An Efficient Federated Learning Framework for Privacy-Preserving Data Aggregation in IoT

With the development of Internet of Things (IoT) technology, smart mobile devices are widely used in daily life. The service providers always extensively collect data from users for training machine learning models in order to improve their accurate services. This also raises users' concerns about data privacy and security. Federated learning, as an extension of centralized machine learning, allows several users working together to train a machine learning model on their own devices without sending their data to the centralized servers. However, existing research suggests that local models also contain privacy related to the users' data. Unfortunately, the current privacy-preserving secure aggregation methods have either poor accuracy or high computational and communication costs in training process which can not afford by the IoT devices. In this work, we propose a federated learning framework supporting privacy-preserving data aggregation against external and internal attackers with lower computational and communication costs, which is suitable for the weak IoT devices. The scheme is also supporting aggregation with fault tolerance and dynamic user set even if a part of users leave the system in the training. Detailed security analysis and extensive experiments using a real dataset confirm the efficacy and efficiency of the proposed schemes.