MULTILEAK: A Multi-Modal Feature Fusion Network for Information Leakage Detection in Short Videos

With the rise of digital content, short video platforms have become a significant vector for the hidden dissemination of sensitive information. Current detection methods struggle with the complexities of multimodal content, where data is embedded in audio, text, and visuals in subtle forms. In this paper, we develop a multimodal dataset comprising 2,000 original videos, 13,387 videos with information leakage, and 15,000 unrelated videos. We present MULTILEAK, a multimodal framework designed to detect information leakage in short videos. MULTILEAK integrates features from text, video frames, and audio to detect leaks in these diverse modalities. It incorporates advanced techniques such as entity recognition weighting and a Siamese network, enhancing the system’s ability to identify leaked information. The experiment results demonstrate that MULTILEAK significantly outperforms existing detection models in complex video content, making it a robust solution for multimodal information leakage detection in short videos.