Dear authors,
We are currently trying to reproduce the ViT experiments using the code available on your GitHub repository. While working with the released implementation, we observed some variability in the backdoor effectiveness. In certain runs, the attack success rate (ASR) reaches close to 100%, as reported in the paper, whereas in others it decreases to very low level.
We believe this may be related to configuration choices or specific experimental settings, and we would like to ensure that we are reproducing the experiments correctly.
We would greatly appreciate your thoughts on the following points:
- Is the configuration provided in the GitHub repository directly suited for the ViT model, or are specific adjustments typically required (particularly in the run_script.py file)?
- Do you have any recommendations or best practices to improve the stability and reproducibility of the backdoor generation for this model?
Otherwise, we would like to know if, in your own experiments with ViT, you observed this kind of variation in backdoor effectiveness?
Thank you very much in advance for your time.
Dear authors,
We are currently trying to reproduce the ViT experiments using the code available on your GitHub repository. While working with the released implementation, we observed some variability in the backdoor effectiveness. In certain runs, the attack success rate (ASR) reaches close to 100%, as reported in the paper, whereas in others it decreases to very low level.
We believe this may be related to configuration choices or specific experimental settings, and we would like to ensure that we are reproducing the experiments correctly.
We would greatly appreciate your thoughts on the following points:
Otherwise, we would like to know if, in your own experiments with ViT, you observed this kind of variation in backdoor effectiveness?
Thank you very much in advance for your time.