Authors: Ms. Neeharika Sengar
Abstract: The rapid advancements in generative adversarial networks (GANs) have enabled the creation of highly realistic deepfake videos, posing significant risks in domains such as politics, cybersecurity, and digital media. Detecting such manipulated content has become a pressing challenge. This study investigates deepfake detection using computer vision techniques by training a convolutional neural network (CNN) model from scratch on the publicly available Face Forensics++ dataset. A systematic methodology involving data preprocessing, model training, and evaluation was adopted. The proposed CNN model achieved a detection accuracy of 92.3% on the test set. Furthermore, the model demonstrated strong generalization across various manipulation methods. The results indicate that custom-built CNN architectures, even without transfer learning, can be effective for deepfake detection when paired with rigorous training protocols. Challenges such as data imbalance and overfitting are discussed, and directions for future research are proposed.
