Computer vision-based CSAM screenin

• Morais, Pablo ^[1] ; Fidalgo, Eduardo ^[1] ; González-Castro, Víctor ^[1] ; Tanveer, Waqar ^[1] ; Mirjalili, Milad ^[1] ; Alegre, Enrique ^[1]
  1. [1] Universidad de León

     Universidad de León

     León, España

     
• Localización: Jornadas de Automática, ISSN-e 3045-4093, Nº. 47, 2026
• Idioma: inglés
• DOI: 10.17979/ja-cea.2026.47.13766
• Títulos paralelos:
  • Cribado de CSAM basado en visión artificial
• Enlaces
  • Texto completo
• Resumen
  • español

    La detección de material de abuso sexual infantil (CSAM) es un desafío relevante en informática forense y protección de menores, cuya investigación está limitada por restricciones legales, éticas y de privacidad. Este trabajo presenta un pipeline modular de visión artificial para el cribado de riesgo de CSAM basado en tres señales visuales proxy: detección facial, estimación de edad aparente y clasificación de contenido adulto. El sistema integra YOLOv12n-face, SwinFace y ViT-B/16, aplicando una regla de decisión que marca una imagen ´cuando coinciden evidencia de contenido adulto y presencia de  menores. Además, se proponen dos datasets proxy de tres clases, APD-2M-3C y NPDI-2K-3C, orientados a diferenciar imágenes pornográficas, no pornográficas sencillas y no pornográficas difíciles. Al no utilizar CSAM real, la evaluación se centra en el rendimiento temporal y los falsos positivos sobre datasets sin CSAM. Los resultados preliminares muestran una capacidad eficiente para procesar grandes volúmenes de imágenes, aunque se requiere validación adicional antes de un uso operativo.

  • English

    The detection of Child Sexual Abuse Material (CSAM) is a major challenge in digital forensics and child protection, while research in this area is constrained by legal, ethical, and privacy restrictions. This paper presents a modular computer vision pipeline for CSAM risk screening based on three proxy visual signals: face detection, apparent age estimation, and adult-content classification. The system integrates YOLOv12n-face, SwinFace, and a ViT-B/16-based pornography classifier, applying a decision rule that flags an image only when evidence of adult content coexists with facial evidence suggesting a possible minor. In addition, two three-class proxy datasets, APD-2M-3C and NPDI-2K-3C, are introduced to distinguish pornographic images, simple non-pornographic images, and challenging non-pornographic images. Since no real CSAM was used, the evaluation focuses on processing performance and false-positive behavior on CSAM-free datasets. Preliminary results show that the pipeline can efficiently process large-scale image collections, although further validation with controlled law-enforcement datasets is required before operational deployment.

• Referencias bibliográficas
  • Amirgaliyev, B., Mussabek, M., Rakhimzhanova, T., Zhumadillayeva, A. (2025). A review of machine learning and deep learning methods for person...
  • Avila, S., Thome, N., Cord, M., Valle, E., ArauJo, A. D. A. (2013). Pooling in image representation: The visual codeword point of view. Computer...
  • Coelho, T., Ribeiro, L. S. F., Macedo, J., dos Santos, J. A., Avila, S. (2024). Transformers-based few-shot learning for scene classification...
  • Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S.,...
  • Gangwar, A., Gonzalez-Castro, V., Alegre, E., Fidalgo, E. (2021). Attm-cnn: Attention and metric learning based cnn for pornography, age and...
  • Gangwar, A., Gonzalez-Castro, V., Alegre, E., Fidalgo, E. (2023). Triple-biggan: Semi-supervised generative adversarial networks for image...
  • Gangwar, A., Gonzalez-Castro, V., Alegre, E., Fidalgo, E., Martínez-Mendoza, A. (2024). Deephsar: Semi-supervised fine-grained learning for...
  • Kaewkorn, H., Zhou, L., Li, W. (2025). Rp-net: A robust polar transformation network for rotation-invariant face detection. Pattern Recognition...
  • Lee, H.-E., Ermakova, T., Ververis, V., Fabian, B. (2020). Detecting child sexual abuse material: A comprehensive survey. Forensic Science...
  • Lin, C., Gou, J., Fan, Z., Liao, Y. (2024). A feature fusion-based resnet using the pooling pyramid for age estimation. Proceedings of the...
  • Liu, X., Qiu, M., Zhang, Z., Shi, Y., Li, Z., Chen, X., Liu, Y., Chen, X., Yu, H. (2025). Enhancing facial age estimation with local and global...
  • Morais, P. (2025). Aprendizaje profundo en la clasificación automatizada de imágenes en la detección de abusos sexuales a menores. (unpublished)
  • Moreira, D., Avila, S., Perez, M., Moraes, D., Testoni, V., Valle, E., Goldenstein, S., Rocha, A. (2016). Pornography classification: The...
  • Pereira, M., Dodhia, R., Anderson, H., Brown, R. (2020). Metadata-based detection of child sexual abuse material. arXiv preprint arXiv:2010.02387.
  • Qin, L., Wang, M., Deng, C., Wang, K., Chen, X., Hu, J., Deng, W. (2023). Swinface: A multi-task transformer for face recognition, expression...
  • Rautela, K., Sharma, D., Kumar, V., Kumar, D. (2024). Obscenity detection transformer for detecting inappropriate contents from videos. Multimedia...
  • Shou, Y., Cao, X., Liu, H., Meng, D. (2025). Masked contrastive graph representation learning for age estimation. Pattern Recognition 158....
  • Steinebach, M. (2023). An analysis of PhotoDNA. Proceedings of the 18th International Conference on Availability, Reliability and Security...
  • Tian, Y., Ye, Q., Doermann, D. (2025). Yolov12: Attention-centric real-time object detectors. URL: https://arxiv.org/abs/2502.12524
  • UNICEF (2026). Artificial intelligence and child sexual abuse and exploitation. URL: https://www.unicef.org/reports/artificial-intelligence-and-child-sexual-abuse-and-exploitation
  • Westlake, B., Brewer, R., Swearingen, T., Ross, A., Patterson, S., Michalski, D., Hole, M., Logos, K., Frank, R., Bright, D., et al. (2022)....
  • Yang, J. (2025). Facial recognition optimization based on adversarial sample generation in the field of artificial intelligence. Discover...
  • Yu, S., Zhao, Q. (2025). Improving age estimation in occluded facial images with knowledge distillation and layer-wise feature reconstruction....
  • Yu, W., Zhou, P., Yan, S., Wang, X. (2025). Inceptionnext: When inception meets convnext. URL: https://arxiv.org/abs/2303.16900
  • Zhang, J., Hu, N. (2025). Accuracy and robustness evaluation of deep learning algorithms in facial recognition systems. Systems and Soft Computing...
  • Zhu, D., Shan, X., Wu, C., Yung, K., Ip, A. (2024). Multi frame obscene video detection with vit: An effective for detecting inappropriate...