Privacy in Practice: Research Challenges in the Deployment of Privacy-Preserving ML

As machine learning (ML) systems become increasingly integrated into sensitive domains such as healthcare, finance, and government, concerns over data privacy have risen significantly resulting in regulations controlling use of individuals’ private data. Privacy-preserving machine learning (PPML) integrating formal guarantees of differential privacy offers a promising solution heralded by many privacy and data analysis experts. However, despite notable theoretical advancements, practical deployment remains constrained by several real-world challenges including hyperparameter tuning for complex privacy-preserving algorithms to balance privacy and model utility, heterogeneity in the privacy preferences of individuals whose data is being used to train ML models, and the usability of PPML tools. Addressing challenges such as these remains critical for widespread adoption of PPML. In this paper, we outline the implications of these pressing challenges and offer research directions towards addressing barriers in the deployment of ML systems incorporating differential privacy in practice.