Hi Daniele,
From your description, I understand that you are trying to achieve Image regression. In the example script that you pointed to “Train Image classification network robust to adversarial examples”, I suggest you to modify the CNN network by removing Softmax layer and add a FullyConnectedLayer with n inputs and single output so that you get a number as output from your network.
Also, depending upon the range on training labels (the numbers associated with an image), if the output is expected to be between 0 and 1, you may use sigmoid function as final layer. If output is expected to be between 0 and any number N, you may use max(0, N) which is ReLu.
Also, do refer the following answer to have training labels as categorical input-
I assume that you already have adversarial examples in your dataset, the trained network should be robust to adversarial examples.
