It's a difficult problem area. I would bet naively normalizing the data would introduce many new concerns. Does the amount of crimes reported scale linearly with police presence? If not, your data is now going to be skewed. How do you count police numbers in an area? Man hours, individual officers, or something else? Keep in mind that not all officers are going to be responding to the same type of crime and so police hours are not interchangeable.

IMO machine learning algorithms are brittle and prone to breaking in unexpected ways. In the past, organizations who should have known better and should have been smart enough have produced bad machine learning systems. Policing is a highly sensitive area. It's okay to be very concerned.