Real-Time Multi-Modal Detection of Harmful Videos in Vietnamese Online Platforms

The rapid growth of online video platforms such as YouTube, TikTok and Facebook has accelerated the spread of harmful content, including violence, explicit material, incitement to violence and superstition. While harmful video detection has attracted increasing global research interest, there remains little work tailored to Vietnam's cultural and societal context, where such content is highly prevalent. This work presents one of the first large-scale studies on harmful video detection in Vietnam, combining the HarmfulVideosVN2023 dataset (1,589 videos across five categories) with additional data collected by our team and introducing a culturally salient superstitious category. We propose ViHarmfulDetect, a multimodal deep learning model with featurelevel fusion that jointly leverages representations from pre-trained visual, audio and Vietnamese text encoders. On the extended dataset, ViHarmfulDetect achieves high F1-scores for superstitious (0.97), pornographic (0.95) and offensive (0.97) content, while performance is comparatively lower for violent (0.86) and horrible (0.73); overall, the model consistently outperforms unimodal baselines. We also build a real-time streaming pipeline using Apache Kafka, Spark Structured Streaming and Cassandra for low-latency analysis. Our contributions are threefold: (i) extending harmful video detection with a culture-specific superstition class, (ii) introducing a high-quality Vietnamese multimodal dataset, and (iii) delivering an accurate, real-time deployable system for content moderation and surveillance. We release the dataset publicly to support future research.11Dataset repository (GitHub): locngocphan12/Real-Time-Multi-Modal-Detection-of-Harmful-Videos-in-Vietnamese-Online-Platforms.