New Attack Framework Fools White-Box Explainable AI Auditors

Post-hoc model explainers like LIME, SHAP, and Integrated Gradients are widely used to audit AI models in high-stakes domains such as finance and healthcare, ensuring transparency and trustworthiness. However, the security of these explainability pipelines has been underexplored, with previous adversarial attacks often relying on detectable out-of-distribution (OOD) scaffolding. These black-box attacks could be neutralized by identifying their anomalous perturbation footprints. This research reveals a critical vulnerability by presenting a more sophisticated white-box evasion attack framework. This framework employs a continuous-embedding dual-penalty mechanism to directly penalize trigger feature gradients during model training on in-distribution data. By embedding the evasion logic directly into the model's parameters, the attack generates smooth, in-distribution predictions that leave no detectable anomaly footprint. Empirical evaluations across four benchmark tabular datasets confirm that this method effectively reduces target feature attribution to near-zero, achieving high attack success rates and fundamentally bypassing existing conditional anomaly detection defenses.