GRAND is a Ultra High Energy (UHE) cosmic particles detection project, consisting in a giant,
self-triggered, antenna array. Wherever we decide to setup the antennas apart from polar areas,
we will generally face a high rate of background signals, orders of magnitude higher than the rate
of extensive air showers. To avoid the saturation of the acquisition, we need to reject a significant
part of the data at the antenna level, with a more sophisticated method than a basic peak-over baseline selection. We present here an attempt to discriminate air showers and background radio
time traces with a convolutional neural network, using experimental data rather than simulations.
These data were produced with TREND, a self-triggered 50-antennas array, which was the seed
for the GRAND project. At the antenna-level, for a set of signals that had triggered TREND, it
was possible to reject 82% of the ultra-dominant background, while preserving 86% of the air
shower signals.