search

AlanPeng0897 / Defend_MI

[KDD 2022] "Bilateral Dependency Optimization: Defending Against Model-inversion Attacks"
https://arxiv.org/pdf/2206.05483.pdf
MIT License
22 stars 5 forks source link

BiDO (lambda_x=0 and lambda_y=0) and No. Def models #8

Open hosytuyen opened 1 year ago

hosytuyen commented 1 year ago

Hi,

Thank you for sharing amazing work.

I am currently working on your code and curious about the difference between models training with BiDO (lambda_x=0 and lambda_y=0) and No. Def model.

Based on my understanding, it should be similar? In your code, I noticed that the difference comes from hyper-parameters (learning rate, learning rate scheduler, etc)?

Thank you very much.

AlanPeng0897 commented 1 year ago

Yes, you are right. For the No. Def model, I follow the same training protocol (optimizer, lr_scheduler, etc) as previous works. While in BiDO, I switch to Adam simply because I find it works better than SGD.