Deepfake Detection System Using Hybrid Deep Learning for Images and Videos

The rapid growth of deepfake technology poses severe threats to digital security, privacy, and media integrity. These AI-generated images and videos, often indistinguishable from real content, enable misinformation, identity theft, and digital fraud. This project proposes a Deepfake Detection System using a hybrid EfficientNet-ResNet50 model to improve accuracy and reliability, with a training accuracy of 0.9906 and validation accuracy of 0.8659. Facial regions are extracted from images and video frames with Haar Cascade, followed by advanced preprocessing and augmentation for robust training on the DFDV dataset. Explainable AI (LIME) enhances transparency by highlighting key facial features influencing predictions. Deployed via a Flask-based web application, the system enables real-time analysis with superior accuracy, reduced false positives, and efficient processing. By strengthening defenses against manipulated media, the framework fosters trust in digital content and supports applications in cybersecurity, media verification, and digital forensics.