This research introduces CLIPure, a novel approach that enhances zero-shot image classifiers' robustness against adversarial attacks without requiring attack-specific training.

• Leverages CLIP's vision-language pre-training for zero-shot classification
• Implements purification in latent space to defend against various attack types
• Achieves superior adversarial robustness compared to existing methods
• Maintains high accuracy on clean images while protecting against attacks

For security professionals, this research offers a significant advancement in building ML systems that can withstand malicious attempts to manipulate image classification results, addressing a critical vulnerability in AI deployment.

CLIPure: Purification in Latent Space via CLIP for Adversarially Robust Zero-Shot Classification