Subject

Federated Learning (FL) is a decentralized training framework that aims to learn collaboratively from multiple partners without sharing sensitive data. Because partners hold various types of data, typically non-Independently Identically distributed (non-IID), this can cause trained models on the client side to drift and drastically jeopardize training performance. Some studies [LS20, SL21] have investigated and relieved the issue of drift in feature extraction aspects by either incorporating a proximal term into the loss function to constrain the drift between the global and local model parameters or by restricting the distance between the representations learned by the local and global models.

Kind of work

The aim of this thesis is to explore the drift problem in federated learning from a deeper perspective. Specifically, the student is expected to investigate which areas of the training data the neural network utilizes for feature extraction and decision-making, and how these areas drift during the optimization process. Through visualizing the drift problem, this thesis aims to propose a new optimization algorithm that can effectively relieve the drift and improve training performance.

Framework of the Thesis

In this thesis, the student is first required to begin by conducting a thorough review of the current literature on federated learning, with a focus on the optimization strategies utilized in the classical federated learning framework. Next, the student should explore and visualize the drift problem in federated learning. Finally, the student is encouraged to investigate potential optimization strategies to mitigate the drift problem.

[ME17] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas, “Communication- efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics. PMLR, 2017, pp. 1273–1282.
[LS20] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In Conference on Machine Learning and Systems, 2020.
[SL21] J. Sun, A. Li, B. Wang, H. Yang, H. Li and Y. Chen, "Soteria: Provable Defense against Privacy Leakage in Federated Learning from Representation Perspective,"?2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 2021, pp. 9307-9315, doi: 10.1109/CVPR46437.2021.00919.

Number of Students

1

Expected Student Profile

Good knowledge of Machine Learning, AI and data processing. Good programming skills in Python (particularly PyTorch)