ege-erdogan / splitguard Goto Github PK

Hey @ege-erdogan , i noticed your work and have some questions as follows:

• In my understanding, the training data (feature + label) are vertically splited in split learning. While in the paper (same in unleashing the tiger), it looks like federeated learning, where each client hold entire features and labels for different training samples.
• In SG-AD, if the gradient list used for train LOF is constructed as: 1) run batch training locally; 2) append the gradients for each layer (or gradients for some specific layer like the last layer?) to the list. Therefore the gradient list is of shape (#num_of_iterations, K, W, H), where K denotes the number of model layers and W, H denote the weight shape for some layer.

Hello, I want to know how to generate the files in these two directories.
cifar10_HONEST/honest_cifar_grads_
cifar10_newAtt_fake/FSHA_newAtt_cifar10_fakes_

An interesting work! Could you please provide the file and the code to generate and save Gradients, which are used in sg_ad_main.ipynb.
Thanks a lot!