SLAFCoM: A Study on Loss Functions for Adversarial Finetuning of Contrastive Models

Aayan Yadav, Shree Sighi

Advisor: Prof. Sanjeev Kumar

Docs

1. Read report at docs/report.pdf
2. Find our presentation at docs/presentation.pptx

Setup

pip install -r requirements.txt

Training

 python apgd_train.py --lr=1e-4 --dataset=imagenet --clean_weight 0.1

References

Schlarmann, C., Singh, N. D., Croce, F., & Hein, M. (2024). Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings for Robust Large Vision-Language Models. http://arxiv.org/abs/2402.12336