AttM-CNN: Attention and metric learning based CNN for pornography, age and Child Sexual Abuse (CSA) Detection in images

Abstract

Nowadays, the growing number of cases of possession and distribution of Child Sexual Abuse (CSA) material pose a significant challenge for Law Enforcement Agencies (LEAs). In this paper, we decompose the automatic CSA detection problem into two simpler ones for which it is feasible to create massive labeled datasets, especially to train deep neural networks: (i) pornographic content detection and (ii) age-group classification of a person as a minor or an adult. We propose a deep CNN architecture with a novel attention mechanism and metric learning, denoted as AttM-CNN, for these tasks. Furthermore, the pornography detection and the age-group classification networks are combined for CSA detection using two different strategies: decision level fusion for binary CSA classification and score level fusion for the re-arrangement of the suspicious images. We also introduce two new datasets: (i) Pornographic-2M, which contains two million pornographic images, and (ii) Juvenile-80k, including 80k manually labeled images with apparent facial age. The experiments conducted for age-group and pornographic classification demonstrate that our approach obtained similar or superior results compared to the state-of-the-art systems on various benchmark datasets for both tasks, respectively. For the evaluation of CSA detection, we created a test dataset comprising one million adult porn, one million non-porn images, and 5,000 real CSA images provided to us by Police Forces. For binary CSA classification, our method obtained an accuracy of 92.72%, which increases the recognition rate by more than 21% compared to a well-known forensic tool, i.e. NuDetective. Furthermore, re-arrangement of the CSA test dataset images showed that 80% of CSA images can be found in the top 8.5% of images in the ranked list created using our approach.