This repository's purpose is to present some productions for my exam...

It is based on the PyTorch-DP framework (frozen in the 19th of June, 2020 version. Under Apache License 2.0)

1. A notebook that presents the budget computation of Differential Privacy with PyTorch-DP and links to properties in several articles for justifications. You can either read it or download it to use on your own server or run it online (quite slow to launch) thanks to

2. An attempt to convert "102-flowers" DL model to a DP version, see flowers/ directory.

• First, once and for all : download data.

  • mkdir data/ && cd data && wget https://s3.amazonaws.com/content.udacity-data.com/nd089/flower_data.tar.gz
  • tar xvf flower_data.tar.gz && cd ..

• Train model: python3 train.py or ./train.py (try -h to list optional arguments, like --cpu if no GPU available, --batch-size, --learning-rate, epochs, --disable-dp...)

• Record stats about accuracy: ./stats.py. Parameters to experiment are hard coded, see the docstring. Results are stored in experiment_stats/ directory.

• Record stats about GPU memory:

  • flowers_mem_monitor.py may be run directly to add one line in .cvs file.
  • flowers_mem_stats.sh launches flowers_mem_monitor.py for different parameters (modify code to chose their values), results are stored in mem_flowers/

3. An other attempt on cifar10 dataset.

• First without DP: cifar10.py
  • (net = home-made simple model imported from my_net.py)
  • net = fully pre-trained VGG16
  • net = not pre-trained VGG16
  • net = pre-trained on "features" only, not on "classifier" layers.
• Then converted to DP version: dp_cifar10.py

4. Other stuff

• Similar functions (than one of 'flowers') about MNIST dataset, see mnist_train.py, mnist_mem_stats.sh, mnist_mem_monitor.py in mnist/ directory.

• Calculate directly a privacy budget without launching any training with torchdp/scripts/compute_dp_sgd_privcacy.py (This adaptation to PyTorch-DP I wrote from an equivalent script for TensorFlow is now included in PyTorch-DP repository 😏).