Real-time Anomaly Detection at IoT-Edge Ingress Port using FPGA based ML Classifiers

Edge based architectures are being widely used for various Internet of Things (IoT) verticals like factories, transportation, healthcare, etc. Intelligent edges have the advantages of quick reaction time, lesser data bandwidth requirements for cloud connectivity, and privacy because of localization of data. Edge based IoT deployments make detection of anomalies in the ingress port which connects to the things, faster compared to the cloud based anomaly detectors. These anomalies can be due to security attacks, malfunction of equipment, degradation of sensor performance, or actual change in physical system. For example, in the case of smart healthcare with remote health monitoring, the anomalies can be due to variation in the patients’ vitals being monitored. It is very important to detect anomalies in real time with minimal latency as it dictates the response time for an appropriate action. Achieving low latencies in large scale IoT systems, wherein large number of sensors need to be monitored continuously, using only software, is challenging. In this paper a real-time anomaly detector architecture, based on a set of Logistic Regression (LR) Machine Learning (ML) classifiers in a Field Programmable Gate Array (FPGA) is proposed. This anomaly detector monitors the ingress traffic at the IoT edge, to predict the anomalies with minimal latency. Hardware-software partitioning used to achieve the real-time performance has been described. A novel LR twin based retraining mechanism, which updates the weights of the set of classifiers, with minimal disruption of real-time operation is discussed. The proposed system has been developed on PYNQ Z2 board for a set of 7 LRs. The system feasibility and timing analysis have been presented, highlighting the performance advantages as compared to pure software implementation.