This repository is the official implementation of [Improving Deep Learning with Differential Privacy using Gradient Encoding and Denoisin]
To install requirements:
conda env create -f env.yml
cd torchsearchsorted/
pip install .
Download https://drive.google.com/file/d/1k04lW5_IGMnHxfJMd-KR7Md9KlOgKtqi/view?usp=sharing to the main directory.
Please refer to https://github.com/aliutkus/torchsearchsorted if you had problem installing torchsearchsorted.
To train the model(s) in the paper, run this command:
python MNIST.py --noisemodel studentt --noiseparams 9 0 1.0 --epochs 100 --microbatch 1 --batch 256 --ngpus 4 --nprocs 50 --clip 1.0 --seed 0 --quantization --quantclip 1.0 --errcrt --quantmultiplier 4 --distancemultiplier 500 --distancethresh 0.700000 #MNIST
python CIFAR10.py --noisemodel studentt --noiseparams 9 0 0.9 --epochs 300 --microbatch 2 --batch 256 --ngpus 4 --nprocs 40 --clip 1.0 --seed 0 --quantization --quantclip 1.0 --errcrt --quantmultiplier 4 --distancemultiplier 500 --distancethresh 0.700000 #CIFAR
noise model will define the probability distributions used for privatizing the gradient vector, noiseparams is the parameters for the noise model.
Our model achieves the following performance on :
| Model name | Accuracy | eps |
|---|---|---|
| This work | 55% | 3.6 |
| DPSGD | 55% | 5 |
| Model name | Accuracy | eps |
|---|---|---|
| This work | 96% | 2.5 |
| DPSGD | 96% | 5 |