Analyzing the Influence of Medical Imbalanced Data on Performance and Fairness in Differentially Private Deep Learning

EasyChair Preprint 11619

Abstract

Deep learning carries a significant potential for a paradigm shift in healthcare and medicine. Unfortunately, deep learning poses privacy risks, as various inference attacks have revealed. Differential Privacy offers robust guarantees and substantial defense against privacy threats, making it a prevalent approach for privacy-preserving deep learning lately. Many recent approaches to deep learning and differentially private deep learning assume identically Distributed data, which is often not the case in real-world situations. In our study, we examine the impact of imbalanced data on differentially private deep learning. We find that imbalanced data negatively affects both the model's performance and fairness. We explore the trade-off between privacy, usefulness, and fairness. Our findings underscore the challenges of using standard deep learning algorithms in a differentially private context to achieve reliable results for underrepresented groups.

Keyphrases: Imbalanced medical data, differentially private deep learning., privacy-utility/fairness trade-off

@booklet{EasyChair:11619,
  author    = {Benladghem Rafika and Hadjila Fethallah and Merzoug Mohammed and Belloum Adam},
  title     = {Analyzing the Influence of Medical Imbalanced Data on Performance and Fairness in Differentially Private Deep Learning},
  howpublished = {EasyChair Preprint 11619},
  year      = {EasyChair, 2023}}