Radiation Anomaly Detection Using an Adversarial Autoencoder

Scintillators are the primary devices used for radiation detection., especially at national borders and other ports of entry. Reducing the size of these detectors for placement on mobile devices such as drones can allow for better detection and localization of radiation sources. Detection of radiation with supervised machine learning can be a challenge when looking for previously unobserved radiation sources. Therefore, anomaly detection methods are investigated. In this work we employ adversarial autoencoders trained to classify spectra from radioactive sources as either background or anomalous. This allows the model to detect anomalies regardless of radiation source and outperform supervised methods on newly encountered sources.