LtDNet-learning-the-difference
Learning the class-wise difference for image classification
Ablation on n_ref (total 10000 samples):
| n_ref | acc |
|---|---|
| 1 | 87.340% |
| 3 | 88.640% |
| 5 | 88.910% |
| 7 | 88.900% |
| 10 | 89.140% |
| 30 | 89.240% |
PGD attack (first 100 samples):
| n_ref | method | acc | asr |
|---|---|---|---|
| 1 | random | 52.941% | 47.059% |
| 3 | random | 33.333% | 66.667% |
| 5 | random | 27.451% | 72.549% |
| 7 | random | 23.529% | 76.471% |
| 1 | fixed | 72.549% | 27.451% |
| 3 | fixed | 56.863% | 43.137% |
| 5 | fixed | 39.216% | 60.784% |
| 7 | fixed | 35.294% | 64.706% |
Quickstart
python preprocess.pypython make_stats.pypython train.pypython test.pypython attack.py
by Armit 2023/02/22