Published Paper Details:

The project focuses on detecting deepfake images using a dataset containing real and fake images. This employs transfer learning, a technique in which a pre-trained model (EfficientNetB0) is used as a starting point and fine-tuned on a new dataset (comprising real and deepfake images). This approach leverages the knowledge learned by the pre-trained model on a large dataset (ImageNet) and adapts it to the specific task of deepfake detection. For face detection and alignment, the MTCNN (Multi-task Cascaded Convolutional Neural Network) algorithm is utilized. MTCNN is a state-of-the-art deep learning model known for its accuracy in detecting faces and facial landmarks. It is used to locate and extract faces from frames of a video, which are then processed and fed into the deepfake detection model. Data augmentation is applied to the dataset using techniques such as rotation, shifting, shearing, zooming, and flipping. This process helps increase the diversity of the training data, improving the model's ability to generalize to unseen images. The Deepfake detection model consists of a base EfficientNetB0 model followed by a Global Average Pooling 2D layer, a Dropout layer for regularization, and a Dense layer with a sigmoid activation function for binary classification (real or fake). The model is trained using the Adam optimizer with a varied range of learning rates and binary cross-entropy loss. During training, the model's performance is monitored using various metrics such as accuracy, precision, recall, and F1-score. Early stopping is employed to prevent overfitting, and the best model is saved using Model Checkpoint to ensure that the model with the lowest validation loss is retained. In conclusion, the project integrates cutting-edge technologies such as transfer learning, deep learning models like EfficientNetB0, and advanced face detection algorithms like MTCNN to develop a robust deepfake detection system.