Data Poisoning Attack on Black-Box Neural Machine Translation to Truncate Translation
Lingfang Li, Weijian Hu, Mingxing Luo · 2024
Read Data Poisoning Attack on Black-Box Neural Machine Translation to Truncate Translation by Lingfang Li, Weijian Hu, Mingxing Luo (2024) 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/pmc-11675270.txt.
Description
Lingfang Li, Weijian Hu, and Mingxing Luo describe in 2024 a poisoning attack that makes translation software cut sentences short. Machine translation systems already tend to drop hard words; poisoned training data can make that far worse. The attack slips a trigger into a small share of training pairs. After the trigger, words grow harder for the model, so it discards them and truncates the output.
The attack worked almost every time while barely denting normal translation scores. The team also built a defense that spots triggers by how they warp a sentence's meaning. Their defense found the triggers and eased the harm. The work warns builders to guard training data.
Catalog record
| CREATOR | Lingfang Li, Weijian Hu, Mingxing Luo |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |