SceneTAP introduces a novel approach to generate visually natural typographic adversarial attacks that can effectively mislead advanced vision-language models while appearing coherent within real-world environments.

• Creates adversarial text that blends seamlessly into scene contexts
• Demonstrates serious vulnerabilities in state-of-the-art vision-language models
• Achieves high attack success rates while maintaining visual naturalness
• Highlights security concerns for AI systems in critical visual interpretation tasks

This research reveals important security implications for deploying vision-language models in sensitive domains like autonomous driving, content moderation, and surveillance systems, emphasizing the need for more robust defenses against adversarial attacks.

SceneTAP: Scene-Coherent Typographic Adversarial Planner against Vision-Language Models in Real-World Environments