This paper proposes an attention Unet model to detect noise in electrodermal activity (EDA). Three databases containing EDA signals collected from 78 participants, together with sample-based expert annotations, are used for training and performance evaluation. The results demonstrate that adding an attentional mechanism in the skip connections of the Unet improves performance. In addition, the proposed attentional model achieved a performance superior to the state of the art by achieving a kappa score of 56%, demonstrating the possibility of detecting noise at the sample level.