ReconGAN: Reconstruction-Based Anomaly Detection using Generative Adversarial Networks.

Abstract

In recent years, anomaly detection has become crucial for gaining insights into important and valuable information related to critical situations across various industries, including finance and IT. However, the vague nature of anomalies and their complex temporal correlations, combined with the common challenge of lacking labels in time series data, make it difficult to identify abnormalities effectively. Many approaches now exist to tackle anomaly detection, including a range of deep learning-based techniques for unsupervised anomaly identification. Despite advancements (Qin et al., 2017), contemporary state-of-the-art unsupervised machine learning methods for anomaly detection often face significant issues, such as high false positive rates and challenges related to scalability and portability. Recent advancements in deep learning have introduced sophisticated techniques, such as Generative Adversarial Networks (GANs), which can be utilized to identify irregularities in time series data. The proposed solution takes a reconstruction-based unsupervised approach using GANs to effectively reconstruct time signals and detect anomalies within them. This reconstruction-based method trains a model to capture a low-dimensional representation of time signals and aims to generate a synthetic reconstruction of the time series data. In this process, the GAN architecture is employed, where the Encoder maps the time series data into latent space, and the Decoder transforms this latent space back into the reconstructed time series signals. A critic is used to assess the differences between the real and generated time series, evaluating the effectiveness of the mapping. This solution utilizes a custom LSTM cell with peephole connections for both the encoder and decoder. Initial experiments conducted on the NASA Telemetry datasets, with an average of 100 epochs, demonstrated that the proposed model was able to efficiently reconstruct time signals, achieving low Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) values, and it was compared favorably with existing approaches