Table 2 Top-1 classification accuracy under the poster-printing attack scenario (M means meter, º means angel. The former is the accuracy of the target model, the latter is the accuracy of the defense model)
From: Defense against adversarial attacks in traffic sign images identification based on 5G
Setting | I-FGSM | C&W | Deepfool | JSMA |
|---|---|---|---|---|
5M 0° | 13.73%/82.91% | 5.91%/90.21% | 22.41%/82.61% | 15.71%/90.66% |
5M 15° | 11.27%/81.75% | 6.06%/84.82% | 23.04%/81.25% | 23.22%/89.34% |
5M 30° | 12.98%/85.96% | 10.31%/88.18% | 30.96%/87.15% | 26.65%/88.52% |
10M 0° | 12.96%/87.65% | 9.58%/87.52% | 25.63%/88.64% | 18.96%/90.72% |
10M 15° | 10.62%/89.51% | 8.69%/82.65% | 27.12%/89.61% | 23.41%/89.26% |
10M 30° | 6.09%/85.63% | 13.37%/85.31% | 29.59%/90.13% | 25.67%/89.17% |
20M 0° | 16.64%/87.69% | 14.24%/84.93% | 20.76%/87.69% | 21.19%/88.93% |
30M 0° | 27.21%/90.39% | 12.17%/87.45% | 19.36%/88.42% | 21.39%/89.93% |