What percentage of training data was poisoned in the paper experiment?

BigML-CS-UCLA / RoCLIP

Robust Contrastive Language-Image Pretraining against Data Poisoning and Backdoor Attacks

11 stars 2 forks source link

What percentage of training data was poisoned in the paper experiment? #3

Open uguess1 opened 9 months ago

uguess1 commented 9 months ago

Hi there, could you please elaborate on how many samples were poisoned by the 1M sized training dataset used to train RoCLIP in the paper? I see that your code seems to have only poisoned a 10k sized training dataset.