Self-Attention-Based Multivariate Anomaly Detection for CPS Time Series Data with Adversarial Autoencoders

Data-driven anomaly detection continues to be challenging due to the increased complexity of modern cyber physical systems (CPSs) and their temporal dependencies. Unsupervised detection techniques are widely used through VAE-based frame-works and RNN-based deep learning techniques. However, VAE and its variants impose too much constraint on extracted latent code, and RNN's autoregressive essence indicates the shortage of parallelism and long-term prediction. To tackle the above is-sues, we propose TransAAE (Transformer-augmented Adversarial Autoencoder), a novel unsupervised approach for multivariate time series anomaly detection. The use of the adversarial autoencoder (AAE) architecture loosens the regularization of latent code, and self-attention mechanism is utilized to extract temporal information. Extensive experiments show an average F1 score over 0.9 on three public datasets, which significantly outperforms among the baselines.