An Object-Centric Approach for Image Analysis to Combat Human Trafficking

Images are often part of a criminal investigation, but for questions like “where was this image taken,” investigators are often limited to using generic tools designed for different purposes like Google reverse image search. This project, developed with feedback from the St. Louis County (Missouri) Police Department and the National Center for Missing and Exploited Children, (NCMEC) aims to develop novel approaches to index and search imagery, and to deploy these in a prototype search tool with a focus on fighting human trafficking. This tool will be an enhanced version of the TraffickCam system that GWU is already developing for law enforcement. TraffickCam consists of a large database of hotel room photographs and an Artificial Intelligence-based platform for law enforcement to compare victim photographs to the photographs in that database in order to determine the hotel where a victim was photographed. The database of hotel room photographs contains several million hotel room images collected from both publicly available travel websites (e.g., Expedia, TripAdvisor) and the over 100,000 users of our mobile application that allows travelers to submit images of hotel rooms they visit. TraffickCam uses neural networks to support image-based search for members of law enforcement. Images containing victims of trafficking in hotel rooms are provided as input, and the system returns the hotels with the most similar images. (TraffickCam is currently being used and tested by NCMEC and the St. Louis County Police Department.

This research seeks to address user input on ways that the tool could better align with their investigative approaches, including searching with a smaller part of the victim photograph, such as just a particular object that seems unique; more easily understood explanations of why the AI system returned a specific result; and ability to determine where a victim was photographed even if the hotel has been renovated since the photograph was captured. Implementing these desired capabilities is non-trivial. In this proposal, GWU addresses the research questions necessary to provide a more useful AI-based investigative tool to law enforcement officers. Specifically, GWU considers the following questions: What is the best image representation to support exploratory investigative search; how can black box AI tools demonstrate why two images or two objects match; and how can learning-based approaches adapt to differences between the query and the database imagery and detect and model changes in the database over time? GWU’s approach will be to improve current Deep Learning computational models to more explicitly focus on and visualize objects in the images. This will support flexible, intuitive ways for investigators to recognize where images were taken, what objects are recognizable within images and what sets of images have the same object. The version of TraffickCam resulting from this project will be provided to the NIJ Testing and Evaluation Center to assess technical performance, conduct operation assessment of the technology, or conduct an impact assessment for the technology.

"Note: This project contains a research and/or development component, as defined in applicable law," and complies with Part 200 Uniform Requirements - 2 CFR 200.210(a)(14).

ca/ncf