Adversarial Attack on 100 source-target pairs standard (test)
83Attack Success Rate (ASR)FRA-Attack
Evaluation results
| Method | Links | ||
|---|---|---|---|
| FRA-AttackDefense=Center crop (r=0.9), Target Model=Claude-Opus-4.62026.05 | 83 | 67.9 | |
| FRA-AttackDefense=JPEG (q=75), Target Model=Claude-Opus-4.62026.05 | 80 | 64.4 | |
| M-Attack-V2Defense=JPEG (q=75), Target Model=Claude-Opus-4.62026.05 | 70.7 | 57 | |
| M-Attack-V2Defense=Center crop (r=0.9), Target Model=Claude-Opus-4.62026.05 | 67 | 54.8 | |
| FRA-AttackDefense=Center crop (r=0.9), Target Model=Gemini-3-flash2026.05 | 63 | 53.8 | |
| FRA-AttackDefense=Center crop (r=0.9), Target Model=GPT-5.42026.05 | 59 | 49.7 | |
| FRA-AttackDefense=JPEG (q=75), Target Model=Gemini-3-flash2026.05 | 55 | 47.8 | |
| M-Attack-V2Defense=Center crop (r=0.9), Target Model=Gemini-3-flash2026.05 | 51 | 47.1 | |
| FRA-AttackDefense=JPEG (q=75), Target Model=GPT-5.42026.05 | 48 | 43.8 | |
| M-Attack-V2Defense=Center crop (r=0.9), Target Model=GPT-5.42026.05 | 47 | 42.6 | |
| M-Attack-V2Defense=JPEG (q=75), Target Model=Gemini-3-flash2026.05 | 45 | 42 | |
| M-Attack-V2Defense=JPEG (q=75), Target Model=GPT-5.42026.05 | 35 | 33.9 | |
| FRA-AttackDefense=Gaussian (σ=0.5), Target Model=Claude-Opus-4.62026.05 | 29 | 28.8 | |
| FRA-AttackDefense=Gaussian (σ=0.5), Target Model=Gemini-3-flash2026.05 | 22 | 21.4 | |
| M-Attack-V2Defense=Gaussian (σ=0.5), Target Model=Claude-Opus-4.62026.05 | 21 | 21.5 | |
| M-Attack-V2Defense=Gaussian (σ=0.5), Target Model=Gemini-3-flash2026.05 | 13 | 17.1 | |
| M-Attack-V2Defense=Gaussian (σ=0.5), Target Model=GPT-5.42026.05 | 5 | 8 | |
| FRA-AttackDefense=Gaussian (σ=0.5), Target Model=GPT-5.42026.05 | 3 | 8 |