GPU Accelerated Fully Homomorphic Encryption Framework for End-to-End Logistic Regression

초록

Machine Learning as a Service (MLaaS) platforms are increasingly adopted in regulated domains such as finance and healthcare, where protecting sensitive data and model parameters is paramount. Fully Homomorphic Encryption (FHE) offers rigorous privacy guarantees by enabling computation directly on encrypted data; however, its high computational overhead has limited practical deployment. Prior research has primarily focused on optimizing encrypted deep learning inference, leaving efficient FHE-based training largely unexplored. This paper addresses this gap by presenting a GPU-accelerated CUDA framework for FHE logistic regression models with full 128-bit security. Logistic regression is selected as a foundational case study because it incorporates a nonlinear activation function, making it an essential building block for extending FHE operations to neural networks. Our implementation executes all computations directly over ciphertexts, offloading heavy FHE operations, such as homomorphic addition, multiplication, and bootstrapping, to optimized CUDA GPU kernels. By leveraging GPU parallelism, the framework achieves significant efficiency gains while maintaining strong privacy guarantees. Our open-source design abstracts the complexities of FHE and CUDA, lowering the technical barrier and enabling practitioners to perform privacy-preserving model training without deep cryptographic or GPU programming expertise.

키워드