Full Citation
Title: A Machine Learning Algorithm With Adaptive Truncated Concentrated Differential Privacy
Citation Type: Conference Paper
Publication Year: 2023
ISBN: 979-8-3503-3421-0
ISSN:
DOI: 10.1109/ITOEC57671.2023.10292102
NSFID:
PMCID:
PMID:
Abstract: The rapid advancement of artificial intelligence technology, particularly driven by machine learning, has substantially impacted different industries and daily life. Regrettably, as the machine learning parameters are obtained from the mathematical expressions of datasets, Attackers can use these parameters to deduce sensitive information that the data owner with no intention to disclose reveal. Therefore, this paper proposes a novel approach that utilizes differential privacy technology to safeguard model parameters. It introduces an adaptive algorithm that provides differential privacy protection to machine learning. More specifically, this paper utilizes truncated concentrated differential privacy (tCDP) to perturb gradients in the local model training process of participants. Additionally, the algorithm incorporates adaptive step sizes and dynamic privacy budget adjustment techniques to enable the flexible distribution of privacy budgets. Based on the experimental and privacy analyses, the algorithm presented in this paper outperforms DP-AGD, thereby ensuring model effectiveness while preventing privacy leakage in the local models of the participating parties.
Url: https://ieeexplore.ieee.org/document/10292102/
User Submitted?: No
Authors: Li, Qiong; Wang, Qingjiang; Liu, Yang; Guo, Fucheng; Li, Zhiqing
Conference Name: 2023 IEEE 7th Information Technology and Mechatronics Engineering Conference (ITOEC)
Publisher Location:
Data Collections: IPUMS International
Topics: Methodology and Data Collection
Countries: