Improving the Efficiency and Security of Fully Homomorphic Machine Learning as a Service

Abstract

Machine learning (ML) services have become increasingly prevalent for both personal and business use. However, despite the widespread adoption of these services, privacy concerns have remained a significant barrier to further integration as an honest-but-curious cloud service can potentially misuse it for commercial gain. Fully Homomorphic Encryption (FHE) provides a solution by allowing computations to be performed directly on encrypted data without exposing its content. However, FHE is resource-constrained, requiring careful consideration of available operations, ciphertext precision, storage size, and computation latency.

This thesis presents substantial progress toward enhancing the efficiency, accuracy, and security of FHE-based machine learning inference and its applications using the TFHE cryptosystem. We first present a set of works enhancing and evaluating binary neural network, multi-exit neural network, and split model architectures. Additionally, this thesis provides a detailed analysis of the security of these new efficiency-improved models hosted in the cloud, both in terms of user privacy protection and the security of the model intellectual property.

Using the lessons learned from developing these encrypted machine learning architectures, the following chapters present new applications and FHE protocols that balance the privacy-efficiency tradeoffs in FHE. The first of these applications is an FHE-encrypted generative AI model, exploring the problem of encrypted probabilistic selection. Next, the dissertation presents a protocol designed for authenticated data storage and private data verification, with applications in private content moderation and biometric authentication.

Together, these contributions advance the practical deployment of privacy-preserving machine learning and strengthen the foundation for secure, real-world FHE applications.