Extending patch based image classification into image classification

I am trying to classify tampered, pristine images from set of images, in that I have built a network in which I would divide the image into multiple overlapping patches and then classify them into pristine or fake(based on the probability outputs), but now I want extend the same to Image level. That is I want to build some model or some rule over output probabilities of patches of each image to get probability that the image is fake or pristine.

ways I am thinking to do is –

1. Build a shallow network over the probabilities of the patch probabilities. In this case problem is all images are of different shape
2. Apply a ML classifier (something like Logistic Regression), with output probabilities by appending zeros to the output probability vector generated so that all image has same sized probability vector as input
3. generate a mask using patches and then build a simple classification network over the masks using original image labels.

I can’t really say which among the above three is better or worse, I don’t even know the possibility of the above three. (Kind of hit a roadblock in thinking)

Now the question am I thinking in right direction, what would be better among the ideas I am considering and why. Is there anything better than what I am thinking. It would helpful in suggesting some resources.