness, thereby validating the reliability of the applied framework. These findings underscore the necessity of incorporating robustness testing as a core component of LSM model evaluation to enhance reliability and reduce uncertainty in diverse application scenarios. The integrated framework, combining adversarial testing with formal verification metrics, offers a scientifically rigorous, efficient, and objective toolset for future LSM research, model benchmarking, and operational deployment.