Replies: 3 comments 5 replies
-
|
GitHub Discussions are available 🎉 |
Beta Was this translation helpful? Give feedback.
-
|
In beta for the moment 😅 |
Beta Was this translation helpful? Give feedback.
-
|
Very interesting paper! Is there a possibility to attack the pretrained language model itself? It might be a stronger attack. |
Beta Was this translation helpful? Give feedback.
-
|
How do you mean? My understanding is that they do start with a pretrained model and simply "fine-tune" it with the RIPPLe objective. |
Beta Was this translation helpful? Give feedback.
-
|
just for clarification: what percentage of the corpus the authors poisoned in that paper? |
Beta Was this translation helpful? Give feedback.
-
|
Yeah, they do 50%. What do you mean by easy to recognize? The dataset they poison is never "released", but rather used to poison the pre-trained model weights which are released. Are you saying it would be easy to recognize just from the model weights that half of the data it was trained on were poisoned? |
Beta Was this translation helpful? Give feedback.
-
|
Ok, that make sense. |
Beta Was this translation helpful? Give feedback.
-
|
Np! 😃 |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hi everyone, this week I wrote up a quick discussion on a great paper from Kurita et al.'s on how pre-trained models can be "poisoned" to exhibit nefarious behavior that persist even after fine-tuning on downstream tasks. Below are a few general discussion questions I'd love to get your input on, but feel free to also bring up anything that's interesting to you!
Discussion Questions
Beta Was this translation helpful? Give feedback.
All reactions