Nikola Konstantinov: Incentivizing Collaboration in Federated Learning

Abstract: Collaborative and federated learning techniques have the potential to enable training powerful machine learning models from distributed data. However, in many cases the potential participants in such collaborative schemes have additional incentives to end-model accuracy. For example, they may be competitors on downstream tasks or may be concerned about the privacy of their data. This creates incentives for individual participants to obfuscate their messages or even damage the training process for the other players, which often undermines the benefits of collaboration. In this talk I will present recent results regarding participation incentives in federated learning (FL). In particular, I will cover several theoretical models for rational data-sharing decision making in the context of market competition and privacy concerns. I will also show how to design FL protocols that provably incentivize honesty during training, even in the presence of conflicting incentives.