Benchmarking Differentially Private Residual Networks for Medical Imagery

TL;DR

This paper benchmarks the effectiveness of differential privacy mechanisms, Local-DP and DP-SGD, in medical imaging, analyzing the privacy-accuracy trade-off and real-world applicability of privacy guarantees.

Contribution

It provides a comparative analysis of Local-DP and DP-SGD for medical imagery, highlighting their performance and practical utility in privacy-preserving medical AI.

Findings

Local-DP and DP-SGD show different privacy-accuracy trade-offs

Theoretical privacy guarantees may not fully translate to real-world effectiveness

Benchmark results inform best practices for privacy in medical imaging

Abstract

In this paper we measure the effectiveness of $\epsilon$-Differential Privacy (DP) when applied to medical imaging. We compare two robust differential privacy mechanisms: Local-DP and DP-SGD and benchmark their performance when analyzing medical imagery records. We analyze the trade-off between the model's accuracy and the level of privacy it guarantees, and also take a closer look to evaluate how useful these theoretical privacy guarantees actually prove to be in the real world medical setting.