Theses

Authority-Aware FHE Program Partitioning Compiler for Edge-Cloud Systems [abstract] (PDF)
Dongkwan Kim
Ph.D. Thesis, School of Electronical and Electronic Engineering, Yonsei University, August 2026.

Privacy-preserving edge-cloud tensor execution is difficult to develop without exposing intermediate values to unauthorized devices because user data and service-side model parameters often belong to different privacy authorities. Fully homomorphic encryption enables computation on encrypted data and allows a cloud server to process data that belong to the edge. Recently proposed FHE compilers support for deciding where encrypted execution is needed, translating tensor operators into ciphertext-level execution, and generating coordinated edge and cloud programs. Fully homomorphic encryption enables computation on encrypted data and allows a cloud server to process edge-owned data without accessing it in plaintext. However, they mainly focus on fully encrypted execution, while existing edge-cloud partitioning methods do not directly handle authority conflicts between private data owned by different parties.

This dissertation presents AION, an end-to-end authority-aware compiler for privacy-preserving edge-cloud tensor execution. AION allows developers to express tensor computation together with privacy intent through source-level placement annotations and authority annotations. From the annotated program, AION analyzes how privacy authority propagates through the program and identifies the values that must be protected during edge-cloud execution. AION lowers the protected tensor computation into FHE primitive operations using tensor-to-slot mapping information, and selects an edge-cloud partitioning plan by estimating encrypted computation and ciphertext communication costs. Finally, AION generates coordinated edge and cloud programs with the required encryption, decryption, and communication operations.

The evaluation shows that AION improves the efficiency and programmability of privacy-preserving edge-cloud tensor execution. AION achieves 4.92× and 3.69× geometric mean speedups over EVA and Hecate, respectively, while reducing edge and cloud memory usage by up to 65.5% and 59.1%. AION also reduces the burden of developers by replacing manual encrypted-region selection, communication insertion, encryption and decryption management, and tensor-to-FHE primitive rewriting with source-level annotations and compiler-generated edge-cloud programs.