Table 2 Performance of the three models under the Vocal Tract Length Normalization attack (UA/WA/ACC)
From: Black-box adversarial attacks through speech distortion for speech emotion recognition
αvtln | CNN-LSTM (%) | GCN (%) | CNN-MAA (%) |
|---|---|---|---|
0.20 | 9.25/10.36/9.80 | 9.54/9.61/9.58 | 8.85/9.12/8.99 |
0.15 | 11.56/11.84/11.70 | 10.22/11.50/10.86 | 10.75/10.94/10.85 |
0.10 | 15.62/17.45/16.54 | 14.12/17.21/15.67 | 14.55/14.85/14.70 |
0.05 | 20.55/24.63/22.59 | 19.31/20.56/19.94 | 19.42/19.78/19.60 |
0.00 | 62.54/64.27/63.41 | 75.23/72.32/73.78 | 76.24/73.32/74.78 |
− 0.05 | 24.77/25.61/25.19 | 25.49/27.38/26.44 | 20.55/21.64/21.10 |
− 0.10 | 16.51/16.35/16.43 | 14.85/15.64/15.25 | 17.43/17.87/17.65 |
− 0.15 | 12.43/14.62/13.53 | 11.26/12.43/11.85 | 12.65/13.44/13.05 |
− 0.20 | 10.45/11.24/10.85 | 10.46/10.65/10.56 | 10.32/10.55/10.44 |