Don't Shift the Trigger: Robust Gradient Ascent for Backdoor Unlearning
Xingyi Zhao, Tian Xie, Xiaojun Qi, Depeng Xu, Shuhan Yuan
Abstract
Backdoor attacks pose a significant threat to machine learning models, allowing adversaries to implant hidden triggers that alter model behavior when activated. Although gradient ascent (GA)-based unlearning has been proposed as an efficient backdoor removal approach, we identify a critical yet overlooked issue: vanilla GA does not eliminate the trigger but shifts its impact to different classes, a phenomenon we call trigger shifting. To address this, we propose Robust Gradient Ascent (RGA), which introduces a dynamic penalty mechanism to regulate GA's strength and prevent excessive unlearning. Our experiments show that RGA effectively removes backdoors while preserving model utility, offering a more reliable defense against backdoor attacks.
BibTeX
@inproceedings{
zhao2026dont,
title={Don't Shift the Trigger: Robust Gradient Ascent for Backdoor Unlearning},
author={Xingyi Zhao and Tian Xie and Xiaojun Qi and Depeng Xu and Shuhan Yuan},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=voqtsqYS6j}
}