SoK: Can Fully Homomorphic Encryption Support General AI Computation? A Functional and Cost Analysis

Artificial intelligence (AI) increasingly powers sensitive applications in domains such as healthcare and finance, relying on both linear operations (e.g., matrix multiplications in large language models) and non-linear operations (e.g., sorting in retrieval-augmented generation). Fully homomorphic encryption (FHE) has emerged as a promising tool for privacy-preserving computation, but it remains unclear whether existing methods can support the full spectrum of AI workloads that combine these operations. In this SoK, we ask: Can FHE support general AI computation? We provide both a functional analysis and a cost analysis. First, we categorize ten distinct FHE approaches and evaluate their ability to support general computation. We then identify three promising candidates and benchmark workloads that mix linear and non-linear operations across different bit lengths and SIMD parallelization settings. Finally, we evaluate five real-world, privacy-sensitive AI applications that instantiate these workloads. Our results quantify the costs of achieving general computation in FHE and offer practical guidance on selecting FHE methods that best fit specific AI application requirements. Our codes are available at https://github.com/UCF-ML-Research/FHE-AI-Generality.

💡 Research Summary

The research paper, “SoK: Can Fully Homomorphic Encryption Support General AI Computation? A Functional and Cost Analysis,” presents a systematic investigation into the feasibility of using Fully Homomorphic Encryption (FHE) to execute the diverse range of operations required by modern Artificial Intelligence. As AI integration expands into highly regulated sectors like healthcare and finance, the demand for privacy-preserving computation has become critical. While FHE is theoretically capable of performing computations on encrypted data, the complexity of modern AI workloads—which blend linear operations (such as matrix multiplications in LLMs) with non-linear operations (such as sorting in Retrieval-Augmented Generation)—poses a significant challenge to its practical implementation.

The authors adopt a “Systematization of Knowledge” (SoK) approach to address this uncertainty. The study is structured into three primary analytical phases. First, a functional analysis was conducted, where the researchers categorized ten distinct FHE approaches. This categorization allowed them to evaluate the mathematical capability of each scheme to handle the heterogeneous operations that constitute “general” AI computation. By distinguishing between the relative ease of linear operations and the computational difficulty of non-linear tasks, the paper identifies the fundamental bottlenecks in FHE-based AI.

Second, the paper performs a rigorous cost analysis. After identifying three promising FHE candidates, the researchers benchmarked workloads that mix linear and non-linear operations. Crucially, they examined these workloads across varying bit lengths (representing security levels) and SIMD (Single Instruction, Multiple Data) parallelization settings (representing computational efficiency). This allows for a granular understanding of how increasing security or attempting to optimize throughput affects the overall computational overhead and latency.

Third, the study validates these findings through real-world application testing. The researchers instantiated five privacy-sensitive AI applications, evaluating how the identified FHE methods perform in practical, high-stakes scenarios. This application-driven approach moves the discussion from theoretical possibility to practical engineering.

In conclusion, the paper provides a quantitative framework for understanding the “cost of generality” in FHE. By quantifying the computational penalties associated with non-linear operations and providing a selection guide based on specific application requirements, this work serves as a vital roadmap for developers building the next generation of secure, privacy-preserving AI systems. The findings offer practical guidance on balancing the trade-offs between security, functionality, and performance in the pursuit of general-purpose encrypted AI.