Unsupervised detection of offending radio transmissions by means of a deep learning autoencoder

With the arrival of deep learning, it is possible for the authorities of each country to automate surveillance tasks in the use of the radio spectrum. We present a method for monitoring radio channels in time and frequency based on an autoencoder trained with sequences that are randomly generated as Rayleigh-distributed signals. A receiver with an omnidirectional antenna records electric field amplitude samples on a set of channels belonging to the terrestrial or maritime mobile service. The samples are fed into the autoencoder, generating a prediction. When the prediction differs greatly from the original samples, we infer that a radio emission has occurred. This method has been successfully used to detect unauthorized radio emissions.