regarding the "poison_injection_rate" and "clean subset percentage for linear classifier"

[cinout]

Dear Authors,

Thank you for sharing the code to facilitate the learning in the backdoor attack community.

I have some questions regarding the training details, and I hope to hear your answer.

The first is the "poison_injection_rate". In the paper, you mentioned that you used 50% poisoning rate for the target class, but I notice that in the HTBA configure files, the default poison_injection_rate is set to 1.0. Therefore, I wonder if the shared backdoored model checkpoints are trained with 50% or 100% poisoning rate?

The other question is regarding the second stage--linear classifier training. In the paper and in your shared checkpoints, you used both 1% and 10% of clean data to train the linear classifier. But in the code, e.g., moco/eval_linear.py, it seems the whole clean training set is used. I cannot find code regarding 1% and 10% subset. Maybe I missed something in the code, but can you tell me where to find the relevant code?

Thank you for reading this message. Look forward to hearing your reply.