What question did this study set out to answer?

This research aims to explore the stability of various explainable AI methods under conditions where predictions remain unchanged despite perturbations.

April 29, 2026Applied AI LettersOpen Access

Cross‐Method Explanation Stability Under Prediction‐Preserving Perturbations in Explainable AI

View Full Paper

Ask AI

Bookmark

Authors

MHMuhammad Hasnain TGTooba Gul MAMuhammad Fawad Ahmed

Discussion

Member takes

Overview

Randomized trial investigates stability of XAI methods under prediction-preserving perturbations, highlighting resilience variations.

Key Points

This research aims to explore the stability of various explainable AI methods under conditions where predictions remain unchanged despite perturbations.
Applied four post hoc explanation methods: Integrated Gradients, SHAP, Vanilla Gradients, and Grad-CAM.
Conducted experiments with a controlled perturbation magnitude of 0.01 maintaining consistent predicted classes.
Used a cosine-based similarity measure to evaluate the divergence of explanation maps across different methods.
SHAP showed significant variance in explanations (mean = 0.6475), while Grad-CAM demonstrated very low variance (mean = 0.0058).
Vanilla Gradients and Integrated Gradients recorded moderate variance (means = 0.4647 and 0.3319, respectively).
The findings suggest that explanation divergence starts before prediction changes, marking explanations as sensitive indicators of model vulnerability.

Cite This Study

Hasnain et al. (2026) studied this question.

synapsesocial.com/papers/69f19f9cedf4b46824806665 https://doi.org/10.1002/ail2.70030

View Full Paper

Ask AI

Bookmark

April 29, 2026Applied AI LettersOpen Access

Cross‐Method Explanation Stability Under Prediction‐Preserving Perturbations in Explainable AI

View Full Paper

Ask AI

Bookmark

Authors

MHMuhammad Hasnain TGTooba Gul MAMuhammad Fawad Ahmed

Discussion

Member takes

Overview

Randomized trial investigates stability of XAI methods under prediction-preserving perturbations, highlighting resilience variations.

Key Points

This research aims to explore the stability of various explainable AI methods under conditions where predictions remain unchanged despite perturbations.
Applied four post hoc explanation methods: Integrated Gradients, SHAP, Vanilla Gradients, and Grad-CAM.
Conducted experiments with a controlled perturbation magnitude of 0.01 maintaining consistent predicted classes.
Used a cosine-based similarity measure to evaluate the divergence of explanation maps across different methods.
SHAP showed significant variance in explanations (mean = 0.6475), while Grad-CAM demonstrated very low variance (mean = 0.0058).
Vanilla Gradients and Integrated Gradients recorded moderate variance (means = 0.4647 and 0.3319, respectively).
The findings suggest that explanation divergence starts before prediction changes, marking explanations as sensitive indicators of model vulnerability.

Cite This Study

Hasnain et al. (2026) studied this question.

synapsesocial.com/papers/69f19f9cedf4b46824806665 https://doi.org/10.1002/ail2.70030

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

View Full Paper

Ask AI

Bookmark