Performance analysis of various machine learning models for membership inference attack

摘要

In order to function correctly during the training phase, many ML models require enormous amounts of labelled data. There is a possibility that the data will contain private information, which must be protected regarding privacy. Membership inference attacks (MIA) are attacks that try to identify if a target data point was utilised for training a particular ML method. These attacks have the potential to compromise users' privacy and security. The degree to which an algorithm for ML divulges user membership information varies from implementation to implementation. Hence, a performance analysis was performed based on different ML algorithms under MIA inference attacks. This study proposed for comparing different ML approaches against MIAs and analyses which ML algorithm is better performing to such privacy attacks. Based on the performance analysis observation, the GAN and DNN models are considered as the best ML models to defend against MIA attacks with better performances.