Attribution

This code is based on the PATE framework presented by Papernot et al. in: arXiv:1610.05755.

Dependencies

This requires torch, torchvision, scipy, numpy, and tensorboard. So far (fingers crossed!) we've been able to use never versions of all these packages without much difficulty.

How to run

For our purposes, we train teachers like so:

python torch_teachers.py

This should output a number of .tch files. After training the teachers, we run the voting aggregation algorithms, like so:

python get_predicted_labels.py

Within that code, there are options to change what the aggregation mechanism is. This code generates a .npy file. After the data has been aggregated, we train the student like so:

python student.py