lmgraves
commented
3 years ago Hi! Thanks for reaching out. The sensitive data can be selected any way you choose - for the CIFAR and MNIST domains I selected an entire class and then had two test dataloaders, one including the sensitive class test data and one including the rest of the test data. To test the membership inference attack success I used a small random selection of training instances as sensitive data.
The Amnesiac ML Evaluation.ipynb file is the testing that we did for figure 7 in the paper, but looking at it right now it's a big buggy. I'll fix it later today and let you know when it's updated.
Thanks!
-Laura Graves
The github isn't the entire code base that we used, I hadn't included a full testing and training cycle
On Wed, Nov 25, 2020 at 2:17 AM Brown Chen notifications@github.com wrote:
Thank you for publishing the code. It's a great attempt for machine unlearning. By the way, how do you select the sensitive data during the unlearning process and I can't find the corresponding code in 'Amnesiac ML Evaluation.ipynb'. I want to reproduce the model accuracy on the target and non-target data. Looking forward to your reply!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/lmgraves/AmnesiacML/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AH2MZMUF5Y4NA5ONHMHA5BLSRSVQTANCNFSM4UCACK7A .
Brownchen
Chrisable
Thank you for publishing the code. It's a great attempt for machine unlearning. By the way, how do you select the sensitive data during the unlearning process and I can't find the corresponding code in 'Amnesiac ML Evaluation.ipynb'. I want to reproduce the model accuracy on the target and non-target data. Looking forward to your reply!