Towards Characterizing Model Extraction Queries and How to Detect Them