Accuracy drops to zero

The security evaluation performed on the attached models returns accuracy values near to zero. In the worst case I expected 10% accuracy.
The models have 10 classes and they are trained on a subset of the google speech commands dataset (you can find it in the attachment).

Perturbation values: [0, 0.01, 0.02, 0.05, 0.1]

Archive: https://app.box.com/s/he5ng229qqulnekkxnnjeule2m8vehbh