C4Censor: A Lightweight Benchmark dataset for Inappropriate Content Detection
摘要
The rapid dissemination of graphic and violent imagery in recent times has outpaced existing moderation tools, and yet most public benchmarks focus on only one type of inappropriate content and treat moderation as a simple binary task—Pornography vs. Neutral. This research introduces C4Censor, a lightweight, multi-class image benchmark for fine-grained censorship across four high-risk categories Blood & Gore, Pornography, Terrorism, and Neutral—each subdivided into three challenging subclasses (e.g., Hentai vs. Anime, Counter-Terrorism vs. War-Crimes) with 500 images per subclass, for a total of 6k images. All visuals were scraped from publicly accessible sources—gaming streams (YouTube, Twitch), medical procedure archives, specialty NSFW (not safe for work) collections, X (formerly Twitter) and Telegram channels tracking extremist activity, and curated public websites—and each image was manually annotated at both the coarse (4-way) and fine (12-way) levels. Unlike existing datasets that are either binary or single-domain (e.g., porn vs. Neutral, violence vs. non-violence), C4Censor presents a unified, balanced, and multi-modal challenge. In benchmarking nine state-of-the-art deep-learning models, even top Vision Transformer variants achieved only 62.1% top-1 accuracy, underscoring both the dataset’s complexity and the pressing need for more robust censorship algorithms.
Disclaimer: The article contains inappropriate content; however, this cannot be avoided owing to the nature of the work.