A Study On Privacy-Preserving Transfer Learning

In the field of machine learning,the quality and quantity of datasets have a huge impact on the effectiveness of machine learning models.Transfer Learning is an effective method to handle the lack of data by leveraging data or knowledge from other sources.However,there are lots of data exchanges in the process of transfer learning.And it brings risks to the privacy of the participants of transfer learning.Hence this is an important problem to be solved about how to conduct effective transfer learning while preserving privacy.Our goal is to design a practical and efficient transfer learning algorithm which satisfies privacy preserving.In this paper,we leverage a public auxiliary dataset,and use importance weighting mechanism to determine the relationship between datasets from different sources and the public dataset.By using this way,we avoid the direct comparisons between sources.And we assign different weights to different sources to process transfer learning.At the same time,we use homomorphic encryption and secure multi-party computation to protect the data privacy.For the different content to transfer,we design different privacy preserving protocols for hypothesis transfer learning and network parameters transfer learning.We show these two protocols in detail using the logistic regression model as an example.We conduct various experiments on the proposed algorithms on real datasets.The experiments show that our algorithms can preserve the data privacy of participants while maintaining the accuracy of transfer learning at the cost of adding a small amount of time and communication overhead.