Run model in reverse in Keras

There is no such thing as "running a neural net in reverse", as a generic architecture of neural net does not define any not-forward data processing. There is, however, a subclass of models which do - the generative models, which are not a part of keras right now. The only thing you can do is to create a network which somehow "simulates" the generative process you are interested in. But this is paricular model specific method, and has no general solution.