I have an input of 288 bits (comprising 4 × 32-bit identity function outputs and 10 × 16-bit integers). I need to hash this to 96 bits with as few collisions as possible. The goal could be stated as key compression with probabilistic collisions.
I'm aware that CRC is a bijective hash, thus ensuring 100% even distribution (as I understand it). In my view, I should be able to run 3 parallel CRC paths through the input, resulting in a 96-bit lossy hash (obviously not bijective) of optimum distribution.
However, I'm also aware that CRC is not used for such applications. An algorithm such as MetroHash would typically be used.
Could someone explain to me why CRC is a bad (or not) idea for this application?
Note: This is not intended for anything secure.