Zero-Shot Pornographic Detection Using Vision-Language Models

The proliferation of pornographic content online challenges content moderation efforts, especially in sensitive contexts. Traditional detection methods often require extensive labeled datasets and struggle with nuanced content. This paper proposes a zero-shot classification approach using Vision-Language Models (VLMs) like CLIP and Open CLIP, which leverage visual-textual alignment to classify pornographic content without task-specific training. Evaluated on the LSPD dataset, our research examined various aspects of using VLMs, including the effects of keyword choice, prompt construction, model size, and pre-training data resolution. Our method achieved comparable or better performance than traditional models, with the best accuracy reaching 91.6% using the key-word “erotica“ and specific descriptive prompts. This approach reduces dependency on large datasets, offering a robust solution for detecting explicit content.