Malicious Prompt Detection on VLGuard & MSSBench
97.48AUROCVLMGuard
Evaluation results
| Method | Links | |
|---|---|---|
| VLMGuardBackbone=Qwen2.5-VL-72B-Instruct2024.10 | 97.48 | |
| VLMGuardBackbone=LLaVA-1.6-7B, Training Supervision=Unsupervised (no labeled data)2024.10 | 90.47 | |
| GuardRankBackbone=LLaVA-1.6-7B, Training Supervision=Fully supervised2024.10 | 88.14 | |
| HiddenDetectBackbone=Qwen2.5-VL-72B-Instruct2024.10 | 83.55 | |
| ASTRABackbone=Qwen2.5-VL-72B-Instruct2024.10 | 83.55 |