Scholarlink

Child abuse is a criminal offense that needs investigation tools to be more efficient as well as to protect investigators. This is the double objective covered by the ODIP application. This tool has been developed by the French Ministry of Interior Cyberspace Command to provide a child abuse images automatic detection tool to analyze a hard disk or any other physical storage device with great precision and efficiency. The architecture is divided into two major components: a Vision Transformer (ViT) and a Classifier. The ViT model enables us to extract a digital vector from an image, capturing its semantic content. The Classifier is trained to distinguish between vectors representing CSAM and those representing non-CSAM. To train the classifier, we created a corpus of positive and negative images, and all the positive data has been preprocessed by specialist investigator. Thus, when analyzing a hard disk, the program is able to classify each new image as either child abuse or not. This result is still validated by the investigator’s or analyst’s expertise.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ODIP : A State-of-the-Art Child Sexual Abuse Material Detection Model

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jérémy Billuart,
  • Romain Gemignani,
  • Patrick Perrot

摘要

Child abuse is a criminal offense that needs investigation tools to be more efficient as well as to protect investigators. This is the double objective covered by the ODIP application. This tool has been developed by the French Ministry of Interior Cyberspace Command to provide a child abuse images automatic detection tool to analyze a hard disk or any other physical storage device with great precision and efficiency. The architecture is divided into two major components: a Vision Transformer (ViT) and a Classifier. The ViT model enables us to extract a digital vector from an image, capturing its semantic content. The Classifier is trained to distinguish between vectors representing CSAM and those representing non-CSAM. To train the classifier, we created a corpus of positive and negative images, and all the positive data has been preprocessed by specialist investigator. Thus, when analyzing a hard disk, the program is able to classify each new image as either child abuse or not. This result is still validated by the investigator’s or analyst’s expertise.