Federated learning for performance behavior detection in a fog-IoT system

Abstract

In fog-based IoT systems, every fog can behave differently due to context and vulnerabilities. In smart irrigation systems, some fog nodes use more or fewer computing resources to analyze the data according to sensor location, soil moisture, plant species, or seasons. Therefore, fog behavior should consider distinct contexts. Federated learning support fog-based IoT systems to detect faster the behavior of fog nodes as it enables them to perceive previous behaviors from their peer nodes. We develop and assess an unsupervised federated learning system to identify fog anomalies. We consider experiments with seven rounds of four minutes, executing K-Means in every node to obtain local centroids, and the system merges them in the cloud to calculate global centroids, sending them back to the fog nodes. This paper evaluates the accuracy and time a fog node needs to predict a behavior already identified by another fog node. We assess the CPU usage and the time the cloud takes to compute global centroids using thousands of local cluster centers and measure the prediction time for different fog hardware. We observe that the cloud CPU usage and time to obtain the global centroids vary according to the number of fog nodes and the number of fog behaviors. Our results also show that, in the worst case, our system predicts a behavior by around 50 ms. In contrast, a non-federated approach must wait for the current round to end, as 51.3 s in our results. Therefore, our approach shows promising results for time-sensitive IoT systems.