Full Citation

Title: Adaptive Distributed Differential Privacy with SGD

Citation Type: Conference Paper

Publication Year: 2020

ISBN:

ISSN:

DOI:

NSFID:

PMCID:

PMID:

Abstract: Privacy leakage is an important issue for machine learning. Existing privacy-preserving approaches with differential privacy usually allow the server to fully control users’ information. This may be problematic since the server itself may be untrusted, leading to serious privacy leakage. Besides, existing approaches need to choose a fixed number of iterations, so that the total privacy budget is finite. Fine-tuning is a common practice, which needs a lot of opportunities to try. However it is not allowed in the actual environment, because multiple access to user data will bring serious privacy risks. In this paper, we aim to address the problem of achieving privacy preserving with distributed differential privacy. In this scenario, different from the traditional need to disclose some local data to the centralized server, each participant keeps the data locally, to achieve better privacy protection effect. We propose a novel algorithm for privacy-preserving training with adjustable iteration steps by sampling techniques. The validity of the algorithm is verified by theoretical analysis and experimental evaluation on real-world datasets

Url: https://www2.isye.gatech.edu/~fferdinando3/cfp/PPAI20/papers/paper_28.pdf

User Submitted?: No

Authors: Cheng, Junhong; Liu, Wenyan; Wang, Xiaoling; Lu, Xingjian

Conference Name: Association for the Advancement of Artificial Intelligence

Publisher Location:

Data Collections: IPUMS USA

Topics: Other, Population Data Science

Countries: