Secure and Efficient Federate Learning for Automated Internet-of-Things Environment
- Sponsor: Institute of Information & Communications Technology Planning &
Evaluation (IITP), South Korea
- Budget: KRW₩ 120,000,000 ($102,358.39)
- Period: May 1, 2021 - October 31, 2022
This project aims to perform comprehensive research on Federated Learning algorithms (FL) to enhance both security and efficiency. FL is a new framework that data is distributed over millions of mobile devices, Edge computing hosts, and servers. Instead of collecting enormous datasets at a central server, FL trains data on each device and transfer a training result as a form of parameters or gradients to the server. Then, the server will be able to obtain the final training model by averaging the results from participating devices.
The major advantage of FL is that it provides highly personalized models and does not compromise user privacy. However, some recent research results introduced the following vulnerabilities of FL which can be exploited to negate the key advantages, such as (a) Model Inversion Attack, (b) Membership Inference Attack, (c) Model Extraction Attack, etc.
This project will focus on enhancing security and efficiency of FL algorithms. Specifically, this project has the following objectives:
(a) Discovering hidden weaknesses or potential vulnerabilities of FL,
(b) Developing new FL algorithms, architectures, or computing models that are secure against those attacks,
(c) Establishing simulation environments to evaluate the proposed scheme,
(d) Demonstrating the effectiveness of the proposed scheme.