This research introduces a Meticulous Adversarial Attack (MAA) method that effectively exposes security vulnerabilities across different vision-language pre-trained models with unprecedented transferability.

• Creates highly transferable adversarial attacks against vision-language models
• Focuses on meticulously selecting shared, region-agnostic features across models
• Demonstrates significantly higher success rates in attacking multiple models with a single adversarial example
• Reveals critical security gaps in current vision-language AI systems

These findings highlight the urgent need for more robust security measures in multimodal AI systems, as attackers could potentially develop universal attacks that compromise multiple vision-language models simultaneously.

MAA: Meticulous Adversarial Attack against Vision-Language Pre-trained Models