papers/hanzo-fhe-inference/sections/03-scheme-selection.tex

\section{Scheme Selection}

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

The FHE inference system automatically selects the optimal scheme based on the
model architecture:

\begin{table}[H]
\centering
\small
\begin{tabular}{lll}
\toprule
\textbf{Model Type} & \textbf{Scheme} & \textbf{Rationale} \\
\midrule
Dense neural network & CKKS & Approximate arithmetic, SIMD batching \\
Convolutional network & CKKS & Rotation-based convolution \\
Decision tree / forest & TFHE & Exact comparisons \\
Gradient-boosted model & TFHE & Parallel tree evaluation \\
Logistic regression & CKKS & Matrix-vector product \\
SVM (linear) & CKKS & Inner product + sign \\
SVM (kernel) & CKKS + TFHE & CKKS for kernel, TFHE for argmax \\
\bottomrule
\end{tabular}
\caption{Automatic scheme selection by model architecture.}
\label{tab:scheme}
\end{table}

For hybrid models (e.g., neural network with tree-based post-processing), the
system uses \emph{scheme switching}: CKKS ciphertexts are converted to TFHE
ciphertexts at the boundary, incurring a one-time cost of $O(N \log N)$
operations.

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