Practical Vertical Federated Learning with Unsupervised Representation Learning

As societal concerns on data privacy recently increase, we have witnessed silos among multiple parties in various applications. Federated learning emerges as a new paradigm that enables to collaboratively train machine model without sharing their raw data. Vertical federated learning, where each party owns different features of the same set samples and only single has label, is an important challenging topic learning. Communication costs been major hurdle for practical vertical systems. In this paper, propose novel communication-efficient algorithm named FedOnce, which requires one-shot communication parties. To improve accuracy provide guarantee, FedOnce unsupervised representations setting privacy-preserving techniques based moments accountant. The comprehensive experiments 10 datasets demonstrate achieves close performance compared state-of-the-art algorithms with much lower costs. Meanwhile, our technique significantly outperforms approaches under budget.