A Machine Learning-based Forensic Discriminator of Pornographic and Bikini Images

The increased use of microcomputers and smart-phones has contributed to social progress, but it has also facilitated the exchange of illegal files, such as child pornography photographs, increasing the demand for digital forensic examination in these devices, which can store more than 300,000 images. Based on the large number of files to be analyzed, it is necessary to use capable algorithms to perform this detection, especially in the most challenging scenarios, such as the distinction between pornographic and bikini images. In this work, we present an approach that improves the "Algorithm for detection of Nudity" proposed by Ap-Apid. Our method improved the performance of this algorithm by using machine learning in extracted features from detected skin regions instead of using static rules to classify this kind of images. In addition, we used detected faces as features in order to mitigate false positives in portrait photos. For conducting the training, validation and testing phases, we used the AIIA-PID4 pornographic data set. Furthermore, we also present a statistical analysis by comparing our approach to two algorithms, one based on the “Algorithm for detection of Nudity” and another proposed by the AIIA-PID4 pornographic data set author. The experimental results showed that we achieved an accuracy of 96.96% and 94.94 in the F1-score metric, increasing the accuracy by 79.19% and 18.21% compared to the referred works, respectively.