A new paper published on arXiv details how visual backdoors can propagate through On-Policy Distillation (OPD) in vision-language models (VLMs). Contrary to initial assumptions, OPD does not inherently resist backdoor propagation, as backdoors can emerge rapidly once a critical poisoning threshold is met. The propagation is driven by the accumulation of anomalous probability mass from the teacher model, rather than explicit malicious responses. Researchers have developed a novel trigger design that facilitates visual backdoor propagation without requiring explicit trigger injection, highlighting potential security risks in OPD. AI
IMPACT Highlights a subtle security risk in vision-language model training that could be exploited.
RANK_REASON Academic paper detailing a security vulnerability in AI model training. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- Influence Flower
- On-Policy Distillation
- ScienceCast
- vision-language model
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →