Differentially Private Generative Adversarial Networks with Model Inversion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Standard differentially private (DP) stochastic gradient descent for Generative Adversarial Networks (GANs) can degrade synthetic data quality and hinder convergence due to added noise.
• Protecting sensitive data during GAN training is crucial for privacy-preserving machine learning.

Purpose of the Study:

• To propose a novel method, Differentially Private Model Inversion (DPMI), to enhance privacy and performance in GAN training.
• To address the limitations of existing DP-GAN methods in terms of synthetic data quality and convergence.

Main Methods:

• The proposed DPMI method maps private data to a latent space using a public generator.
• A lower-dimensional DP-GAN with improved convergence properties is then applied in the latent space.
• The approach was evaluated on CIFAR10, SVHN, and a facial landmark dataset for Autism screening.

Main Results:

• DPMI demonstrated superior performance over the standard DP-GAN method.
• Improvements were measured using Inception Score, Frechet Inception Distance, and classification accuracy.
• The proposed method maintained the same level of privacy guarantee.

Conclusions:

• Differentially Private Model Inversion (DPMI) offers a more effective approach for privacy-preserving GAN training.
• DPMI significantly improves the quality of synthetic data and the convergence of GANs under differential privacy.
• This method shows promise for applications requiring sensitive data protection, such as Autism screening.