Could you please release the adversarial pre-trained CLIP weights?

[LYMDLUT]

Great job! Could you please release the adversarial pre-trained CLIP weights? This would be more beneficial for subsequent research and progress in the community.

[HashmatShadab]

Kindly provide an update on this

[zw615]

@LYMDLUT @HashmatShadab Hi, thanks for your interest in our work! I have just released ViT-L and ViT-H adversarially pre-trained CLIP model weights at https://github.com/zw615/Double_Visual_Defense. Apologize for the delay. The model weights was under some internal review process, which took a much longer time than expected. Let me know if you have any other questions, and I shall be happy to help.

[HashmatShadab]

@zw615 Thank you for releasing the ViT-L and ViT-H adversarially pre-trained CLIP model weights, even after an extensive internal review process that spanned nearly a year. Due to limited computational resources, I focused on analysis of previously available weights to demonstrate improvements in robustness within Multi-Modal Large Language Models. For more details, you can explore the project page for our technical report, codebase and the Robust MLLM weights.

I appreciate the release of the model weights and your insightful work in Double Visual Defense, which explores a similar direction to my research. It’s great to see efforts in incorporating robust backbones into MLLMs, and I look forward to potential discussions in this space.