Generative Adversarial Models for Privacy and Fairness

Explore generative adversarial models for privacy and fairness in this 38-minute lecture by Peter Kairouz from Google AI. Delve into zero-sum games, Generative Adversarial Networks (GANs), and their limitations in privacy preservation. Learn about membership inference attacks on GANs and how Differential Privacy (DP) can mitigate these risks. Discover TensorFlow Privacy and differentially private GANs (DP-GAN), including noisy Wasserstein GAN implementations and their results. Investigate context-aware fair data publishing and empirical risk minimization. Examine Generative Adversarial Privacy & Fairness (GAPF) concepts, including data-driven approaches and real-life applications using the GENKI dataset. Understand adversary's neural networks, feedforward and transposed convolution neural network encoders, and the trade-offs between fairness and utility. Explore Siamese-GAPF (S-GAPF) for handling sensitive labels with multiple values, and analyze fairness vs. utility in the Human Activity Recognition (HAR) dataset. Conclude with insights into Gaussian mixture data models for privacy and fairness in machine learning.

Intro
Zero-sum games
Generative Adversarial Networks (GANS)
Why not using GANs as they are?
Membership inference attacks for GANS
Differential Privacy (DP) to the rescue!
TensorFlow Privacy
Differentially private GANS (DP-GAN)
DP-GAN: noisy Wasserstein GAN
DP-GAN results
Context-aware fair data publishing
Empirical risk minimization with MI?
Generative Adversarial Privacy & Fairness (GAPF)
Example: GAPF under log-loss
Data-driven GAPF
Penalty method
Real-life data: GENKI dataset
Adversary's neural network
Feedforward Neural Network (FNN) encoder
Transposed Convolution Neural Network (TCNNP) encoder
GENKI fairness vs utility
Siamese-GAPF (S-GAPF) What if the sensitive label can take many values?
Real-life data: HAR dataset
HAR fairness vs utility
Gaussian mixture data model