Paper Image

Mixup data augmentation for differentially private image classification

Published on:

2 November 2023

Primary Category:

Machine Learning

Paper Authors:

Wenxuan Bao,

Francesco Pittaluga,

Vijay Kumar B G,

Vincent Bindschaedler

Bullets

Key Details

Proposes two techniques, DP-Mixup-Self and DP-Mixup-Diff, to enable mixup data augmentation with differential privacy

DP-Mixup-Self applies mixup to self-augmentations of individual training examples

DP-Mixup-Diff incorporates synthetic images from diffusion models into the mixup process

Achieves new state-of-the-art accuracy with differential privacy across datasets

AI generated summary

Mixup data augmentation for differentially private image classification

This paper proposes two novel techniques to enable multi-sample data augmentation like mixup to be used with differentially private learning for image classification. The techniques work by applying mixup to self-augmentations of individual training examples. This allows mixup to be used while preserving the differential privacy guarantee. Experiments show the techniques achieve state-of-the-art accuracy across datasets and privacy budgets when training models from scratch and fine-tuning pretrained models.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up