search

HanxunH / Unlearnable-Examples

[ICLR2021] Unlearnable Examples: Making Personal Data Unexploitable
https://hanxunh.github.io/Unlearnable-Examples/
MIT License
151 stars 17 forks source link

A problem about noise generating. #15

Open zhangyisheng0629 opened 1 year ago

zhangyisheng0629 commented 1 year ago

Hello author, the noise in your experiments was generated after data augmentation, but the noise was added to the clean datasets before the data augmentation. Is this not rigorous ? Thank you very much!

HanxunH commented 1 year ago

Hi,

The noise was generated without augmentation by setting the transform to the same one as the test set transform. This is usually just the ToTensor operation.

https://github.com/HanxunH/Unlearnable-Examples/blob/28e1e417259ce5baf0d3e68d3408d9f8ecc3024f/dataset.py#LL113C1-L114C1

The noise added before data augmentation for training is due to the experimental setting. Users who wish to protect their data, usually do not have access to the training process, including data augmentation. Considering that user post their photo on social media, they do not control how these images are augmented if it is being used to train models.