FedProLs: federated learning for IoT perception data prediction

With the development of Internet of Things, sensor devices collect massive amounts of data. However, due to privacy protection requirement, data cannot be shared and collected. How to integrate independent perception data into deep learning is one of the most challenging problems. In this paper, we present a novel framework (FedProLs) for IoT perception data prediction based on a horizontal federated learning model. The framework is constructed by the client nodes and the server nodes, and the training data of the federated learning system is deployed on the client nodes. Each client uses its own data to train machine learning models locally and encrypts its training model parameters and sends it to the server nodes. The server node uses the federated averaging method to construct a global model for prediction. In addition, we propose a new multi-feature factor model (ProLs) as a client-node machine learning model. Finally, the proposed FedProLs and ProLs models are compared with the single model Prophet, LSTM and BP Neural Networks, and combine model CNN-LSTM, ARIMA. The experimental results using two real-life IoT perception data sets demonstrate that the FedProLs and the participants' ProLs achieves better results in terms of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) than existing methods. The FedProLs model is suitable for distributed independent data protection when predicting the perception data of Internet of Things (IOT).