Functional encryption based approaches for practical privacy-preserving machine learning

Machine learning (ML) is increasingly being used in a wide variety of application domains. However, deploying ML solutions poses a significant challenge because of increasing privacy concerns, and requirements imposed by privacy-related regulations. To tackle serious privacy concerns in ML-based applications, significant recent research efforts have focused on developing privacy-preserving ML (PPML) approaches by integrating into ML pipeline existing anonymization mechanisms or emerging privacy protection approaches such as differential privacy, secure computation, and other architectural frameworks. While promising, existing secure computation based approaches, however, have significant computational efficiency issues and hence, are not practical. In this dissertation, we address several challenges related to PPML and propose practical secure computation based approaches to solve them. We consider both two-tier cloud-based and three-tier hybrid cloud-edge based PPML architectures and address both emerging deep learning models and federated learning approaches. The proposed approaches enable us to outsource data or update a locally trained model in a privacy-preserving manner by employing computation over encrypted datasets or local models. Our proposed secure computation solutions are based on functional encryption (FE) techniques. Evaluation of the proposed approaches shows that they are efficient and more practical than existing approaches, and provide strong privacy guarantees. We also address issues related to the trustworthiness of various entities within the proposed PPML infrastructures. This includes a third-party authority (TPA) which plays a critical role in the proposed FE-based PPML solutions, and cloud service providers. To ensure that such entities can be trusted, we propose a transparency and accountability framework using blockchain. We show that the proposed transparency framework is effective and guarantees security properties. Experimental evaluation shows that the proposed framework is efficient.