Deepfake Web Video Detection using Deep Neural Networks and LSTM Architectures

Abstract

This research focuses on detecting deepfakes using deep neural networks (DNNs). The objective is to design and evaluate a binary classifier capable of distinguishing between real and fake content. The model is trained on a subset of the DF40 dataset due to its inclusion of various deepfake synthesis techniques. The model is trained using Binary Cross-Entropy loss, and its performance is evaluated using accuracy, precision, recall, F1-score, and ROC-AUC metrics. ResNet50 and VGG16 are used to classify the human digital images, where the former outperforms VGG16 across all performance metrics in deepfake detection on the DF40 dataset. It achieves higher accuracy, precision, recall, F1-score and ROC-AUC, with fewer classification errors. Its superior generalization and stability make it more reliable and suitable for real world deepfake detection applications. Experimental results demonstrate that the model effectively captures subtle artefacts introduced by manipulation techniques, achieving strong classification performance.