AIM: Adversarial Information Masking for Evaluating EEG-DL Interpretations

ICLR 2025 Conference Submission12578 Authors

Keywords: Explainable AI, Post-hoc explanation, EEG, Feature attribution method, Saliency map, in-distribution imputation

TL;DR: A robust framework for quantifying the faithfulness of post-hoc explanations in EEG deep learning models, employing adversarial information masking for effective in-distribution imputation.

Abstract:

We identify significant gaps in the existing frameworks for assessing the faithfulness of post-hoc explanation methods, which are essential for interpreting model behavior. To overcome these challenges, we propose a novel adversarial information masking (AIM) approach that enhances in-distribution information masking techniques. Our study conducts the first quantitative comparison of faithfulness assessment frameworks across different architectures, datasets, and domains, facilitating a comprehensive evaluation of post-hoc explanation methods for deep learning of human electroencephalographic (EEG) data. This work lays a foundation for further developments of reliable applications of explainable artificial intelligence (XAI).
The code and sample data for this work are available at https://anonymous.4open.science/r/EEG-explanation-faithfulness-5C05.

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12578