Exposing Video Forgeries: A Deep Learning Analysis of SVM vs. CNN

Abstract

Introduction: The rapid proliferation of digital media has heightened concerns regarding video forgeries, necessitating robust detection mechanisms for security and forensic applications. Traditional methods often struggle to detect complex forgeries, particularly in low-quality or compressed videos.
Objectives: This study conducts an analysis of Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) for video forgery detection. The dataset underwent preprocessing, including resizing, grayscale conversion, and normalization, to enhance frame uniformity and reduce noise. Videos were segmented, and the difference of consecutive frames (DOCFs) was computed to extract feature vectors for SVM classification..
Methods: CNN was trained to detect complex spatial-temporal forgery patterns, including edge inconsistencies and visual artifacts, by utilizing convolutional and pooling layers, which were analyzed by fully connected layers.
Results: Experimental results demonstrated that SVM achieved an accuracy of 46.67%, while CNN significantly outperformed it with an accuracy of 73.33% on a dataset of 57 videos. Further analysis on a smaller dataset of 33 videos showed CNN achieving 81.82% accuracy, reinforcing its adaptability and robustness in detecting forgeries..
Conclusions: The study highlights CNN's superiority in handling intricate manipulations, positioning it as a viable approach for scalable and precise video forgery detection in evolving digital landscapes.