Anomaly Detection-Reconstruction Trade-off in Autoencoder-based Compression

Monitoring systems generate and transmit large volumes of data to facilities capable of effectively performing multiple tasks. Data is often compressed and autoencoders have emerged as a promising neural network-based approach. This work focuses on a scenario in which the receiver performs reconstruction and anomaly detection tasks. We examine how two autoencoder-based compression strategies administer the trade-off between reconstruction and anomaly detection. The experiments consider two scenarios: ECG time series and CIFAR-10 images. Each dataset is corrupted by five anomalies with different intensities and assessed with two detectors. We highlight the pros and cons of the two approaches showing that that their efficacy depends on a specific anomaly and setting.