notesum.ai
Published at October 30Effective and Efficient Adversarial Detection for Vision-Language Models via A Single Vector
cs.CV
cs.CL
cs.CR
Released Date: October 30, 2024
Authors: Youcheng Huang1, Fengbin Zhu2, Jingkun Tang1, Pan Zhou3, Wenqiang Lei1, Jiancheng Lv1, Tat-Seng Chua2
Aff.: 1Sichuan University; 2National University of Singapore; 3Singapore Management University

| Paper | Scale | Harmful Types | Open Source | Data Filtering |
|---|---|---|---|---|
| (Zhang et al., 2023a) Arxiv | 200 | Harmful queries | \usym2713 | \usym2717 |
| (Tu et al., 2023) Arxiv | 3 | Toxic words | \usym2713 | \usym2717 |
| (Carlini et al., 2023) Neurips 2023 | - | Toxic words | \usym2717 | \usym2717 |
| (Qi et al., 2024a) AAAI 2024 | 3 | Toxic words | \usym2713 | \usym2717 |
| (Luo et al., 2024) ICLR 2024 | - | Harmful queries | \usym2717 | \usym2717 |
| (Shayegani et al., 2024) ICLR 2024 | 8 | Toxic words | \usym2717 | \usym2717 |
| RADAR (Ours) | 4,000 | Both | \usym2713 | \usym2713 |