CNRF defends vision-language models with 86 to 89 percent success
It checks how much data supports a risk score, and then it handles only selected requests.
Researchers show CNRF, a safety method for vision-language models. It compares pairs of safe and dangerous input in the same context. On 3 models and 5 jailbreak families, it got 86.06 to 89.30 percent macro defense success rate.
Sources
Posted
