A naukri.com initiative
Arxiv
4d
101
Image Credit: Arxiv
To Shuffle or not to Shuffle: Auditing DP-SGD with Shuffling
- The Differentially Private Stochastic Gradient Descent (DP-SGD) algorithm allows the training of machine learning (ML) models with formal Differential Privacy (DP) guarantees.
- A novel DP auditing procedure has been introduced to analyze DP-SGD with shuffling and it has been shown that DP models trained with this approach have considerably overestimated privacy guarantees.
- The study assesses the impact on privacy leakage of several parameters, including batch size, privacy budget, and threat model.
- The usage of shuffling instead of Poisson sub-sampling in DP-SGD can lead to significant privacy leakage, as observed in this research.
Read Full Article
6 Likes
For uninterrupted reading, download the app