Exploring efficient Federated Learning via Graph Signal Processing
In the domain of Federated Learning (FL), which facilitates model training among distributed clients without sharing data directly with each other, this thesis aims to explore efficient federated learning through using a graph signal processing technique to selectively sample clients during model training. We use the MNIST dataset, and we show that even when only a subset of similar clients is sampled and used in training, we can achieve comparable, if not similar, results. We select clients based on cosine similarity score, motivated by the idea that updates from clients with similar data can be sub-sampled and still maintain the quality and accuracy of the global training model. The sampling theory, which is rooted in graph signal recovery, provides a foundation for this approach, showing that perfect signal (or model) recovery is possible, even with somewhat imperfect sampling, when we use 2-dimentional interpolation technique or the graph Fourier transformation technique. In this thesis, we explore how the sampling technique can make federated learning processes better and more efficient during training as sharing data across different clients can be costly. We show that we can keep, or even improve, model performance when a group of better sampled clients is involved in the training process.
Embargo status: Restricted until 01/2030. To request the author grant access, click on the PDF link to the left.