MTN: Forensic Analysis of MP4 Video Files Using Graph Neural Networks

Ziyue Xiang, Amit Kumar Singh Yadav, Paolo Bestagini, Stefano Tubaro, Edward Delp

Abstract

MP4 video files are stored using a tree data structure. These trees contain rich information that can be used for forensic analysis. In this paper, we propose MP4 Tree Network (MTN), an approach based on an end-to-end Graph Neural Networks (GNNs) that is used for forensic analysis of MP4 trees. MTN does not use any video pixel data. MTN is trained using Self-Supervised Learning (SSL), which generates semantic-preserving node embeddings for the nodes in an MP4 tree. We also propose a data augmentation technique for MP4 trees, which helps train MTN in data-scarce scenarios. MTN achieves good performance across 3 video forensics tasks on the EVA-7K dataset. We show that MTN can gain more comprehensive understanding about the MP4 trees and is more robust to potential attacks compared to existing methods.
BibTeX
@inproceedings{xiang2024mtn,
  author = {Ziyue Xiang and Amit Kumar Singh Yadav and Paolo Bestagini and Stefano Tubaro and Edward Delp},
  title = {{MTN: Forensic Analysis of MP4 Video Files Using Graph Neural Networks}},
  booktitle = {Workshops @ CVPR},
  year = {2024},
  url = {https://openreview.net/forum?id=Bi588HGymx},
}

Loading related content...

Figure