Edge Computing-Enabled Secure Federated Learning for Anomaly Detection in Industrial IoT Sensor Streams
TL;DR
A federated learning framework to collaboratively develop an anomaly detection model from multiple edge nodes, without sharing the raw sensor data, so that data sovereignty is preserved, and demonstrates that practical, deployable security for federated learning in IIoT is achievable without sacrificing accuracy or responsiveness.
Abstract
The rise of the Industrial IoT (IIoT) will result in a surge of IIoT devices with high-velocity data streams requiring rapid, real-time analysis of these data streams to power predictive maintenance and assure cybersecurity. Centralized cloud-based approaches to anomaly detection are hindered by their inherent latency and privacy issues while independent or stand-alone approaches to edge-based anomaly detection do not have enough data to effectively detect anomalies. This paper presents a federated learning framework to collaboratively develop an anomaly detection model from multiple edge nodes, without sharing the raw sensor data, so that data sovereignty is preserved. Two major contributions of this research include a lightweight hybrid secure aggregation method that utilizes pairwise additive masking and differential privacy, which mitigates the threat of inversion attacks achieved via a reconstruction SSIM < 0.05, and successfully detects >90% of model poisoning attempts, where as traditional homomorphic encryption solutions incur excessively high computational costs (approximately 68 ms/node) and 70 KB of communication overhead/round). Our architecture was validated using a synthetic IIoT data set containing 600,000 rows of data correlated across 12 edge nodes and 10 different types of sensors, and further validated using the real-world SWaT data set. The experimental results demonstrated the effectiveness of the framework, achieving an F1 score of 0.944 for anomaly detection, an F2 from centralized training of only 2.3%, and a 60-70% reduction in communication costs compared to the use of homomorphic encryption. The experimental results also demonstrate that local edge inference latency is <50 ms, meeting the real-time requirements of IIoT systems. Overall, the framework demonstrates that practical, deployable security for federated learning in IIoT is achievable without sacrificing accuracy or responsiveness, and open-source implementations are provided to ensure full reproducibility.