papers/hanzo-fhe-inference/sections/01-introduction.tex

\section{Introduction}

\label{sec:intro}
% ============================================================================

AI-as-a-service requires clients to send sensitive data---medical records,
financial transactions, legal documents---to cloud infrastructure for inference.
This creates a fundamental privacy conflict: the service provider must see the
data to process it, but the client must not reveal the data to the provider.

Existing approaches have significant limitations:
\begin{enumerate}[leftmargin=1.4em]
    \item \textbf{Trusted execution environments (TEEs)}: Hardware enclaves
    (Intel SGX, AMD SEV) protect data in use but rely on hardware vendor trust
    and are vulnerable to side-channel attacks~\cite{sgxattacks}.
    \item \textbf{Differential privacy}: Adds noise to outputs, protecting
    individual records but degrading model quality. Does not protect the input
    data itself during computation.
    \item \textbf{Secure multi-party computation (MPC)}: Requires communication
    between parties during computation, introducing latency proportional to
    circuit depth.
    \item \textbf{Federated learning}: Keeps data local during training but
    does not address the inference privacy problem.
\end{enumerate}

FHE eliminates this conflict entirely: computation occurs on ciphertexts, the
server never sees plaintext data, and correctness is mathematically guaranteed.
The historical barrier---computational cost---has been reduced by four orders
of magnitude through algorithmic improvements and GPU
acceleration~\cite{fhesurvey2024}.

\subsection{Contributions}

\begin{enumerate}[leftmargin=1.4em]
    \item A production FHE inference system supporting both CKKS and TFHE
    schemes, with automatic scheme selection based on model architecture
    (\S\ref{sec:schemes}).
    \item GPU-accelerated NTT and bootstrapping kernels implemented in the
    Candle framework, achieving 14$\times$ speedup over CPU baselines
    (\S\ref{sec:gpu}).
    \item Integration with Hanzo ACI for verifiable encrypted computation and
    Lux TFHE for on-chain proof verification (\S\ref{sec:integration}).
    \item Production benchmarks on real workloads: encrypted medical image
    classification, encrypted credit scoring, and encrypted document
    embedding (\S\ref{sec:results}).
\end{enumerate}

% ============================================================================