Efficient Deep Learning Forensics Detection System for Traffic Video Surveillance

Multimedia data, including images and videos, constitutes a critical and pervasive form of global information exchange. Within video forensics, the accurate and robust detection of forged or manipulated video content is essential for maintaining the integrity of digital evidence. This study focuses on detecting presentation attacks and temporal forgeries specifically within traffic surveillance contexts using the FRAUD1 and FRAUD2 datasets. Evaluation of a variety of models had been conducted in this paper, including 2D and 3D Convolutional Neural Networks (CNNs), Random Forest, and Multi-Layer Perceptron (MLP). Experimental results demonstrate that while the 3D CNN achieves a peak accuracy of 100% by capturing temporal dependencies, the 2D CNN maintains a high accuracy of 98.85% with significantly lower computational overhead. The study further proposes a practical deployment guideline that balances detection performance against available computational resources, identifying the 2D CNN as the optimal candidate for real-time edge-based surveillance, whereas the 3D CNN is best suited for centralized forensic analysis. Comparative evaluations against state-of-the-art methods confirm the effectiveness of the proposed system in terms of accuracy, robustness, and architectural efficiency for modern traffic monitoring applications.