Skip to main content

Privacy-robustness sensitivity analysis for secure development of differentially privacy deep learning under distribution shift

2026, vol.18 , no.2, pp. 71-82

Article [2026-02-07]

Authors
B. Lavanya
S. Janani
Abstract

The widespread use of deep learning on sensitive data has positioned Differential Privacy (DP) as a central mechanism for protecting training confidentiality. Existing evaluations emphasize the privacy-utility trade-off measured on clean data, overlooking robustness under distribution shift. This study introduces Privacy-Robustness Sensitivity (PRS), a metric that explicitly quantifies the degradation in robustness induced by differential privacy by measuring the relative loss in corruption robustness with respect to clean accuracy as privacy constraints tighten. A privacy–utility–robustness evaluation framework is validated on CIFAR-10, CIFAR-10-C, CelebA, and a systematically constructed CelebA-C biometric corruption benchmark, enabling robustness-aware evaluation of privacy-preserving deep learning models for secure and resilient deployment scenarios.

Keywords

differential privacy, distribution shift robustness, three-dimensional evaluation, privacy-robustness sensitivity (PRS), privacy-preserving deep learning

DOI

https://doi.org/10.59035/LBBT3945

Download full article

Citation of this article:

B. Lavanya, S. Janani. Privacy-robustness sensitivity analysis for secure development of differentially privacy deep learning under distribution shift. International Journal on Information Technologies and Security, vol.18 , no.2, 2026, pp. 71-82. https://doi.org/10.59035/LBBT3945