Advancing Wildlife Conservation
Through innovative data science and machine learning approaches to combat wildlife trafficking
Project Overview
Wildlife trafficking is a global crisis affecting biodiversity, ecological systems, and public health. This illicit trade generates $7-23 billion annually and encompasses fashion, exotic pets, traditional medicine, and accessories. During the COVID-19 pandemic, traffickers shifted from face-to-face to online interactions, creating new challenges but also opportunities for digital detection.
Our interdisciplinary project combines computer science, criminology, and environmental science to develop innovative tools for discovering, analyzing, and disrupting wildlife trafficking networks operating online.
Our Impact
Our research directly addresses the critical need for effective wildlife trafficking detection and prevention. By developing cost-effective, scalable solutions, we empower researchers and law enforcement agencies worldwide to combat this global crisis more effectively.
The Challenge
Online marketplaces publish millions of ads, but identifying wildlife-related products is like finding a needle in a haystack. For example, searching for "shark" on eBay returns toys, shirts, vacuum cleaners, and only a few actual shark products. Even specific queries like "shark jaw" return irrelevant fossil items. This makes data collection and analysis extremely challenging for researchers studying wildlife trafficking patterns. What if you could achieve 95% accuracy in wildlife trafficking detection in online marketplaces while slashing labeling costs?
Our Solution
Learn to Sample (LTS) makes this possible by combining clustering with multi-armed bandit sampling and leveraging LLMs to label the data. The derived labeled data can be used to train specialized classifiers. this approach tackles the notorious challenge of highly imbalanced datasets—turning the detection of illegal endangered species trade from needle-in-a-haystack to precision targeting.
Our Approach
Our Learn to Sample methodology addresses the fundamental challenge of identifying relevant wildlife advertisements in massive online datasets. We use clustering to ensure diverse sample selection, multi-armed bandit strategies to balance exploration and exploitation, and active learning to iteratively refine our classifiers.
This approach enables researchers to create specialized models for different wildlife trafficking research questions at a fraction of the cost of traditional methods, making large-scale wildlife trafficking analysis accessible to researchers and law enforcement agencies worldwide.
This project is funded by the NSF