Fog-based Federated Time Series Forecasting for IoT Data

Abstract

Federated learning (FL) allows multiple nodes or clients to train a model collaboratively without actual sharing of data. Thus, FL avoids data privacy leakage by keeping the data locally at the clients. Fog computing is a natural fit for decentralized FL where local training can take place at fog nodes using the data of connected Internet of Things (IoT) or edge devices. A cloud-based node can act as the server for global model updates. Although FL has been utilized in fog and edge computing for a few applications, its efficacy has been demonstrated majorly for independent and identically distributed (IID) data. However, real-world IoT applications are generally time-series (TS) data and non-IID. Since there has not been any significant effort towards using FL for non-IID time-series data, this paper presents a fog-based decentralized methodology for time series forecasting utilizing Federated Learning. The efficacy of the proposed methodology for the non-IID data is evaluated using a FL framework Flower. It is observed that the FL based TS forecasting performs at par with a centralized method for the same and yields promising results even when the data exhibits quantity skew. Additionally, the FL based method does not require sharing of data and hence, decreases the network load and preserves client privacy.