Erase the Injected Region: Multimodal Prompt-Injection Defense with On-Policy Self-Distillation
ERASE, an on-policy self-distillation method that constructs its reference by erasing the injected region in place, is proposed, an on-policy self-distillation method that constructs its reference by erasing the injected region in place.
Shu-Lian Hong, Jia-Xin Ma, Xi-Di Cai et al.
· 0 citations