3 ms·
They don't solve the problem completely, but they address the problem in their publication as follows. The inference step uses a lot of layers and the first and
by ftxbro 4y ago
They don't solve the problem completely, but they address the problem in their publication as follows. The inference step uses a lot of layers and the first and last layers have more 'interpretability' so some of these are the ones that you solve locally on your own computer. If they wanted to prank you, they would have to use some sophisticated thing that probably doesn't exist yet but it could still be possible. Also if it becomes a problem they could do usual methods like run on multiple other computers and use the majority solution.
- sebzim4500 4y ago>If they wanted to prank you, they would have to use some sophisticated thing that probably doesn't exist yet but it could still be possible. Isn't the attack straightforward? i) Take the model, freeze all the weights except the ones you expect to be responsible for ii) Finetune to produce whatever output you are looking for. iii) Profit. Or mainly just annoy people, but it could be funny.
- ftxbro 4y agoOK sure, any time someone says on the internet that something is sophisticated someone will say it's straightforward. So I guess it's a continuum. The thing you describe is more sophisticated than what they would have to do if they had access to the final layer.