Fooling Contrastive Language-Image Pre-Trained Models with CLIPMasterPrints.

Models leveraging both visual and textual data such as Contrastive Language-Image Pre-training (CLIP), are the backbone of many recent advances in artificial intelligence. In this work, we show that despite their versatility, such models are vulnerable to what we refer to as fooling master images. Fooling master images are capable of maximizing the confidence score of a CLIP model for a significant number of widely varying prompts, while being either unrecognizable or unrelated to the attacked prompts for humans. We demonstrate how fooling master images can be mined using stochastic gradient descent, projected gradient descent, or gradient-free optimisation. Contrary to many common adversarial attacks, the gradient-free optimisation approach allows us to mine fooling examples even when the weights of the model are not accessible. We investigate the properties of the mined fooling master images, and find that images trained on a small number of image captions potentially generalize to a much larger number of semantically related captions. Finally, we evaluate possible mitigation strategies and find that vulnerability to fooling master examples appears to be closely related to a modality gap in contrastive pre-trained multi-modal networks.