mirror of
https://github.com/zoo-apps/papers.git
synced 2026-11-02 17:38:34 +00:00
24 lines
1.4 KiB
TeX
24 lines
1.4 KiB
TeX
\section{Case Studies}
|
|
|
|
\label{sec:case_studies}
|
|
|
|
\subsection{Cross-Continental Tiger Monitoring}
|
|
|
|
Three sites in the consortium monitor tiger populations (Sundarbans, Western Ghats, SE Asian sites).
|
|
Local-only models achieved 67\% individual identification accuracy due to small per-site sample sizes.
|
|
ZooFed's federated training, with species embedding sharing for Panthera tigris, improved identification to 84\%, enabling the first cross-site population connectivity analysis without sharing raw location data.
|
|
|
|
\subsection{Acoustic Biodiversity in the Amazon}
|
|
|
|
The Amazon Basin site contributed only acoustic data.
|
|
Through HSA cross-modal alignment with visual data from other Neotropical sites, the acoustic model for bird species improved from 71\% to 88\% accuracy, demonstrating that visual knowledge transfers meaningfully to acoustic identification through the shared embedding space.
|
|
|
|
\subsection{Marine eDNA Privacy}
|
|
|
|
The Great Barrier Reef site contained eDNA detection records for 12 critically endangered marine species.
|
|
Under CCDP with $\gamma=2$, these species received $\epsilon=0.2$ privacy guarantees---sufficient to prevent re-identification of sampling locations even given strong auxiliary information---while overall marine species detection accuracy degraded by only 1.8\% compared to no-privacy training.
|
|
|
|
% ============================================================
|
|
% 10. Discussion
|
|
% ============================================================
|