RN-Net: A Video Source Identification Method Based on Deep Learning
Kumar Shubham Singh, Vinod Pankajakshan · 2024
In today's digital age, video content is ubiquitous, and its authenticity has become a major concern for law enforcement agencies, legal proceedings, and media outlets. In recent years, deep learning-based methods have shown promising results for video source identification. This paper proposes a deep learning-based approach for source camera identification using a CNN-based denoiser and the MISLnet architecture. The proposed model, Residual Noise-Network (RN-Net), extracts noise residue from patches of video frames to identify the camera used to record the video. We evaluate the performance of RN-Net on the VISION dataset consisting of videos captured with 28 mobile devices. The performance of RN-Net is compared with the conventional Photo responsive non-uniformity (PRNU)-based method and other deep learning networks proposed for video source identification. The experimental results show that the RN-Net outperforms existing methods and accurately identifies the source camera in different scenarios. The proposed method is also robust to in-camera digital stabilization and correctly attributes strongly stabilized videos. Furthermore, we demonstrate the effectiveness of RN-Net against compression algorithms used by social media platforms such as WhatsApp and YouTube. The experimental results show that the proposed approach achieves source identification accuracy up to 97%.