The School of Computer Science is pleased to present…
Practical Secure Aggregation in Federated Learning using Additive Secret Sharing
MSc Thesis Defense by: Hamid Fazli Khojir
Date: Wednesday, January 11, 2023
Time: 9:30 am – 11:00 am
Location: Essex Hall, Room 122
Reminder: Recording of your attendance is mandatory – Part I: QR Code, Part II: Sign-in sheet.
Two-part attendance is mandatory; arrive 5-10 minutes prior to the event starting – LATECOMERS WILL NOT BE ADMITTED once the door has been closed and the presentation has begun. Please be respectful of the presenter by NOT knocking on the door for admittance.
Abstract:
Deep learning has enabled many industries to use data and train models with unlimited applications. However, data can include sensitive and private information of individuals, companies, or even hospitals. Therefore, data cannot simply be shared with a third party for training the model as it breaks the privacy of data owners, which is strongly prohibited by laws. Google addressed this problem in 2016 by introducing Federated Learning (FL), allowing users to train a model collaboratively by aggregating locally-computed updates while the dataset is kept in the local device. However, recent works have shown that the central server can infer sensitive information about the local dataset as it has access to updates of each client. Researchers are trying to provide an efficient solution for this problem, known as secure aggregation, in terms of the model’s accuracy, communication, and computation cost for clients. Motivated by this problem, we will propose a scalable, highly efficient framework for clients that provides guaranteed privacy using additive secret sharing.
Keywords: Deep Learning, Federated Learning, Secure Aggregation, Additive Secret Sharing
MSc Thesis Committee:
Internal Reader: Dr. Saeed Samet
External Reader: Dr. Ning Zhang
Advisor: Dr. Dima Alhadidi
Chair: Dr. Alioune Ngom
Bookings
Bookings are closed for this event.
No Responses