Explainable Ensemble Models for Child Sexual Abuse Imagery Classification

• Grant number: 25/08423-5
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: September 01, 2025
• End date: August 31, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Leo Sampaio Ferraz Ribeiro
• Grantee: Clara Ernesto de Carvalho
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated research grant: 23/12086-9 - Araceli: Artificial Intelligence in the Fight Against Child Sexual Abuse, AP.R

Abstract

This research proposal investigates Child Sexual Abuse Imagery (CSAI) detection. The spread of CSAI continues to pose serious challenges to online safety, leading re-victimization and an increase in real world cases of abuse. According to the National Center for Missing and Exploited Children's CyberTipline, over 36 million reports of suspected CSAM were received in 2023, making it a record year. While automated tools for detecting such material have improved, progress remains limited by strict and necessary legal and ethical barriers surrounding access to CSAI. In this project, we focus on the development of explainable ensemble models built on Proxy Tasks --- sub-problems that reflect key aspects of CSAI recognition without requiring access to sensitive content by being built on public data. By combining multiple Proxy Tasks such as age estimation, nudity detection, and scene understanding under a unified ensemble framework, we aim to more closely approximate the complexity of real-world CSAI while maintaining transparency and interpretability. This approach is aligned with responsible AI practices in sensitive domains. We furthermore consider explainable frameworks, allowing for better support for decision making and for the evaluation of how features from each Proxy Task contribute to CSAI recognition. The project is part of the multi-institutional Araceli project and contributes both methodologically and ethically to the advancement of automation in the fight against child sexual abuse.