Beyond Black Box Densities: Parameter Learning for the Deviated Components