Towards Generalized and Distributed Privacy-Preserving Representation Learning