Adversarial Robustness on Figstep audio
6ASRSARSteer
Evaluation Results
| Method | Links | |
|---|---|---|
| SARSteerAttack type=PGD (ϵ=0.005), Base Model=Qwen2-Audio2025.10 | 6 | |
| SARSteerAttack type=GCG (audio), Base Model=Kimi-Audio2025.10 | 6 | |
| SARSteerAttack type=Gaussian (σ=0.01), Base Model=Qwen2-Audio2025.10 | 7.2 | |
| SARSteerAttack type=PAIR (audio), Base Model=Kimi-Audio2025.10 | 10 | |
| Kimi-AudioAttack type=GCG (audio), Defense condition=No Defense2025.10 | 14 | |
| Qwen2-AudioAttack type=PGD (ϵ=0.005), Defense condition=No Defense2025.10 | 24.4 | |
| Qwen2-AudioAttack type=Gaussian (σ=0.01), Defense condition=No Defense2025.10 | 27.2 | |
| Kimi-AudioAttack type=PAIR (audio), Defense condition=No Defense2025.10 | 58 |