search

ybdai7 / Backdoor-indicator-defense

[Usenix Security 2024] Official code implementation of "BackdoorIndicator: Leveraging OOD Data for Proactive Backdoor Detection in Federated Learning" (https://www.usenix.org/conference/usenixsecurity24/presentation/li-songze)
26 stars 1 forks source link

How to run BDIndicator on Google Colab? #6

Open HideLakitu opened 2 weeks ago

HideLakitu commented 2 weeks ago

well I use the command !python main.py --GPU_id "0" --params utils/yamls/indicator/params_vanilla_Indicator.yaml under T4 GPU try to run, then it said: _FileNotFoundError: [Errno 2] No such file or directory: 'saved_models/saved_model_global_model1200.pt.tar'

But isn't this .tar file automatically generated during training？ or I should edit resumed_model: "Jun.05_06.09.03/saved_model_global_model_1200.pt.tar" to something else which u meationed in redame file.

ybdai7 commented 1 week ago

Hi,

Yes. You should follow the instructions in the README.md. Generally speaking, the "resumed_model" parameter provides you a checkpoint, where you can quickly start poisoning regardless of the previous training. You can simply run a program without poisoning first, and save any checkpoint you want to start with. Then replace the "resumed_model" with the one you saved to start poisoning.

HideLakitu commented 1 week ago

Hi,

Yes. You should follow the instructions in the README.md. Generally speaking, the "resumed_model" parameter provides you a checkpoint, where you can quickly start poisoning regardless of the previous training. You can simply run a program without poisoning first, and save any checkpoint you want to start with. Then replace the "resumed_model" with the one you saved to start poisoning.

THX for reply, just set the resumed_model to False in file to run then it worked.

BTW, I checked corresponding .yaml file and found the total training round is 1900 ( # server training setting start_round: 1 end_round: 1900, ) correct? so do u have impression about how long u take for running before --- under TPU-G4 on Google, about nearly 3 hours by me to only complete less than 350 rounds, feel a little time consuming for just reach the limitation of computing resource within a day, I didn't pay for that.

ybdai7 commented 1 week ago

Yes, and you can modify the end_round in .yaml to different configurations. For the needed time, I think it is reasonable for you to run less than 350 rounds in 3 hs.