Towards Secure and Efficient Outsourcing of Machine Learning Classification

Abstract

Machine learning classification has been successfully applied in numerous applications, such as healthcare, finance, and more. Outsourcing classification services to the cloud has become an intriguing practice as this brings many prominent benefits like ease of management and scalability. Such outsourcing, however, raises critical privacy concerns to both the machine learning model provider and the client interested in using the classification service. In this paper, we focus on classification outsourcing with decision trees, one of the most popular classifiers. We propose for the first time a secure framework allowing decision tree based classification outsourcing while maintaining the confidentiality of the provider’s model (parameters) and the client’s input feature vector. Our framework requires no interaction from the provider and the client—they can go offline after the initial submission of their respective encrypted inputs to the cloud. This is a distinct advantage over prior art for practical deployment, as they all work under the client-provider setting where synchronous online interactions between the provider and client is required. Leveraging the lightweight additive secret sharing technique, we build our protocol from the ground up to enable secure and efficient outsourcing of decision tree evaluation, tailored to address the challenges posed by secure in-the-cloud dealing with versatile components including input feature selection, decision node evaluation, path evaluation, and classification generation. Through evaluation we show the practical performance of our design, and the substantial client-side savings over prior art, say up to four orders of magnitude in computation and 163$\times$ in communication.