Are Your Explanations Reliable? Investigating the Stability of LIME in Explaining Text Classifiers by Marrying XAI and Adversarial Attack

LIME explanations can be unreliable and manipulated: The research demonstrates that LIME, one of the most widely-used tools for explaining AI decisions in critical applications like healthcare and finance, produces unstable explanations that can be significantly altered through small, carefully crafted changes to input text—raising concerns about whether users can trust these explanations.

Adversarial attacks can fool explanation systems: The researchers developed XAIFooler, an algorithm that can perturb text inputs in ways that preserve the original meaning and the model's prediction while completely changing which features LIME identifies as important—demonstrating a fundamental vulnerability in how LIME explains text classifiers.

Semantic preservation masks manipulation: XAIFooler's ability to maintain text meaning and original predictions while changing explanations is particularly concerning because it means the manipulation could go undetected by human reviewers who might assume that similar inputs should have similar explanations.

Text-based explanations face unique challenges: Unlike other data types, text has special constraints that make explanation stability harder to achieve and easier to exploit, suggesting that XAI tools designed for text classifiers may require different safeguards than those used for other domains.

Current XAI tools may need validation before deployment: Given that LIME's explanations can be unreliably manipulated while maintaining semantic validity, organizations using LIME in high-stakes decisions should implement additional verification methods rather than relying solely on these explanations for model validation or debugging.