Federated learning is a type of privacy-preserving, collaborative machine learning. Instead of sharing raw data, the federated learning process cooperatively exchanges the model parameters and aggregates them in a decentralized manner through multiple users. In this study, we designed and implemented a hierarchical blockchain system using a public blockchain for a federated learning process without a trusted curator. This prevents model-poisoning attacks and provides secure updates of a global model. We conducted a comprehensive empirical study to characterize the performance of federated learning in our testbed and identify potential performance bottlenecks, thereby gaining a better understanding of the system.
Keywords: blockchain; federated learning; model-poisoning attack; smart contract.