VRAG introduces a novel retrieval-augmented framework that detects adversarial patches in images without requiring model retraining or fine-tuning.

• Leverages Vision-Language Models to identify malicious patches by comparing them to known attacks
• Achieves training-free defense through similarity matching with stored attack patterns
• Offers practical, real-world protection for vision systems with minimal implementation overhead
• Demonstrates improved security for computer vision applications against sophisticated visual attacks

This research provides security teams with an efficient, deployable solution to protect AI vision systems from adversarial manipulation in production environments.

Don't Lag, RAG: Training-Free Adversarial Detection Using RAG