The increasing amount of available User-Created Content (UCC) has resulted in the need of efficient systems for the filtering of malicious content. In this paper, we discuss a system capable of detecting malicious content by relying on the use of semantic features. The semantic features are represented by the likelihood rate of elementary concepts (e.g., `breast', `bottom'). These elementary concepts together constitute a high-level concept (e.g., `naked people'). To do so, the proposed system first segments video into shots. In a next step, semantic features are automatically extracted from key frames that are representative for the shots found. The system then determines the nature of the video content using the extracted semantic features. To verify the usefulness of the proposed filtering system, experiments were performed with UCC available on social network sites.