Skip to content

InstaHide: An instance-hiding scheme for privacy-preserving distributed learning

InstaHide: Instance-hiding Schemes for Private Distributed Learning

arXiv· arXiv:2010.02772· Yangsibo Huang, Zhao Song, Kai Li, Sanjeev Arora·Indexed by this site · 20 days ago· Originally publishedDefenses
AI summary

The researchers propose InstaHide, a training image encryption scheme for privacy-preserving distributed learning that can be inserted into existing distributed deep learning pipelines, allowing multiple parties to collaboratively train a shared model on private data with little impact on test accuracy. The scheme encrypts each training image with a one-time key by mixing several randomly selected images and applying a random per-pixel mask, and can also mix in large public datasets such as ImageNet during encryption to improve security. Experiments show the scheme resists known privacy attacks, and theoretical analysis shows that successfully attacking privacy requires solving a hard computational problem; the authors also prove that the per-pixel mask is essential for security, and that using only Mixup would be broken by certain efficient attacks. The challenge dataset is released at https://github.com/Hazelsuko07/InstaHide_Challenge,代码见 https://github.com/Hazelsuko07/InstaHide。

Related events and news

0 events · 1 news mentions

News that mentions this paper

Read originalarxiv.org