A Secure Algorithm for Deep Learning Training under GAN Attacks

Abstract

Deep neural networks have outperformed traditional machine learning approaches for many tasks, and are the tool of choice in many fields. However, directly applying these techniques in fields that deal with private data is challenging. The reason is that a third-party, which organizations may not trust, usually needs to centrally collect private data. To overcome this challenge, researches have proposed distributed training algorithms that allow multiple users to collaboratively train their local deep learning models without sharing private datasets. However, these approaches are vulnerable to recently proposed attacks where a malicious user can replicate private data from another user by compromising the collaborative training algorithm. In this paper, we propose a privacy-preserving distributed deep learning algorithm that allows a user to leverage the private datasets from a group of users while protecting its privacy. Our algorithm prohibits this user from ever sharing the parameters of its model, and thus it prevents malicious users from compromising the training and replicating the user's private data. We conduct extensive experiments and observe that our algorithm can achieve a model accuracy of 95.18 %, which is the same accuracy that previous approaches that are vulnerable to attacks can achieve.