Collaborative machine learning allows multiple participants to get a global and valuable insight over their joint data. Nonetheless, in data-sensitive applications, it is crucial to maintain confidentiality across the end-to-end path the data follows from model training phase to the inference phase, preventing any form of information leakage about training data, the learned model, or the inference queries. In this paper, we present our approach to address this problem through PrivML, a framework for end-to-end outsourced privacy-preserving data classification over encrypted data. We provide some preliminary results comparing our proposal with state of the art solutions as well as some insight on our prospective research plan.