RealVision – Deepfake Detection of Videos and Images: A Comparative Study of Custom CNN for Detection of Deep faked for Human Videos and Images

Abstract

In the recent times rise in the content of the manipulated media even particularly deepfakes poses a serious threats to our digital integrity and privacy and the public trust. This study introduces a model called RealVision which is a lightweight and robust deepfake detection system capable of analyzing both images and video content of the humans. This designed to function efficiently on CPU-based systems. And the framework incorporates a custom-built Convolutional Neural Network (CNN) optimized for performance without relying on GPU acceleration keeping in mind the lower end devices . RealVision is trained using the Celeb DF V2 dataset and follows a multi-phase pipeline including dataset preprocessing, automated face extraction, model training and real-time inference. And also the evaluation metrics such as accuracy, precision, recall, and F1-score were used to assess performance across modalities. Our model’s experimental results demonstrate competitive detection accuracy while maintaining low latency so making it suitable for the deployment in resource-constrained environments. RealVision takes into account to contribute to media authenticity verification tools thus in promoting digital safety in journalism, legal forensics, and public information platforms for betterment.