Privacy-Preserving Outsourced Deep Neural Network Training: New Dual-Cloud Framework and Efficient Construction

View at publisher Summarise with VIRAT
Authors: Chen-Fan Chang, Kuohsiu David Huang, Bo-Jia Chen, Yu-Chi Chen
Cited by: 0
References: 41

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Recently, deep neural networks (DNNs) have achieved excellent results in face recognition, natural language processing, and object classification. With the development of data science and the continuous increase in the size of datasets, the computing power and storage requirements for DNN training have also increased. Although GPUs and training methods have been continuously optimized for decades, the storage requirements remain a significant cost. To reduce the cost, datasets are often centrally placed in cloud storage for interactive training. Data security is undoubtedly the primary concern for such applications. Datasets for DNN training are usually collected by various institutions and often contain a lot of sensitive information. To prevent confidential information from leaking, data in cloud storage must be encrypted or protected. In this paper, we use secret sharing, a very lightweight method, to implement privacy-preserving outsourced DNN training. Using secret sharing, the data and models can be split into two shares and outsourced to two cloud service providers for training. Since the calculations are performed by the cloud service providers, the model trainer can be a device with limited computational power, such as a lightweight portable device.

Fields of Study & Subjects

Cloud Data Security Solutions Cryptography and Data Security Privacy-Preserving Technologies in Data

Publication details

Journal / source
IEEE Transactions on Emerging Topics in Computing
Volume
14
Issue
2
Pages
467–482
Publication date
27 March 2026
Language
EN