Detecção de Pornografia Infanto-juvenil Baseada em Pornografia Adulta e Estimativa de Idade Facial.

Abstract

Surely the technological evolution has been bringing a huge social and economic advancement to our generation, however, this development has also been used by some individuals to commit new kinds of crimes or even support the old ones. This is what happens with sexual child and teen exploitation, a kind of crime that was committed through the years without technological support, but in the last decades has been boosted by the possession and sharing of digital files with child and teen pornographic content. The increase in these crimes influences directly the demand for digital exams that look for child pornography. These exams almost always are made non-automatically in police scientific departments of Brazil. Because of this need, it was developed a technique, without using any illicit pornographic content, which allows the users to detect pornographic content and infer, through detected human faces, the likelihood of this kind of image to portray child or teens. In order to detect pornography, it was proposed a novel dataset (Pornographic and Explicit Dataset 376K) used with a greed strategy to, respectively, choose and fine-tune a convolutional neural network and its hyperparameters. In order to estimate if the related people are underage, it was presented a novel technique also based on deep learning to estimate the real age from human faces using data from apparent age to improve its results. In the end, it was still proposed a machine learning technique that improved the results related to the likelihood of a human face to belong to an underage person. This approach, composed of innovative modules in pornography detection and real age estimation, which outperforms state-of-the-art researches in those respective fields, also achieved compatible results with the state-of-the-art of child and adolescent pornography detection.