Privacy-Preserving Lightweight Time-Series Anomaly Detection for Resource-Limited Industrial IoT Edge Devices
Lei Chen, Yepeng Xu, Miao Li, Bowen Hu, Haomiao Guo, Zhaohua Liu
Abstract
Identifying anomalies directly on edge devices rather than in the cloud, known as edge computing, is essential for Industry 4.0. However, the limited computing and storage resources on edge devices render traditional cloud-based anomaly detection models ineffective. To solve this issue, a privacy-preserving lightweight time-series anomaly detection model, named PPLAD, is proposed for resource-limited industrial Internet of Things (IoT) edge devices via global and local similarity discrepancy. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">First</i>, PPLAD directly uses data similarity instead of raw data as model input, to achieve privacy-preserving. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Second</i>, PPLAD applies trainable Gaussian distribution rather than deep neural network as model structure, to achieve high timeliness and low cost. Specifically, PPLAD constructs a trainable Gaussian distribution with only one parameter for each timestamp to model its similarity with neighbors. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Third</i>, a global and local adversarial learning strategy is developed to amplify the discrepancy between local similarity and global similarity for each timestamp. Finally, the discrepancy is used to accurately identify timestamp-level anomalies. To the best of authors' knowledge, this is the first work to build an industrial anomaly detection model using only learnable Gaussian distributions. Extensive experiments on eight public industrial multisensor datasets and three edge devices demonstrate that PPLAD outperforms several state-of-the-art models.