Table 5 Robust defense: we add noise based on different SNR, the lower the SNR, the heavier the added noise
From: Towards the universal defense for query-based audio adversarial attacks on speech recognition system
SNR(dB) | CS-attack | DW-attack | IRTA-attack | DS-attack | Average |
|---|---|---|---|---|---|
150 | 100/3.92/98.00 | 100/1.70/84.74 | 100/56.00/84.00 | 100/11.00/82.5 | 100/17.44/87.31 |
100 | 100/3.92/98.00 | 100/1.70/84.74 | 100/56.50/84.75 | 100/11.00/82.50 | 100/17.57/87.50 |
75 | 100/3.92/98.00 | 100/1.70/84.74 | 100/58.50/87.75 | 100/12.50/93.75 | 100/19.16/91.06 |
50 | 100/3.92/98.00 | 100/1.70/84.74 | 100/63.00/94.50 | 100/13.00/97.50 | 100/20.41/93.69 |
25 | 100/3.92/98.00 | 100/1.70/84.74 | 100/63.00/94.50 | 100/13.00/97.50 | 100/20.41/93.69 |
0 | 15.10/3.70/92.50 | 22.50/1.50/75.00 | 0/10.00/15.00 | 5.20/3.00/22.50 | 10.70/4.55/51.25 |