Video Duplication Forgery Detection Using EfficientNetB0 with Motion-Based Verification

Abstract

The authenticity of digital videos has become a critical concern due to the widespread availability of advanced editing tools that enable subtle manipulations. Frame duplication is a common form of video tampering in which frames are copied and reinserted at different temporal positions to hide or alter events. Detecting such manipulations is challenging. This paper introduces a dual-stage detection framework that addresses this challenge by combining deep feature representations with motion-consistency analysis. The proposed method first employs a lightweight EfficientNetB0 model to extract discriminative features from video frames. A temporal-constrained cosine similarity module then identifies potential duplicate candidates by comparing features only beyond a minimum frame gap and reducing false positives from adjacent frames. In the second stage, a dense optical flow verification module analyzes motion patterns between candidate pairs, confirming duplications only when high visual similarity is accompanied by negligible inter-frame motion. The framework is rigorously evaluated on multiple benchmark datasets, including TDTVD, Fadl, and SULFA. Experimental results demonstrate that the method achieves state-of-the-art performance, attaining an average accuracy of 99.82% and a recall of 100% on the TDTVD dataset. Comparative analysis shows consistent superiority over existing techniques in both detection accuracy and operational efficiency. The architecture ensures computational practicality by limiting resource-intensive optical flow computation to a small subset of high-similarity frames. This work offers significant improvements in reliably identifying duplication forgeries while maintaining feasibility for real-world applications.