Source camera model identification in smartphone videos using fused features from separated noise residuals

The amount of video data shared online has increased significantly in the era of the internet, primarily due to the widespread availability of cameras and smartphones, which raises concerns about the legitimacy and authenticity of such data in sensitive areas, such as courtrooms and intellectual property rights. As technology advances, so do methods of video manipulation techniques, making it increasingly difficult to determine the origins of digital information. In this paper, a new method for source camera model identification on video data was proposed by using an adaptive Wiener filter to extract noise residuals from video I-frames, separating them into positive and negative components to train two independent deep neural networks. The learned features are fused in two steps before classification: first, independently, and second, with another method features, to enhance robustness against compression artifacts. We also explore two fusion functions for final classification: softmax-based voting and fully connected feature-level aggregation. Our approach aims to maximise the discriminative power of sensor-specific artifacts while mitigating the effects of lossy video compression. We evaluate the method on three smartphone video datasets —QUFVD, Daxing, and VISION —demonstrating consistent improvements in accuracy compared to state-of-the-art methods. The effectiveness of the proposed dual-stream residual learning and feature fusion strategy for robust and scalable source camera model identification in compressed video data is proven by the results.