Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification

Page view(s)
28
Checked on Sep 24, 2023
Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification
Title:
Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification
Journal Title:
Proceedings of 2022 the 12th International Workshop on Computer Science and Engineering
Keywords:
Publication Date:
27 July 2022
Citation:
Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification. (2022). Proceedings of 2022 the 12th International Workshop on Computer Science and Engineering. https://doi.org/10.18178/wcse.2022.06.014
Abstract:
Over recent years, machine learning (ML) methods have enabled considerable progress to be made within a variety of data-rich research domains. Genomics is one prominent field, with ML-based approaches achieving exciting results in a range of historically difficult tasks. One such area where ML approaches have achieved particular success is the classification of cancerous tissues using gene expression data. Despite this success, recent advances share a common issue with many other areas of ML research - state-of-the-art performance relies upon the aggregation of large training datasets. Many have raised concerns over the implications such large-scale data aggregation practice has for data privacy, and the nature of genomic data makes such concerns all the more pertinent within a cancer classification setting. Because of the sensitivity of such data, strict policies are enacted to protect patient privacy. Whilst currently unavoidable, such policies are a key roadblock to advances in ML research in this domain. Federated learning (FL) aims to solve this issue by allowing ML models to be trained on decentralized data. Despite the promise held by this approach, recent work has demonstrated that FL is not sufficient for fully private training of ML models. The incorporation of differential privacy (DP) into an FL framework has previously been shown to enable decentralized model training in a privacy-preserving manner. We introduce a method to incorporate custom differentially private algorithms directly into federated learning workflows that are not currently implemented by privacy libraries such as Opacus.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the A*STAR - Advanced Manufacturing and Engineering (AME) Programmatic Programme
Grant Reference no. : A19E3b0099
Description:
ISBN:
978-981-18-3959-7
Files uploaded:

File Size Format Action
14.pdf 1.61 MB PDF Open