Closed nonstopfor closed 1 year ago
While you wait, you might consider reading https://arxiv.org/abs/2202.06382?context=cs.CR I found it insightful ://
I see in the readme that "We selected 100 total errors because this corresponds to a ~35% recall for existing attacks, leaving significant room for improvement". Does this mean the scores for existing attacks are about 0.35?
I am also wondering whether the authors are going to release paper regarding the benchmark!
Whoops we never closed this, but the paper has been in the github repo for the last month.
I wonder when the arxiv paper will be released. I hope to see more details of the data construction process and the performance of current methods.