ksreenivasan / ood_federated_learning Goto Github PK

Hi, I have run the experiment, once I change the total number of clients, but stay the same number of clients participate per round, the defense's accuracy of detecting attacker drop. I think the different is the datasize per client now have been change, but why it cause the attacker more effective?

We ran the simulated_averaging.py, but found the min_size is always 0 ,1 or 2. With the loop termination condition is min_size<10 or dataset=='mnist' and min_size <100, the program cannot jump out of the loop.

Hello,

When trying to generate the adversarial dataset for the EMNIST+ARDIS setting I noticed that the honest (EMNIST) images and the ARDIS images have a different orientation. Specifically, the images from torchvision EMNIST dataset are 90 degrees rotated and mirrored (in generating_poisoned_DA.py), whereas the ARDIS-7's have a normal orientation. The images are then saved together into one attacker dataset. Therefore, it seems that you are training a model on 90deg+mirrored EMNIST while inserting an edge-case ARDIS backdoor that is not rotated/flipped.

Am I missing something here (are you correcting these EMNIST images later somewhere)?

Thanks,
Hidde

First of all, I want to thank you all cause you share this repo to all of us. I've read your paper and try to improve something using your code. However, I find out that it seems like you haven't implement the last 2 experiments with Sentiment140 and Reddit datasets, have you?

In utils.py line 719 and 722, two dataset files are mentioned: new_green_cars_train.pkl and new_green_cars_test.pkl. But they are not available in the repo. I am assuming they are related to the semantic backdoor attack. In any case, is it possible to share these files?

Thanks,
Ehsanul

When I run the program, I need to give the eps for which process？For pgd or for repleacement?

Hello! I encountered an issue while reproducing your code from "https://github.com/ksreenivasan/OOD_Federated_Learning". Specifically, when running "./src/datasets/build_twitter_backdoor.py" in the "language-tasks-fl" directory, I encountered the following error message: "No such file or directory: '../../data/sentiment-140/vocabGood_0.25_0.pkl'." It seems that this file is missing. Could you please verify if you have uploaded this file?
Thanks!