A comparative analysis of zebra optimization algorithm and chaotic sinusoidal zebra optimization algorithm for video forgery detection system

Abstract

Video forgery, including deletion, duplication, and insertion, threatens multimedia integrity, yet CNN-based Video forgery detection models often suffer from suboptimal hyperparameter tuning. This study compares the Zebra Optimization Algorithm (ZOA) and aChaotic Sinusoidal variant (CSZOA) for optimizing CNN performance in forgery detection. ZOA was chosen for its balanced search capability, while Chaotic Sinusoidal mapping was integrated to improve population diversity, avoid local optima, and accelerate convergence. The framework embedded the optimizer in CNN transfer learning layers to fine-tune parameters such as learning rates, Number of filters, filter sizes, and batch size configurations. A dataset of 270 forged videos acquired from kaggle.com underwent preprocessing through frame extraction, resolution normalization and histogram equalization. Results show CSZOA-CNN outperforms ZOA-CNN and CNN, achieving 99.51% accuracy, 0.32% False Positive Rate, and 39.86 s detection time. These findings highlight the benefit of Chaotic Sinusoidal dynamics in enhancing CNN training efficiency and robustness for real-world video forgery detection