Evaluating the Reliability of Post Hoc Explanation Methods under Adversarial Perturbations in High-Stakes Predictive Modeling

Abstract

The integration of deep neural networks into high-stakes decision-making environments, such as clinical diagnostics, financial risk assessment, and criminal justice, has necessitated the development of post hoc explanation methods to ensure transparency and accountability. However, the reliability of these interpretability tools—most notably Local Interpretable Model-agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP)—remains a critical systemic vulnerability when subjected to adversarial perturbations. This research provides a comprehensive evaluation of how adversarial entities can manipulate post hoc explanations to mask underlying biases or systematic errors without altering the primary predictive output of the model. Through a socio-technical lens, we analyze the structural trade-offs between model performance and interpretability robustness, arguing that current explainable artificial intelligence (XAI) frameworks lack the formal guarantees required for deployment in critical infrastructures. Our findings suggest that perturbation-based methods are particularly susceptible to scaffolding attacks that exploit the out-of-distribution characteristics of synthetic data samples used during the explanation process. Furthermore, we discuss the governance and policy implications of these vulnerabilities, emphasizing the need for standardized auditing protocols and robust, integrated transparency mechanisms. The paper concludes by proposing a forward-looking transition toward multi-layered verification and validation frameworks that align technical explainability with institutional accountability and regulatory mandates such as the European Union’s Artificial Intelligence Act.