mirror of
https://github.com/hanzoai/papers.git
synced 2026-11-03 02:27:30 +00:00
zeekay
f9db4f5737
Bylines, paper authorship, correspondence addresses and the founder blurb under each post named an individual. They name the organisation that published the work instead: the brand on posts and pages, the research byline on papers, and research@ in place of a personal address. Both are what these repositories already used elsewhere — nothing new was coined. Co-authors are untouched. Where a paper named someone else alongside, they stay named; only the one byline changed. Bibliography entries, running heads, PDF metadata and affiliations move with it, so a citation still resolves to whoever the cited paper names. Sample and fixture data uses the placeholder name these tests already use elsewhere, so no assertion changed meaning.
78 lines
3.0 KiB
TeX
78 lines
3.0 KiB
TeX
% Hanzo FHE Inference: Fully Homomorphic Encryption for Confidential AI
|
|
\documentclass[11pt]{article}
|
|
\usepackage{../shared/hanzocover}
|
|
\usepackage{amsmath, amssymb, amsthm}
|
|
\usepackage{mathtools}
|
|
\usepackage{bm}
|
|
\usepackage{graphicx}
|
|
\usepackage{booktabs}
|
|
\usepackage{hyperref}
|
|
\usepackage{enumitem}
|
|
\usepackage{algorithm}
|
|
\usepackage{algpseudocode}
|
|
\usepackage{xcolor}
|
|
\usepackage{float}
|
|
\usepackage{multirow}
|
|
|
|
\hypersetup{colorlinks=true,linkcolor=black,citecolor=blue,urlcolor=blue}
|
|
|
|
\newtheorem{theorem}{Theorem}
|
|
\newtheorem{lemma}[theorem]{Lemma}
|
|
\newtheorem{proposition}[theorem]{Proposition}
|
|
\newtheorem{corollary}[theorem]{Corollary}
|
|
\newtheorem{definition}{Definition}
|
|
\newtheorem{remark}{Remark}
|
|
|
|
\title{HANZO FHE INFERENCE\\[0.3em]
|
|
{\large Fully Homomorphic Encryption for\\Confidential AI Inference}}
|
|
\author{
|
|
Hanzo AI Research \\
|
|
\textit{Hanzo Industries Inc (Techstars '17)} \\
|
|
\texttt{[email protected]}
|
|
}
|
|
\date{March 2026}
|
|
|
|
\begin{document}
|
|
|
|
\hanzocoverpage
|
|
|
|
% ============================================================================
|
|
% ABSTRACT
|
|
% ============================================================================
|
|
\begin{abstract}
|
|
We present Hanzo FHE Inference, a system for running machine learning models on
|
|
encrypted data using Fully Homomorphic Encryption (FHE). The system enables
|
|
privacy-preserving AI-as-a-service: clients encrypt their inputs locally, the
|
|
Hanzo Cloud evaluates AI models on ciphertexts without decryption, and only the
|
|
client can decrypt the result. We implement two FHE schemes optimized for
|
|
different model architectures: (i)~CKKS (Cheon--Kim--Kim--Song) for neural
|
|
network inference, supporting approximate fixed-point arithmetic on encrypted
|
|
vectors with SIMD batching, and (ii)~TFHE (Torus FHE) for decision tree and
|
|
gradient-boosted model evaluation, supporting exact Boolean and integer
|
|
operations with programmable bootstrapping. We describe GPU acceleration of FHE
|
|
operations using the Candle runtime~\cite{hanzocandle}, achieving 14$\times$
|
|
speedup over CPU-only evaluation on Number Theoretic Transform (NTT) kernels.
|
|
Production deployment on Hanzo ACI~\cite{hanzoaci} demonstrates encrypted
|
|
inference on a 4-layer neural network in 1.8 seconds (CKKS) and encrypted
|
|
XGBoost evaluation with 500 trees in 0.9 seconds (TFHE), enabling practical
|
|
confidential AI for healthcare, finance, and regulated industries. The system
|
|
integrates with the Lux TFHE infrastructure~\cite{luxtfhe} for on-chain
|
|
verification of encrypted computation proofs.
|
|
\end{abstract}
|
|
|
|
% ============================================================================
|
|
|
|
\input{sections/01-introduction}
|
|
\input{sections/02-fhe-background}
|
|
\input{sections/03-scheme-selection}
|
|
\input{sections/04-gpu-acceleration}
|
|
\input{sections/05-encrypted-neural-network-inference}
|
|
\input{sections/06-encrypted-decision-tree-inference}
|
|
\input{sections/07-integration-with-hanzo-and-lux-infrastru}
|
|
\input{sections/08-production-results}
|
|
\input{sections/09-security-analysis}
|
|
\input{sections/10-related-work}
|
|
\input{sections/99-conclusion}
|
|
|
|
\end{document}
|