GANobfuscator: Privacy-Preserving Synthetic Data Generation with GANs

GANobfuscator is a differentially private framework for generating high-quality synthetic data from sensitive datasets using Generative Adversarial Networks (GANs). It achieves this by incorporating differential privacy into the GAN training process through carefully adding noise to the model gradients. GANobfuscator allows leveraging the generative power of GANs while providing formal privacy guarantees within practical privacy budgets, without compromising data utility. Key features of it includes a gradient pruning strategy to improve training stability and scalability.

Dependencies:

• pytorch
• torchmetrics
• torch-fidelity
• torchvision
• matplotlib
• scikit-learn
• imageio
• numpy
• tqdm

Installation guide:

# Install torch-fidelity (provides the metrics to evaluate GANs)
!pip install torch-fidelity

# Install PyTorchThis (provides the evaluation metric)
!pip install torchmetrics


# Install PyTorch
pip install torch torchvision

# Install tqdm (Progress bar library)
pip install tqdm

# Install imageio (Library for reading and writing image data)
pip install imageio

# Install scikit-learn (Machine Learning library)
pip install scikit-learn

# Install torchmetrics
pip install torchmetrics

# Install matplotlib (Plotting library)
pip install matplotlib

# Install numpy
pip install numpy

import warnings

Running the file

• Download and open the code folder and then run the 'GANobfuscator_dp.ipynb file on Google Collab',

• Alternatively you can run this directly on collab with this link

https://colab.research.google.com/drive/1miVuyKg9Y6mTSTuM8Qri4BFOtMzwShvo?usp=sharing

Parameters used

Parameter	Value	Description
workers	2	Number of data loading threads
batch_size	32	Batch size for training
nz	100	Size of the latent noise vector
ngf	64	Number of feature channels in the generator
ndf	64	Number of feature channels in the discriminator
beta1	0.5	Beta1 parameter for Adam optimizer
ngpu	1	Number of GPUs to use for training
num_test_samples	16	Number of test samples
nc	1	Number of channels in the input data
d_lr	5e-5	Learning rate for the discriminator
g_lr	5e-5	Learning rate for the generator
num_fake_images	1000	Number of fake images to generate
epsilon	10000,10,4	Epsilon value for differential privacy
delta	1e-5	Delta value for differential privacy
num_epochs	15	Number of training epochs
max_per_sample_grad_norm	1e-2	Maximum per-sample gradient norm
max_grad_norm	1e-2	Maximum gradient norm (same as max_per_sample_grad_norm)