An Efficient Attribute-Based Participant Selecting Scheme with Blockchain for Federated Learning in Smart Cities

Abstract

In smart cities, large amounts of multi-source data are generated all the time. A model established via machine learning can mine information from these data and enable many valuable applications. With concerns about data privacy, it is becoming increasingly difficult for the publishers of these applications to obtain users’ data, which hinders the previous paradigm of centralized training through collecting data on a large scale. Federated learning is expected to prevent the leakage of private data by allowing users to train models locally. The existing works generally ignore architectures designed in real scenarios. Thus, there still exist some challenges that have not yet been explored in federated learning applied in smart cities, such as avoiding sharing models with improper parties under privacy requirements and designing satisfactory incentive mechanisms. Therefore, we propose an efficient attribute-based participant selecting scheme to ensure that only someone who meets the requirements of the task publisher can participate in training under the premise of high privacy requirements, so as to improve the efficiency and avoid attacks. We further extend our scheme to encourage clients to take part in federated learning and provide an audit mechanism using a consortium blockchain. Finally, we present an in-depth discussion of the proposed scheme by comparing it to different methods. The results show that our scheme can improve the efficiency of federated learning by enabling reliable participant selection and promote the extensive use of federated learning in smart cities.