Hide and Seek: on the Stealthiness of Attacks against Deep Learning Systems