FEAT Faster Greedy Coordinate Gradient algorithm (GCG)

[romanlutz]

Is your feature request related to a problem? Please describe.

There are several papers or blogs claiming to speed up GCG.

• https://arxiv.org/html/2403.01251v2 with repo https://github.com/zhaoyiran924/Probe-Sampling/tree/main
• https://www.haizelabs.com/technology/making-a-sota-adversarial-attack-on-llms-38x-faster

Describe the solution you'd like

Having a faster version of GCG would help a lot because it currently takes at least an hour to generate single suffix.

[riyosha]

Hi, I'm interested in taking up this issue if no one else is working on it!

[0xm00n]

would like to work on this as well. @riyosha would you be open to collaboration?

[romanlutz]

Note: This is potentially very challenging. You would have to have GPUs to run GCG and the alternatives, and even run it several times to show that it works well and is truly faster.

[0xm00n]

yes, i have access to an A100 gpu and also H100s.

[romanlutz]

Great! Feel free to try this out then! One thing I'll say up front: the current implementation is essentially a copy of the original research code with light modifications. There are no tests. For obvious reasons, that is not ideal. Hopefully, for the next iteration of GCG we can improve that. I think knowing that you're not breaking things also makes development easier.

@riyosha feel free to coordinate with @0xm00n depending on whether you have GPUs. There are multiple alternatives so you could each pick one, for example. If you don't have GPUs it may be better to pick a different issue.

[riyosha]

@0xm00n hi! lmk your preferred method to communicate

[0xm00n]

@riyosha i prefer discord if you have. will connect with you on linkedin to send it.

[4B5F5F4B]

I think this nanoGCG()https://github.com/GraySwanAI/nanoGCG has already implemented a faster GCG, in nanoGCG's README.md, you can also see it recite this blog https://www.haizelabs.com/technology/making-a-sota-adversarial-attack-on-llms-38x-faster.