SenBen: Sensitive Scene Graphs for Explainable Content Moderation

TL;DR

SenBen introduces a large-scale scene graph benchmark for sensitive content, enabling explainable content moderation with improved models that outperform existing systems in accuracy and efficiency.

Contribution

The paper presents the first large-scale sensitive scene graph benchmark and a compact, efficient model that enhances explainability and performance in content moderation.

Findings

SenBen benchmark contains 13,999 frames with detailed annotations.

The proposed model improves SenBen Recall by 6.4 percentage points.

The model achieves faster inference and lower GPU memory usage than competitors.

Abstract

Content moderation systems classify images as safe or unsafe but lack spatial grounding and interpretability: they cannot explain what sensitive behavior was detected, who is involved, or where it occurs. We introduce the Sensitive Benchmark (SenBen), the first large-scale scene graph benchmark for sensitive content, comprising 13,999 frames from 157 movies annotated with Visual Genome-style scene graphs (25 object classes, 28 attributes including affective states such as pain, fear, aggression, and distress, 14 predicates) and 16 sensitivity tags across 5 categories. We distill a frontier VLM into a compact 241M student model using a multi-task recipe that addresses vocabulary imbalance in autoregressive scene graph generation through suffix-based object identity, Vocabulary-Aware Recall (VAR) Loss, and a decoupled Query2Label tag head with asymmetric loss, yielding a +6.4 percentage point improvement in SenBen Recall over standard cross-entropy training. On grounded scene graph metrics, our student model outperforms all evaluated VLMs except Gemini models and all commercial safety APIs, while achieving the highest object detection and captioning scores across all models, at $7.6\times$ faster inference and $16\times$ less GPU memory.