+
+ Chengzhi Mao* + · + Scott Geng* + · + Junfeng Yang + · + Xin Wang + · + Carl Vondrick
+https://arxiv.org/abs/2212.07016
+ + +Pretrained vision-language foundation models like CLIP have exhibited strong generalization over unseen tasks, yet imperceptible adversarial perturbations can significantly reduce their performance. Given that such large-scale models are becoming a form of infrastructure in practice, understanding and defending their robustness has become an important new problem space. In particular, our recent work demonstrates that existing standard adversarial training techniques suffer from a catch-22 when applied to zero-shot models: without adversarial training, the model is vulnerable to attacks, but with adversarial training, the model loses its zero-shot capabilities. This problem is partially addressed by our introduced text-guided contrastive adversarial training loss, but a gap still remains. To spur further advances in this important space, we propose a defence challenge. ## Zero-Shot Adversarial Robustness Challenge | Defence Method | Submitted By | Accuracy