Privacy Funnel

The Privacy Funnel (PF) framework aims to learn data representations that balance utility for downstream tasks with privacy preservation against adversarial attacks, quantified through information-theoretic measures like self-information loss. Current research focuses on developing efficient solvers for the PF optimization problem, exploring both discriminative and generative model architectures (including variational autoencoders and generative adversarial networks), and extending the framework to handle various data types and learning scenarios, such as semi-supervised learning. This work is significant for its potential to improve data privacy in machine learning applications, particularly in sensitive domains like face recognition, by providing a principled approach to controlling information leakage while maintaining data utility.