Using expert-reviewed CSAM to train CNNs and its anthropological analysis

Machine learning methods for the identification of child sexual abuse materials (CSAM) have been previously studied, however, they have serious limitations. Firstly, the training sets used to train the appropriate machine learning algorithms were not previously annotated by a forensic expert in anthropology. Secondly, previously presented solutions have rarely used models trained using real pornographic content involving children. Thirdly, previous studies have not presented a detailed justification for the classification decisions made, which is important due to the recent guidelines of the European Commission (Artificial Intelligence Act).The aim of the study was to train convolution neural networks (CNNs) using expert-labelled CSAM images and thereby identify the elements of the body and/or the environment that are critical for classifications by the neural network. To train and evaluate machine learning models, we used 60,000 images equally divided into four classes (CSAM images, images displaying sexual activity involving adults, images of people without sexual activity, and images not containing people). We used four neural network architectures: MobileNet, ResNet152, xResNet152 and its modification ResNet-s, designed for the purpose of research.The trained models provided high accuracy of classifying CSAM images: xResNet152 (F1 = 0.93, 92,8%), xResNet-s (F1 = 0.93, 93,1%), ResNet152 (F1 = 0.90, 91,39%), MobileNet (F1 ranged from 0.85 to 0.87, ac-curacy ranged from 86% to 87%).The results of the conducted research suggest that using expert knowledge (in sexology and anthropology) significantly improved the accuracy of the models. In regard to further anthropological analysis, the results indicate that the breasts, face and torso are crucial areas for the classification of pornographic content with children's participation. Results suggests that the ResNet-s neural network may be a reliable tool for clinical work and to support the work of experts witnesses in the field of anthropology.The study design received a positive opinion of the Ethics Committee of the Faculty of Mathematics, Informatics and Mechanics of the University of Warsaw. The clinical material was used for research purposes with the consent of the relevant prosecutor's offices. Authors provided free version of Windows application to classify CSAM for forensic experts, policemen and prosecutors at the OSF repository (DOI: 10.17605/OSF.IO/RU7JX).

Automated summary

Machine learning methods for identifying child sexual abuse materials have limitations, including the lack of annotated training sets by forensic experts, rare use of models trained on real pornographic content, and absence of detailed justification for classification decisions. This study trained convolution neural networks using expert-labelled CSAM images, achieving high accuracy in classifying CSAM. The study also highlighted the crucial areas for classifying pornographic content involving children and suggested the ResNet-s neural network as a reliable tool for clinical work and supporting expert witnesses in the field of anthropology.