Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing

Large vision-language models (LVLMs) derive their capabilities from extensive training on vast corpora of visual and textual data.