Plato’s Form: Toward Backdoor Defense-as-a-Service for LLMs with Prototype Representations

Chen Chen, Yuchen Sun, Jiaxin Gao, Yanwen Jia, Xueluan Gong, Qian Wang, Kwok-Yan Lam · 2026

Read Plato’s Form: Toward Backdoor Defense-as-a-Service for LLMs with Prototype Representations by Chen Chen, Yuchen Sun, Jiaxin Gao, Yanwen Jia, Xueluan Gong, Qian Wang, Kwok-Yan Lam (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-2602-06887.txt.

Description

Chen Chen and six co-authors clean hidden backdoors out of language models so the models can be offered as a service. A backdoor is a secret trigger that makes a model misbehave on cue, and old defenses demand things a service rarely has: clean task data, knowledge of the trigger, or task details. Their tool, PROTOPURIFY, works by editing model weights under few assumptions.

It first pools backdoor vectors from pairs of clean and tainted models, groups them into candidate patterns, and picks the pattern that best fits the target model. It then finds the layer where that pattern bites and damps the matching parts there. In tests across models and tasks, it beat six rival defenses against six attack kinds, cutting attack success below 10 percent, and to 1.6 percent in the best case, with under a 3 percent loss on clean tasks.

Catalog record

CREATORChen Chen, Yuchen Sun, Jiaxin Gao, Yanwen Jia, Xueluan Gong, Qian Wang, Kwok-Yan Lam
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE84 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-06887

Browse the full library · About the Tower of Babel