Reformulation is All You Need: Addressing Malicious Text Features in DNNs
Yi Jiang, Oubo Ma, Yong Yang, Tong Zhang, Shouling Ji · 2025
Read Reformulation is All You Need: Addressing Malicious Text Features in DNNs by Yi Jiang, Oubo Ma, Yong Yang, Tong Zhang, Shouling Ji (2025) 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-2502-00652.txt.
Description
Yi Jiang, Oubo Ma, Yong Yang, Tong Zhang, and Shouling Ji defend in a 2025 paper text models against hidden malicious features in inputs. Attackers exploit subtle wording traits to fool models or plant backdoors in them. Model side defenses grow costly with scale, while input side ones chase single attack types.
Both attack kinds trace to one root, the encoding step that overweights faint text traits. Their unified defense rewrites inputs to strip such traits while keeping meaning intact. The rewrite modules adapt across attack styles rather than targeting one scheme. Broad trials show the frame beating sample side baselines across many malicious features. Semantic integrity of clean text is a stated guardrail of the design.
Catalog record
| CREATOR | Yi Jiang, Oubo Ma, Yong Yang, Tong Zhang, Shouling Ji |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-00652