Zero-Sacrifice Persistent-Robustness Adversarial Defense for Pre-Trained Encoders

Zhuxin Lei, Ziyuan Yang, Yi Zhang · 2026

Read Zero-Sacrifice Persistent-Robustness Adversarial Defense for Pre-Trained Encoders by Zhuxin Lei, Ziyuan Yang, Yi 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-2602-11204.txt.

Description

Public pretrained encoders can be fooled by attacks built without knowing downstream tasks. Zhuxin Lei and two co-authors propose ZePAD, a defence that needs only one tuning to protect many downstream tasks while preserving benign results. It pairs two branches: one strengthens adversarial resistance, the other preserves benign memory.

Two adversarially tuned encoders power the enhancement branch, while the memory branch trains on local data. Surprisingly, ZePAD can also detect downstream-agnostic attacks from branch confidence alone, with no separate finding training. Tests across many self-supervised methods and data sets validate the design, with gains in both benign and adversarial results in some cases. The paper claims a zero-sacrifice defence that persists across tasks.

Catalog record

CREATORZhuxin Lei, Ziyuan Yang, Yi Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-11204

Browse the full library · About the Tower of Babel