Loukhaoukha Khaled
École Supérieure de la Gendarmerie Nationale, Zéralda, Algeria.
Development Research Center, Algiers, Algeria.
J Forensic Sci. 2025 Jul;70(4):1359-1374. doi: 10.1111/1556-4029.70043. Epub 2025 May 13.
The widespread use of multimedia editing tools has facilitated the creation of realistic video forgeries, jeopardizing the trust in video content. To address frame duplication forgery, a prevalent technique, this paper introduces a novel algorithm leveraging QR decomposition (orthogonal-triangular decomposition) and Minkowski distance. The algorithm extracts frame features using QR decomposition and compares them with a reference frame using Minkowski distance. Candidate duplicates are identified through random block matching. We evaluate the proposed method on standard datasets (TDTVD, LASIESTA, and IVY LAB) and a self-generated dataset. Our method achieves exceptional performance, attaining a perfect -score for video-level detection on both the TDTVD and our self-generated datasets. Notably, for frame-level detection, it achieves an average accuracy of 0.9943, precision of 0.9752, recall of 0.9858, and -score of 0.9803 across all datasets. Our analysis demonstrates the proposed method demonstrates promising performance in detecting multiply-duplicated frames and shows robustness against post-processing, potentially outperforming existing approaches.
多媒体编辑工具的广泛使用促进了逼真视频伪造的产生，损害了人们对视频内容的信任。为了解决一种普遍存在的技术——帧复制伪造问题，本文介绍了一种利用QR分解（正交三角分解）和闵可夫斯基距离的新颖算法。该算法使用QR分解提取帧特征，并使用闵可夫斯基距离将其与参考帧进行比较。通过随机块匹配来识别候选重复帧。我们在标准数据集（TDTVD、LASIESTA和IVY LAB）以及一个自行生成的数据集上对所提出的方法进行了评估。我们的方法取得了卓越的性能，在TDTVD和我们自行生成的数据集上的视频级检测均获得了满分。值得注意的是，对于帧级检测，在所有数据集中它实现了平均准确率0.9943、精确率0.9752、召回率0.9858以及F1分数0.9803。我们的分析表明，所提出的方法在检测多重复制帧方面表现出了良好的性能，并且对后处理具有鲁棒性，可能优于现有方法。
