Low Rank Adaptation for Adversarial Perturbation

Han Liu, Shanghao Shi, Yevgeniy Vorobeychik, Chongjie Zhang, Ning Zhang · 2026

Read Low Rank Adaptation for Adversarial Perturbation by Han Liu, Shanghao Shi, Yevgeniy Vorobeychik, Chongjie Zhang, Ning Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-27487.txt.

Description

Liu and four colleagues ask whether adversarial perturbations, like model updates, live in a low-dimensional space, and they show through theory and experiments that they do. Their method uses a reference model plus auxiliary data to steer gradient projection into a low-dimensional subspace, then confines black-box perturbation search to that subspace. Across attack methods, model architectures, datasets, and threat models, their low-rank attacks improve over conventional methods in efficiency and effectiveness. The low-rank insight opens opportunities for both stronger attacks and better defenses.

Catalog record

CREATORHan Liu, Shanghao Shi, Yevgeniy Vorobeychik, Chongjie Zhang, Ning Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75700
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.27487

Browse the full library · About the Tower of Babel