A Spatio-Temporal Forgery Localization Network for Robust Copy Move Detection in Videos

Abstract

With the rapid advancement of digital editing tools, copy-move forgery has become a prevalent form of tampering in video content, posing serious challenges in domains such as surveillance, media forensics, and legal evidence. This work presents a novel deep learning-based framework designed for efficient and accurate detection of copy-move forgeries in videos. The proposed model utilizes a spatial-temporal encoder-decoder architecture that combines convolutional feature extraction and temporal consistency analysis to identify duplicated regions. A fully convolutional segmentation head is employed to localize manipulated areas at the pixel level, ensuring precise forgery localization. Unlike traditional methods that rely heavily on handcrafted features or are limited to spatial analysis, our approach integrates both frame-wise and sequence-level information, enhancing robustness against various transformations such as flipping, scaling, and rotation. Extensive experiments conducted on three benchmark datasets SULFA, GRIP, and VTD demonstrate that the proposed method outperforms existing state-of-the-art techniques in terms of F1-score, MCC, accuracy, and inference time. Furthermore, the model shows strong generalization capabilities across different tampering scenarios. The lightweight nature of the architecture also allows for real-time deployment, making it a practical solution for real-world video forgery detection applications.