davidtjones / diffp

Differential Privacy

Running the code is simple, simply perform python run.py active|expert|gan train|evaluate dataset

Where dataset is some folder located in ./datasets

Useful run script commands

Currently in refactor

Notes/To do/Future work

• Expert Classifier

  • learning rate scheduling: current model converges after 2-3 epochs to around 73%
  • class balancing?
  • data augmentation? (probably through lighting changes?)

• GAN

  • class balancing?
  • larger kernel size in earlier convolutional layers
  • better regularization
  • increase output resolution??
  • potentially do more epochs

• Need to make some damn plots

  • GAN sample
  • expert classifier confusion matrix