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

Area of Science:

• Forensic Science
• Computer Science
• Anthropology

Background:

• Previous machine learning models for CSAM identification had limitations, including lack of expert annotation and use of real child exploitation imagery.
• Existing solutions often failed to provide detailed justifications for classification decisions, crucial for regulatory compliance like the EU's AI Act.

Purpose of the Study:

• To train convolutional neural networks (CNNs) using expert-labeled CSAM images to identify critical visual elements for classification.
• To evaluate the effectiveness of different neural network architectures in detecting CSAM.
• To provide a transparent and justifiable AI tool for CSAM identification.

Main Methods:

• Trained four neural network architectures (MobileNet, ResNet152, xResNet152, ResNet-s) on 60,000 images across four classes: CSAM, adult sexual content, non-sexual human images, and images without people.
• Utilized images annotated by forensic experts in anthropology and sexology.
• Employed expert knowledge to improve model training and interpretation.

Main Results:

• Achieved high accuracy in CSAM classification, with xResNet152 and ResNet-s reaching F1 scores of 0.93 (92.8% and 93.1% accuracy, respectively).
• ResNet152 and MobileNet also demonstrated strong performance (F1=0.90, 91.39% and F1=0.85-0.87, 86-87% accuracy).
• Identified breasts, face, and torso as critical areas for classification, supporting anthropological analysis.

Conclusions:

• Expert-annotated data significantly enhances the accuracy of machine learning models for CSAM detection.
• The ResNet-s neural network shows potential as a reliable tool for forensic experts and anthropology-related legal cases.
• A free application for CSAM classification is available for forensic experts, law enforcement, and prosecutors.