Using Ordered Learning to Mitigate Privacy Risks of Neural Networks

Abstract

Despite being ubiquitous, machine learning algorithms and services were not initially designed with privacy and security in mind. Therefore, models are prone to revealing information about their training data. In this paper, ordered learning is proposed as a defense against black-box membership inference attacks in which an adversary wants to infer whether data was used to train a model. Ordered learning consists of introducing training data to a model in a specific order, at a specific pace, defined by a scoring and pacing function. The results show that this defense is competitive and comparable to MemGuard and adversarial regularization while boosting utility. This work is the first to use ordered learning as a defense against privacy risks to show that privacy and utility can be improved jointly.