Efficient multi-objective peer-to-peer federated learning

Abstract

Machine learning (ML) had proliferated in recent years, leading to higher scrutiny of how the training dataset is collated from multiple sources. Due to privacy concerns, Federated Learning is implemented to ensure that users’ privacy is not violated in the process of using their data for ML model training. However, datasets collected from different users or devices are not identically and independently distributed. Furthermore, training on datasets that are irrelevant to the ML model’s objective is detrimental to the model’s performance in the long run. As such, the contributions for this report are as follows: 1. implementing a modified version of Floyd Warshall algorithm to include the vertex’s objective for path retrieval, 2. modifying from existing peer-to-peer federated learning algorithm to factor in vertices’ objective when ML model is transmitted from source to destination vertex.