Synonymity-Weighted Similarity for Robustness Evaluation of Text-Based Explainable AI
A family of similarity measures for judging adversarial attacks on text-based XAI. Standard ranked-list measures (Jaccard, Kendall, Spearman, RBO) treat every word substitution as equally disruptive; synonymity weighting amends them so a perturbed word and its original are compared via an inner measure of their semantic similarity, using the single-word-substitution mapping between original and perturbed feature-importance explanations. This prevents overestimating attack success and gives a more faithful measure of an XAI method's true robustness.
Quantifying True Robustness: Synonymity-Weighted Similarity for Trustworthy XAI Evaluation
Evaluations of adversarial attacks on text-based Explainable AI (XAI) judge attack success by comparing the ranked feature-importance explanations (e.g. from LIME) before and after perturbing input w…