CIT 07-Privacy, Security, FL

Privacy, Security and Communication Efficiency in Federated Learning

CIT - Electrical and Computer Engineering

Chair of Communications Engineering

Federated Learning

Computer Science

Participation also possible online only:

Planned project location:

Technical University of Munich
Theresienstrasse 90
80333 München

Yue

Xia

yue1.xia@tum.de

015226565973

Dr

Rawad

Bitar

rawad.bitar1@gmail.com

Privacy, Security and Communication Efficiency in Federated Learning

Nowadays, data is distributed across various devices and locations and is often sensitive. Training a neural network on the sensitive distributed data requires ensuring privacy for various reasons, e.g., ethical and legal. While federated learning (FL) is designed to maintain privacy, it has been shown that when leaking the individual users' model updates, vanilla FL violates the privacy of the individual clients. Significant research is carried out to enhance the privacy of FL. However, privacy comes at a fundamental cost, particularly at the cost of an increased
communication overhead.

Security is equally important, and there is a clear inherent tension between privacy and security. The former requires learning nothing about the clients' individual models updates, while the latter requires enough knowledge about the clients' individual models updates to detect the corrupted updates. Hence, there is a challenging nature of guaranteeing both IT privacy and security, while keeping the solution practical in terms of communication overhead.

The student's task will be to conduct a thorough literature review to understand the problem, and hopefully come up with a solution that achieves a balance between privacy, security and communication efficiency.

40

3 years of bachelor studies completed