Generating transferable attacks across large vision-language models using adversarial deformation learning

Key Points

• Transferable adversarial attacks were successfully generated in multiple vision-language models, enhancing model vulnerability.
• The study achieved a notable 85% effectiveness in deceiving target models with generated adversarial examples.
• Analysis of various vision-language model architectures revealed consistent weaknesses under adversarial deformation learning.
• These findings may indicate a critical need for improved robustness strategies in AI systems managing multimodal inputs.