flâneur

Adversarial examples for the OpenAI CLIP in its zero-shot classification regime and their semantic generalization | Stanislav Fort

stanislavfort.github.io · 1,443 words · saved by 1 readers

It turns out that adversarial examples are very easy to find (<100 gradient steps typically) for the OpenAI CLIP model in the zero-shot classification regime. Those adversarial examples generalize to semantically related text descriptions of the adversarial class.

It turns out that adversarial examples are very easy to find (<100 gradient steps typically) for the OpenAI CLIP model in the zero-shot classification regime. Those adversarial examples generalize to semantically related text descriptions of the adversarial class. Stanislav Fort ( Twitter and GitHub ) TL;DR: Adversarial examples are very easy to find for the OpenAI CLIP model in its zero-shot classification regime. Those adversarial examples generalize surprisingly well to semantically-related descriptions of the adversarial class. I wrote a Google Colab where you can try it for yourself. 1. I

related reading