papers/zoo-federated-wildlife/sections/07-implementation.tex

\section{Implementation}

\label{sec:implementation}

\subsection{System Architecture}

\begin{figure}[H]
\centering
\begin{tikzpicture}[
    node distance=1.5cm,
    box/.style={rectangle, draw, rounded corners, minimum width=2.5cm, minimum height=0.7cm, text centered, font=\footnotesize},
    edge/.style={->, thick},
    dedge/.style={<->, thick, dashed}
]
    \node[box, fill=blue!10] (coord) {ZooFed Coordinator};
    \node[box, fill=green!10, below left=1.5cm and 1cm of coord] (s1) {Site 1: Camera Traps};
    \node[box, fill=green!10, below=1.5cm of coord] (s2) {Site 2: Acoustic + Visual};
    \node[box, fill=green!10, below right=1.5cm and 1cm of coord] (s3) {Site 3: eDNA + GPS};

    \draw[edge] (coord) -- node[left, font=\tiny] {$\phi, \psi_m$} (s1);
    \draw[edge] (coord) -- node[left, font=\tiny] {$\phi, \psi_m$} (s2);
    \draw[edge] (coord) -- node[right, font=\tiny] {$\phi, \psi_m$} (s3);
    \draw[edge] (s1) -- node[right, font=\tiny] {$\nabla\phi$} (coord);
    \draw[edge] (s2) -- node[right, font=\tiny] {$\nabla\phi$} (coord);
    \draw[edge] (s3) -- node[left, font=\tiny] {$\nabla\phi$} (coord);
\end{tikzpicture}
\caption{ZooFed federated learning architecture.}
\label{fig:architecture}
\end{figure}

\subsection{Software Components}

\begin{table}[H]
\centering
\caption{ZooFed software components.}
\label{tab:software}
\begin{tabularx}{\textwidth}{@{}llcX@{}}
\toprule
\textbf{Component} & \textbf{Language} & \textbf{Lines} & \textbf{Description} \\
\midrule
Coordinator & Python & 6,400 & Aggregation, clustering, privacy accounting \\
Site Client & Python & 8,200 & Local training, gradient computation, DP noise injection \\
Sensor Adapters & Python & 4,800 & Modality-specific preprocessing and encoding \\
Privacy Accountant & Rust & 3,100 & Moments accountant, per-species budget tracking \\
Communication & Rust & 2,700 & Secure aggregation, gradient compression \\
Monitoring Dashboard & TypeScript & 5,200 & Training progress, privacy budget visualization \\
\bottomrule
\end{tabularx}
\end{table}

\subsection{Communication Efficiency}

Wildlife monitoring sites often have limited bandwidth.
ZooFed reduces communication cost through:

\begin{enumerate}[leftmargin=*]
    \item \textbf{Gradient Sparsification:} Only the top-$p\%$ gradient entries are transmitted (default $p=10$), reducing communication by 10x with <1\% accuracy loss.
    \item \textbf{Quantized Transmission:} Gradient values are quantized to 8-bit fixed point for transmission.
    \item \textbf{Asynchronous Updates:} Sites with poor connectivity can participate asynchronously, submitting updates when bandwidth is available rather than requiring synchronous rounds.
\end{enumerate}

The total communication cost per round per site is:

\begin{equation}
    C_{\text{comm}} = p \cdot |\phi| \cdot 8 \text{ bits} \approx 0.1 \cdot 2.5\text{M} \cdot 8 = 2 \text{ Mbit}
\end{equation}

This is feasible even over satellite links commonly used at remote monitoring sites.

% ============================================================
% 8. Evaluation
% ============================================================