A Federated Learning Framework with Blockchain-Based Auditable Participant Selection

Abstract

Federated learning is an important distributed model training technique in Internet of Things (IoT), in which participant selection is a key component that plays a role in improving training efficiency and model accuracy. This module enables a central server to select a subset of participants to perform model training based on data and device information. By doing so, selected participants are rewarded and actively perform model training, while participants that are detrimental to training efficiency and model accuracy are excluded. However, in practice, participants may suspect that the central server may have miscalculated and thus not made the selection honestly. This lack of trustworthiness problem, which can demotivate participants, has received little attention. Another problem that has received little attention is the leakage of participants' private information during the selection process. We will therefore propose a federated learning framework with auditable participant selection. It supports smart contracts in selecting a set of suitable participants based on their training loss without compromising the privacy. Considering the possibility of malicious campaigning and impersonation of participants, the framework employs commitment schemes and zero-knowledge proofs to counteract these malicious behaviors. Finally, we analyze the security of the framework and conduct a series of experiments to demonstrate that the framework can effectively improve the efficiency of federated learning.