git-disl / Lockdown

A backdoor defense for federated learning via isolated subspace training (NeurIPS2023)
14 stars 0 forks source link

ASR #1

Open wydwdy opened 3 months ago

wydwdy commented 3 months ago

Thank you for providing the code, but I encountered a problem while running it. I hope you can help me solve it. I used the 'lockdown' method to run the 'fmnist' dataset according to your code, but the results were biased. The 'ASR' I obtained was as high as 50% and 60%, respectively, for both IID and non IID scenarios. May I ask if this result is reasonable, but I have run it multiple times and it has been the same result. What kind of modifications need to be made to the runtime of this dataset if it is unreasonable

huangtiansheng commented 3 months ago

For fashionMnist, the dense_ratio in the options.py should be set lower. We recommend to use Cifar10 -resnet9 for default benchmark. The default hyper-parameter should work for this dataset.

wydwdy commented 3 months ago

Thank you for your answer, it has been very helpful to me.