Trustworthy-AI-Group / TransferAttack

TransferAttack is a pytorch framework to boost the adversarial transferability for image classification.
MIT License
268 stars 41 forks source link

关于 L2T 的结果以及一些论文链接上的问题 #60

Closed the-full closed 4 months ago

the-full commented 4 months ago
您好!首先表达一下感谢,这个仓库的存在帮我节省了不少跑代码和查文献的时间。然后是我的问题,我在尝试复现 L2T 的结果时发现 TransferAttack 上的评估结果和其开源的仓库上的评估结果在 cnn 模型上有明显差异(单卡, batchsize=2 下测试),具体结果如下: 模型名称 ASR in TransferAttack (%) ASR in L2T
resnet18 100.0 100.0
resnet101 87.7 96.2
resnext50_32x4d 89.7 97.2
denssenet121 98.8 98.4
vit_base_patch16_224 51.4 48.2
pit_b_224 64.7 64.0
visformer_small 80.0 78.0
swin_tiny_patch4_window7_224 82.5 80.4
我注意到 L2T 的代码这里,和 TransferAttack 的这里 对 cnn 模型权重设置不同,将其设置为 DEFAULT 后的结果为: 模型名称 ASR under DEFAULT (%)
resnet18 100.0
resnet101 87.6
resnext50_32x4d 88.6
denssenet121 98.4
vit_base_patch16_224 48.2
pit_b_224 64.0
visformer_small 80.4
swin_tiny_patch4_window7_224 82.5

我想 README 中所报告的结果可能需要更新一下,另外还应该补充一下这篇论文的链接 https://arxiv.org/abs/2405.14077 。此外,在 ir.py 的这里给的链接是一篇差分隐私的论文,我想正确的链接应该是这个:https://arxiv.org//abs/2009.11729

xiaosen-wang commented 4 months ago

Thank you for pointing out the mistakes. We have updated the L2T's results and IR's link. Also, we find that there might be other wrong URLs and will check all the URLs.