Research activities focus on the automated detection of Child Sexual Abuse Material (CSAM) using scene recognition as a proxy task. The work emphasizes Scene Graph Generation (SGG) to capture the semantic and relational structure of visual scenes, enabling more accurate and interpretable classification and detection. By modeling objects, attributes, and their relationships, the approach seeks to develop robust detection pipelines that function effectively under the legal, ethical, and technical constraints inherent to CSAM analysis, such as limited training data, restricted evaluation environments, and hardware limitations faced by law-enforcement agencies. The research is conducted in close collaboration with forensic experts to ensure practical applicability and strict adherence to security protocols.