Resource-Efficient Anomaly Detection in Industrial Control Systems With Quantized Recurrent Variational Autoencoder

Abstract

This work presents a novel solution for multivariate time series anomaly detection in industrial control systems (ICSs), specifically tailored for resource-constrained environments. At its core, the quantized gated recurrent unit variational autoencoder (Q-GRU-VAE) architecture, a significant evolution from conventional methods, offers an extremely lightweight yet highly effective solution. By integrating gated recurrent units (GRUs) in place of long short-term memory (LSTM) cells within a variational autoencoder (VAE) framework, and employing channel-wise dynamic post-training quantization (DPTQ), this model dramatically reduces hardware resource demands. The proposed solution exhibits performance on par with existing methods on the widely used secure water treatment (SWaT) and water distribution (WADI) benchmarks, while being tailored towards applications where computational resources are limited. This dual achievement of minimal resource consumption and preserved model efficacy paves the way for deploying advanced anomaly detection in resource-constrained environments, marking a significant leap forward in enhancing the resilience and efficiency of ICSs.