mirror of
https://github.com/zoo-apps/papers.git
synced 2026-10-30 14:56:54 +00:00
zeekay
1474380e17
Locked canonical naming for the Lux FHE product line: Lux FHE = FHE framework umbrella Torus = compiler / runtime TorusVM = core FHE runtime virtual machine TorusEVM = encrypted EVM (this rename) TorusML = private ML TorusNet = threshold/decryption network Zoo FHE / zoo-fhe-ai paper now references TorusEVM as the on-chain FHE-EVM that Zoo extends with CKKS support + ML inference at the T-Chain layer.
90 lines
2.7 KiB
TeX
90 lines
2.7 KiB
TeX
\section{Performance Evaluation}
|
|
|
|
\label{sec:evaluation}
|
|
|
|
\subsection{Testnet Configuration}
|
|
|
|
All benchmarks were performed on Zoo testnet (chain ID 200201):
|
|
|
|
\begin{itemize}
|
|
\item \textbf{Validators}: 21 nodes with NVIDIA A100 GPUs
|
|
\item \textbf{FHE Parameters}: $N = 2^{14}$, $\log Q = 438$ (\CKKS{}); $n = 1024$ (\TFHE{})
|
|
\item \textbf{Security Level}: 128-bit post-quantum security
|
|
\item \textbf{Block Gas Limit}: 1B (Zoo inherits Lux's GPU-accelerated EVM~\cite{lux-evmgpu-benchmark})
|
|
\end{itemize}
|
|
|
|
\subsection{Inference Latency}
|
|
|
|
\begin{table}[t]
|
|
\centering
|
|
\begin{tabular}{lcccc}
|
|
\toprule
|
|
\textbf{Model} & \textbf{Params} & \textbf{Plaintext} & \textbf{Encrypted} & \textbf{Slowdown} \\
|
|
\midrule
|
|
MLP-3 & 1.2K & 0.3ms & 45ms & 150$\times$ \\
|
|
CNN-Small & 62K & 2.1ms & 120ms & 57$\times$ \\
|
|
ResNet-18 & 11.7M & 8.4ms & 340ms & 40$\times$ \\
|
|
ViT-Tiny & 5.7M & 12ms & 890ms & 74$\times$ \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Encrypted inference latency on Zoo testnet. \CKKS{} for linear layers, \TFHE{} for activations.}
|
|
\label{tab:inference}
|
|
\end{table}
|
|
|
|
\subsection{Gradient Aggregation Throughput}
|
|
|
|
\begin{table}[t]
|
|
\centering
|
|
\begin{tabular}{lccc}
|
|
\toprule
|
|
\textbf{Gradient Size} & \textbf{Participants} & \textbf{Updates/s} & \textbf{Gas/Update} \\
|
|
\midrule
|
|
1K params & 10 & 4,200 & 85K \\
|
|
10K params & 10 & 1,800 & 210K \\
|
|
100K params & 10 & 1,200 & 480K \\
|
|
1M params & 10 & 340 & 1.2M \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Encrypted gradient aggregation throughput using \TFHE{} on Zoo testnet.}
|
|
\label{tab:gradients}
|
|
\end{table}
|
|
|
|
\subsection{Comparison with Existing Systems}
|
|
|
|
\begin{table}[t]
|
|
\centering
|
|
\begin{tabular}{lcccc}
|
|
\toprule
|
|
\textbf{System} & \textbf{FHE on-chain} & \textbf{Dual Scheme} & \textbf{Threshold Dec} & \textbf{GPU Accel} \\
|
|
\midrule
|
|
TorusEVM & \checkmark & \checkmark & \checkmark & \checkmark \\
|
|
Sunscreen & $\times$ & $\times$ & $\times$ & $\times$ \\
|
|
\textbf{Zoo FHE} & \checkmark & \checkmark & \checkmark & \checkmark \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Feature comparison. Zoo FHE is the first system combining all four capabilities.}
|
|
\label{tab:comparison}
|
|
\end{table}
|
|
|
|
\subsection{Gas Cost Analysis}
|
|
|
|
\begin{table}[t]
|
|
\centering
|
|
\begin{tabular}{lcc}
|
|
\toprule
|
|
\textbf{Operation} & \textbf{Gas Cost} & \textbf{USD (at \$0.01/Mgas)} \\
|
|
\midrule
|
|
\texttt{ckksAdd} & 15,000 & \$0.00015 \\
|
|
\texttt{ckksMul} & 65,000 & \$0.00065 \\
|
|
\texttt{ckksMatMul} (256$\times$256) & 420,000 & \$0.0042 \\
|
|
\texttt{tfheAdd} & 8,000 & \$0.00008 \\
|
|
\texttt{tfheLt} & 45,000 & \$0.00045 \\
|
|
\texttt{ckksToTfhe} & 120,000 & \$0.0012 \\
|
|
\texttt{requestDecrypt} & 50,000 & \$0.0005 \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Gas costs for FHE precompile operations on Zoo Network.}
|
|
\label{tab:gas}
|
|
\end{table}
|
|
|