DeepPatrol: Finding Illicit Videos for Law Enforcement

• NCJ Number: 254636
• Author(s): Marco Alvarez Vega
• Date Published: June 2019
• Length: 11 pages
• Grant Number(s): 2016-MU-CX-K015
• Agencies: NIJ-Sponsored
• Publication Type: Research (Applied/Empirical); Report (Study/Research); Report (Grant Sponsored); Program/Project Description

Annotation

This is the Final Summary Report on the methodology and results of the development of DeepPatrol, an innovative software tool to assist law enforcement agencies in investigating child pornography cases.

Abstract

This project used recent advances in computer vision and machine learning to automate the process of identifying sexually exploitative Imagery of children (SEIC), so as to reduce the time investigators spend in collecting image evidence. Traditional machine learning methods for this task rely on manual feature engineering and tend to produce many false positives. When many files are on a suspect's hard drive, the agent reviewing evidence may be overwhelmed with imagery that is falsely flagged as SEIC, especially when a variety of pornographic images are present. This occurs because pornographic content is difficult to distinguish from SEIC when using traditional techniques. Researchers have recently had greater success in detecting nudity; however, these models have difficulty in differentiating nudity in SEIC from nudity in adult pornography. In addressing this issue, the current project fused the predictions of these more accurate deep-learning models for nudity detection with apparent age estimation to identify SEIC images. This novel approach provides a framework that can be as fine or coarse a filter as an investigator specifies. It can distinguish between challenging examples of pornographic and SEIC videos with 89-percent accuracy, using the default thresholds. Since this approach relies on automatic representation learning with the use of convolutional neural networks, investigators are not required to be directly exposed to pornographic or SEIC images. The models were tested on a series of challenging datasets in analyzing their performance prior to presenting results on data collected from actual cases. 2 figures