Privacy-Preserving Distributed Learning in the Analog Domain