Effective Cost Sharing for Federated Learning Framework

Abstract

In a federated learning setup the quality of the datasets of the participants may vary greatly. Sharing costs associated with training a model equally across all participants may not appeal participants with high quality data. In this work we implemented Shapley values to compute the contributions of participants in a federated learning framework. We evaluate different methods to compute the contributions and propose methods to share costs based on said contributions. Furthermore we implemented Shapley values for features of participants.

Yann Girsberger

Bachelor's Thesis