Fast and Lightweight Backdoor Detection via Head Random Probing
Yinbo Yu, Xueyu Yin, Jing Fang, Chunwei Tian, Qi Zhu, Jiajia Liu, and, Daoqiang Zhang · 2026
Read Fast and Lightweight Backdoor Detection via Head Random Probing by Yinbo Yu, Xueyu Yin, Jing Fang, Chunwei Tian, Qi Zhu, Jiajia Liu, and, Daoqiang 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-2605-18908.txt.
Description
Yinbo Yu and six colleagues offer a data-free backdoor detector probing only a model's prediction head with random latent inputs. Backdoored models concentrate responses on the target class under such probes, a unified signature avoiding trigger reconstruction. HTell builds architecture-aware probes, feeds them into the head, and reads class response statistics. On a benchmark of over six thousand backdoored and seven hundred clean models it reaches 99.03 percent true positives at 2.11 percent false positives. It runs over thirty thousand times faster than gradient-based rivals.
Catalog record
| CREATOR | Yinbo Yu, Xueyu Yin, Jing Fang, Chunwei Tian, Qi Zhu, Jiajia Liu, and, Daoqiang Zhang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68409 |
| LICENSE | http://creativecommons.org/publicdomain/zero/1.0/ |
| RIGHTS | CC0 |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.18908