Fixing attention of CNNs for image classification

I’d like to fix attention of a CNN to a particular region of a given image.

For example, given 28x28 image, I’d like my model to only consider only the top-left 5x5 region.
What’s the proper way of doing this, if possible?

A naive work-around I can think of is that, I could overwrite the unwanted regions with black-pixels. Not sure if this is a nice solution though.

Would it work, if you just crop the input image to the desired shape and pass this small patch to the model?

Worth trying. Thanks.