Is there a way to change/override the expected loss function of Foolbox's 'attack' function?

Hi,
I am currently trying to generate adversarial images of a CIFAR10 dataset that can fool a CNN in a simple color estimation task.

As a little background for this color estimation task, I created a CNN to estimate the average global color of an image as a 3d column vector in terms of [average red channel, average blue channel, average green channel]. To train this CNN, I used a relabeled CIFAR10 dataset (where each label is no longer a number denoting it's class but a 3d vector of the average color) and MSE as my loss function.

But now, when I tried to use foolbox's _attack_ function
`raw_advs, clipped_advs, success = attack(fmodel, images, labels, epsilons=epsilons)`
to create an adversarial image, I ran into this error:


From this error trace, it seems that the _attack_ function is automatically geared toward object recognition models that use cross entropy loss and output a 1d label.

So my question is: **is there any way to modify the _attack_ function's expected loss function to be Mean Square Error instead of Cross Entropy? Or perhaps I should not use the built-in _attack_ function, and instead generate the adversarial images manually?**
Thank you for any thoughts or tips.