Dynamic Model Editing to Rectify Unreliable Behavior in Neural Networks

ICLR 2025 Conference Submission1227 Authors

Keywords: model vulnerability, model editing, feature attribution

TL;DR: A dynamic model editing technique is proposed for correcting the model's misbehavior.

Abstract: The performance of neural network models deteriorates due to their unreliable behavior on corrupted input samples and spurious data features. Owing to their opaque nature, rectifying models to address this problem often necessitates arduous data cleaning and model retraining, resulting in huge computational and manual overhead. This motivates the development of efficient methods for rectifying models. In this work, we propose leveraging rank-one model editing to correct model's unreliable behavior on corrupt or spurious inputs and align it with that on clean samples. We introduce an attribution-based method for locating the primary layer responsible for the model's misbehavior and integrate this layer localization technique into a dynamic model editing approach, enabling dynamic adjustment of the model behavior during the editing process. Through extensive experiments, the proposed method is demonstrated to be effective in correcting model's misbehavior observed for neural Trojans and spurious correlations. Our approach demonstrates remarkable performance by achieving its editing objective with as few as a single cleansed sample, which makes it appealing for practice.

Primary Area: interpretability and explainable AI

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: 1227