Robust Contrastive Language-Image Pre-training against Data Poisoning and Backdoor Attacks

Large-scale vision-language models such as CLIP (Radford et al., 2021) and ALIGN (Jia et al., 2021) are trained using a multimodal contrastive