A privacy-preserving and correctness audit method in multi-party data sharing
Junping Wan, Hao Xun, Xianrong Zhang, Jiyuan Feng, Zhe Sun · 2020
In general, deep learning can perform better if much training data is introduced. Recently, as a technology that uses large amounts of data, federated learning is more and more popular, where multiple parties cooperate in training a common model. However, data sharing in cooperative training raises the risk of privacy leakage. Besides, the encryption applied in federated learning may cause the participants to behave maliciously. To solve this problem, we design a practical framework based on blockchain to protect privacy and prevent malicious behavior. If anyone tries to cheat in cooperative learning, his behavior is discovered in the audit process. Our framework is composed of several message transmission processes. The transmitted content contains the evaluation and correctness verification of participants and so on. We conduct experiments on a dataset of messages and verify that our scheme works.