A blockchain-infused framework for secure federated learning

Abstract Federated learning (FL) provides an effective framework for distributed data analysis with the model trained on raw data on local devices. While sharing aggregated data or gradient with the server, preserving data privacy is an important requirement in many applications such as healthcare, insurance, legal and social networks. Blockchain (BC) has shown potential in distributed data storage with immutability, transparency and consensus based protocol/transaction execution. In this paper, a novel Blockchain-Infused Framework for Secure Federated Learning, referred to as BIF-SFL is proposed that uses a round-based training mechanism, where trained models are securely recorded on a BC network. A reputation-based incentive mechanism is implemented to build trust among participating devices, where rewards are managed through smart contracts. Data privacy is preserved in BIF-SFL by employing differential privacy. The experimental results show the effectiveness of the proposed framework, achieving a notable accuracy of around 97%.