sclbd / dbd Goto Github PK

Hi,

Thanks for the nice work.
Do you plan to also add the code for the other attacks (such as WANet) used in the paper to the repo?

Sincerely,
Hadi.

I think your comparison with Neural Cleanse is unfair.
我认为您和NC的对比并不公平。

Neural Cleanse only uses a clean dataset. However, your method needs the poisoned training dataset.
NC只需要一个干净的小数据集。然而您的方法需要完整的投毒训练集。

In fact these are totally two different scenarios, it is hard to say which setting is simple or needs extra requirements.
我认为这完全是两个不同的场景，很难说哪个场景下需要的额外信息更多。

If you have to compare, I think your method requires more 'defender capability' in practice.
如果硬要比的话，我也认为在实际应用中，您的方法比NC实现起来困难。

Your claim that Neural Cleanse needs 'extra requirements' is inexact and misleading.
您说NC需要额外的信息，这并不确切，而且具有误导性。

In WaNet original codes and BackdoorBench, the index of poisoned samples are changeable in different epochs. But for the poisoned-based attacks, the index of poisoned samples should be pre-defined and unchangeable during the training. So the re-implementation of WaNet in this paper follows the former setting or the latter?