mirror of
https://github.com/lux-apps/papers.git
synced 2026-10-30 18:15:21 +00:00
zeekay
abf99d05fc
Restructured from flat .tex/.pdf files into per-paper subdirectories. Added Antje Worring as co-author on all papers. Synced PDFs to pdfs/ for download site.
85 lines
2.8 KiB
TeX
85 lines
2.8 KiB
TeX
\section{Performance}
|
|
|
|
\label{sec:performance}
|
|
|
|
All benchmarks on AMD EPYC 7763, single core, Go~1.22 with AVX-512 SIMD NTT.
|
|
|
|
\subsection{TFHE Boolean Operations}
|
|
|
|
\begin{table}[ht]
|
|
\centering
|
|
\caption{TFHE gate evaluation benchmarks ($N = 1024$, 128-bit security).}
|
|
\label{tab:tfhe-benchmarks}
|
|
\begin{tabular}{lr}
|
|
\toprule
|
|
\textbf{Operation} & \textbf{Latency} \\
|
|
\midrule
|
|
AND / OR / XOR / NAND / NOR / XNOR & 12\,ms \\
|
|
NOT & 0.02\,ms \\
|
|
MUX (conditional select) & 24\,ms \\
|
|
8-bit addition & 310\,ms \\
|
|
8-bit comparison & 280\,ms \\
|
|
16-bit addition & 680\,ms \\
|
|
16-bit comparison & 620\,ms \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
\subsection{BFV Batched Arithmetic}
|
|
|
|
\begin{table}[ht]
|
|
\centering
|
|
\caption{BFV batched arithmetic benchmarks ($N = 4096$, 128-bit security).}
|
|
\label{tab:bfv-benchmarks}
|
|
\begin{tabular}{lrr}
|
|
\toprule
|
|
\textbf{Operation} & \textbf{16-bit} & \textbf{32-bit} \\
|
|
\midrule
|
|
Encrypted addition & 45\,ms & 52\,ms \\
|
|
Encrypted subtraction & 45\,ms & 52\,ms \\
|
|
Encrypted multiplication & 180\,ms & 340\,ms \\
|
|
Plaintext-ciphertext multiply & 22\,ms & 28\,ms \\
|
|
Relinearization & 85\,ms & 120\,ms \\
|
|
Bootstrapping & 1.2\,s & 2.4\,s \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
\subsection{Comparison to Cleartext EVM Execution}
|
|
|
|
\begin{table}[ht]
|
|
\centering
|
|
\caption{FHE overhead relative to cleartext EVM operations.}
|
|
\label{tab:overhead}
|
|
\begin{tabular}{lrrr}
|
|
\toprule
|
|
\textbf{Operation} & \textbf{Cleartext} & \textbf{FHE (BFV)} & \textbf{Overhead} \\
|
|
\midrule
|
|
\texttt{uint256 add} & 3 gas (5\,ns) & 65K gas (45\,ms) & $9 \times 10^6$ \\
|
|
\texttt{uint256 mul} & 5 gas (8\,ns) & 320K gas (180\,ms) & $2.3 \times 10^7$ \\
|
|
\texttt{>=} comparison & 3 gas (5\,ns) & 480K gas (480\,ms) & $9.6 \times 10^7$ \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
The overhead is large in absolute terms ($10^6$--$10^8\times$). This is inherent to
|
|
FHE and consistent across all implementations. The relevant comparison is not FHE
|
|
vs.\ cleartext, but FHE vs.\ the \emph{alternative for achieving the same privacy
|
|
guarantee}:
|
|
|
|
\begin{itemize}[leftmargin=2em]
|
|
\item Off-chain computation with on-chain verification (ZK proofs): requires a
|
|
trusted prover and does not support arbitrary computation.
|
|
\item Trusted execution environments (TEEs): requires trusting hardware vendors
|
|
and is vulnerable to side-channel attacks.
|
|
\item Multi-party computation (MPC): requires interaction between parties for
|
|
every operation and does not scale to public smart contracts.
|
|
\end{itemize}
|
|
|
|
FHE is the only approach that enables \emph{non-interactive} computation on encrypted
|
|
data by \emph{untrusted} parties. The overhead is the price of this unique property.
|
|
|
|
%% ========================================================================
|
|
%% 8. SECURITY ANALYSIS
|
|
%% ========================================================================
|