Apple paper asks: What do your logits know?

苹果新论文发出惊人一问：What do your logits know?

Apple researchers posted an arXiv paper testing whether VLM top-k logits leak image details. Using CLEVR, MSCOCO, and probes, 30–80 logits recover noise, target traits, and some background attributes. The key risk is gray-box APIs exposing top-k log probabilities.

Why it matters: HKR-H/K/R all pass: the Apple paper turns VLM logit outputs into a concrete privacy risk, with CLEVR/MSCOCO probes and a 30–80 logit range. It is strong research, not a same-day platform event, so it stays in 78–84.