A Multimodal Anonymization Framework for MP4 Videos

The increasing prevalence of video data containing sensitive information necessitates robust anonymization systems capable of safeguarding privacy across multiple modalities. This paper presents a comprehensive multimodal anonymization framework for MP4 videos, addressing privacy concerns related to text, facial features, and audio. The proposed system processes input videos by independently anonymizing embedded text, human faces, and audio streams through specialized pipelines while maintaining temporal and spatial coherence. Text anonymization leverages optical character recognition (OCR) for detection and employs redaction or pseudonymization techniques to ensure privacy. Facial anonymization uses advanced generative models to replace sensitive facial regions with realistic, privacy-preserving synthetic alternatives. Audio anonymization modifies speaker-specific features to eliminate re-identification risks while retaining semantic integrity. The anonymized components are recombined into a coherent output video, ensuring usability for downstream tasks. Experimental results demonstrate the system’s ability to achieve high privacy protection without significant degradation in video utility. This framework is highly relevant for privacy-preserving applications in healthcare, legal systems, social media, and machine learning datasets.