학술논문
DC-VAE, Fine-grained Anomaly Detection in Multivariate Time-Series with Dilated Convolutions and Variational Auto Encoders
Document Type
Conference
Author
Source
2022 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW) EUROSPW Security and Privacy Workshops (EuroS&PW), 2022 IEEE European Symposium on. :287-293 Jun, 2022
Subject
Language
ISSN
2768-0657
Abstract
Due to its unsupervised nature, anomaly detection plays a central role in cybersecurity, in particular on the detection of unknown attacks. A major source of cybersecurity data comes in the form of multivariate time-series (MTS), representing the temporal evolution of multiple, usually correlated measurements. Despite the many approaches available in the literature for time-series anomaly detection, the automatic detection of abnormal events in MTS remains a complex problem. In this paper we introduce DC-VAE, a novel approach to anomaly detection in MTS, leveraging convolutional neural networks (CNNs) and variational auto encoders (VAEs). DC-VAE detects anomalies in time-series data, exploiting temporal information without sacrificing computational and memory resources. In particular, instead of using recursive neural networks, large causal filters, or many layers, DC-VAE relies on dilated convolutions (dc) to capture long and short term phenomena in the data, avoiding complex and less-efficient deep architectures, simplifying learning. We evaluate dc-vae on the detection of anoma-lies on a large-scale, multi-dimensional network monitoring dataset collected at an operational mobile internet service provider (isp), where anomalous events were manually labeled during a time span of 7-months, at a five-minutes granularity. Results show the main properties and advantages introduced by VAEs for time-series anomaly detection, as well as the out-performance of dilated convolutions as compared to standard VAEs for time-series modeling.