Distortions of GNSS signals can lead to unacceptable pseudo-range errors. The object of study is a certain type of distortion — evil waveforms (EWF) — which is a rare perturbation occuring at the stage of signal generation. Detecting those distortions post-correlation traditionally involve designing hand crafted structure tests on a densely sampled autocorrelation function (ACF). However, traditional hand crafted tests have to be designed for specific scenarios hence lack flexibility compared to data-based methods. A neural network architecture capable of processing the structure of the ACF is proposed, implicitly learning structure tests, in order to tackle the evil waveforms detection problem.