Adversarial Attack on closed-source thinking MLLMs (evaluation set)
3.3ASR (GPT-5.4 Thinking)FRA-Attack
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FRA-AttackSimilarity Threshold=0.92026.05 | 3.3 | 41.6 | 15 | 61.4 | 4.6 | 45.6 | |
| M-Attack-V2Similarity Threshold=0.92026.05 | 2.1 | 37.7 | 8 | 54.7 | 2.2 | 42.6 | |
| FOA-AttackSimilarity Threshold=0.92026.05 | 0.8 | 30.1 | 3.9 | 45.3 | 1.8 | 34.8 | |
| M-AttackSimilarity Threshold=0.92026.05 | 0.6 | 26.4 | 2.6 | 41.4 | 1.4 | 30.6 |