This research demonstrates how CLIP models can be leveraged to create effective adversarial attacks against AI systems without any prior knowledge of the target model or training data.

• Introduces MF-CLIP, a novel approach utilizing multiple CLIP models as surrogates to generate transferable adversarial examples
• Achieves higher attack success rates than traditional methods in no-box attack scenarios
• Reveals critical security vulnerabilities in widely deployed vision systems, even when attackers have minimal information
• Demonstrates that diverse CLIP model ensembles significantly enhance attack transferability

Why it matters: This work exposes serious security implications for AI deployments in high-stakes environments, showing that sophisticated attacks can succeed even with severe knowledge constraints, necessitating more robust defense mechanisms.

MF-CLIP: Leveraging CLIP as Surrogate Models for No-box Adversarial Attacks