Deepfake Video Detection using Transfer Learning

: Deepfake technology enables the creation of hyper-realistic fake videos, posing significant threats in domains like politics, law enforcement, and cybersecurity. This project proposes a deepfake detection framework leveraging facial feature embeddings using FaceNet512, followed by classification through transfer learning models. Unlike traditional CNN-based detectors that analyze full video frames, this method isolates and processes facial data, ensuring higher efficiency and accuracy. Additionally, a real-time email alert system notifies users upon deepfake detection. Evaluation across benchmark datasets like DFDC and FaceForensics++ demonstrates high accuracy, robustness, and suitability for real-time deployment. The proposed system is designed to detect subtle facial manipulations using efficient feature extraction and classification techniques. By focusing on key facial landmarks and minimizing the amount of data analyzed, it ensures quick and effective identification of forged content. The use of pre-trained models via transfer learning significantly reduces the training time while maintaining high generalization across various datasets. Integration with email alerts also provides timely responses for practical security applications. The solution is versatile, scalable, and offers real - world utility in environments like social media platforms, digital forensics, and content authentication. It offers a balanced approach to performance and usability, setting a foundation for future improvements like audio detection, multimodal input support, and deployment on mobile or edge device.