Evidence · DiggingBeagle record
DPOison-VL: clean image poisoning of vision-language preference tuning
October 13 preprint abstract reports imperceptible modifications to training images that redirect preferences while leaving preference labels and text semantically intact. Original paper direct retrieval failed; indexed abstract and critical third-party reading used.
- Published
- Oct 13, 2026
- Publisher
- Chengtao Jian and Kai Yang
- Source role
- primary disclosure
Evidence record
October 13 preprint abstract reports imperceptible modifications to training images that redirect preferences while leaving preference labels and text semantically intact. Original paper direct retrieval failed; indexed abstract and critical third-party reading used.
Claim-level citations (1)
- supportsDPOison-VL redirects preference tuning with altered training images: The authors report attack success up to 98 percent with 2 percent modified preference-training images on tested vision-language settings.
Abstract: extensive experiments and poison budget