mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
44 lines
2.1 KiB
TeX
44 lines
2.1 KiB
TeX
\section{Introduction}
|
|
|
|
\label{sec:intro}
|
|
% ============================================================================
|
|
|
|
Every major AI cloud platform operates on the same trust model: the user sends
|
|
plaintext data to the provider, the provider runs a model on it, the provider
|
|
returns a plaintext result. The provider sees everything. The provider's
|
|
privacy policy is the only barrier between the user's data and the provider's
|
|
training pipeline, the provider's employees, the provider's government
|
|
subpoenas, and the provider's security breaches.
|
|
|
|
This is not a theoretical concern. OpenAI's data retention policy permits
|
|
logging all API inputs for 30 days~\cite{openaiapi}. Anthropic retains inputs
|
|
for safety evaluation. Replicate processes inputs on shared GPU instances with
|
|
no hardware isolation between tenants. The user has no cryptographic guarantee
|
|
that their prompt---containing patient records, financial models, legal
|
|
strategy, classified intelligence---is not stored, read, or reused.
|
|
|
|
Hanzo AI Cloud solves this problem at the architectural level, not the policy
|
|
level. The system is designed so that the provider \emph{cannot} see the
|
|
user's data, not that the provider \emph{promises} not to.
|
|
|
|
\subsection{Contributions}
|
|
|
|
\begin{enumerate}[leftmargin=1.4em]
|
|
\item Confidential inference via CKKS fully homomorphic encryption with
|
|
threshold decryption (\S\ref{sec:fhe}).
|
|
\item Agent hosting with capsule isolation: vault, DID, capabilities,
|
|
budget, and private memory (\S\ref{sec:agents}).
|
|
\item Model serving with verifiable inference receipts anchored to the
|
|
Lux A-Chain (\S\ref{sec:serving}).
|
|
\item Hanzo Candle: a Rust ML framework with GPU acceleration integrated
|
|
into Base Functions (\S\ref{sec:candle}).
|
|
\item Jin architecture: unified multimodal inference across vision,
|
|
language, and audio (\S\ref{sec:jin}).
|
|
\item Edge inference: local models on mobile/IoT with CRDT result sync
|
|
(\S\ref{sec:edge}).
|
|
\item Provider marketplace with stake-weighted selection and on-chain SLA
|
|
enforcement (\S\ref{sec:marketplace}).
|
|
\end{enumerate}
|
|
|
|
% ============================================================================
|