Building Machine Learning models for a law enforcement triage tool to detect illicit images

Kristen L McCooey, University of Rhode Island

Abstract

Child pornography is an atrocious crime that is continuing to grow at such a rapid pace that law enforcement (LE) is struggling to control the problem. One of the major problems faced by LE in conducting these types of investigations arises during the execution of the search warrant. Due to a lack of appropriate tools, LE investigators do not have a good way of triaging digital media on site. This thesis examines taking the RedLight Pornography Scanner, an existing tool that was designed for automatically scanning and detecting pornography on digital media, and tailoring it to better detect child pornography based on LE requirements in a triage environment. ^ This research examines the current algorithms used by the RedLight pornography scanner and modifies them to better assist in finding child pornography on digital media based on law enforcement criteria. This includes enhancing a variety of Computer Vision and Machine Learning techniques for detecting skin tone pixels in images and analyzing the disbursement of those pixels across an image. This research has resulted in a method for detecting child pornography in images on digital media that has a 79.70% true negative rate, a 68.70% true positive rate, and meets law enforcement criteria for a triage situation. ^

Subject Area

Sociology, Criminology and Penology|Artificial Intelligence|Computer Science

Recommended Citation

Kristen L McCooey, "Building Machine Learning models for a law enforcement triage tool to detect illicit images" (2012). Dissertations and Master's Theses (Campus Access). Paper AAI1516558.
http://digitalcommons.uri.edu/dissertations/AAI1516558

Share

COinS