Published Paper Details:

Our research introduces a robust method to address the growing prevalence of deepfake content, which poses a significant threat to the integrity of multimedia. We present an integrated approach that encompasses both the generation and detection of deepfake technology. Our system leverages state-of-the-art few-shot learning techniques to create personalized and highly realistic talking head models from a limited number of photographs. By employing deep Convolutional Neural Networks (ConvNets) trained on a vast video dataset, our system can produce convincing video sequences mimicking facial expressions and vocal nuances from just a single image. Through extensive meta-learning and adversarial training, our system initializes the parameters of both the generator and discriminator in a person-specific manner, facilitating rapid adaptation and training despite the intricacies of the task. Building upon this foundation, we propose a novel deepfake detection framework that integrates convolutional neural networks (CNNs) to capture temporal dependencies, residual networks (ResNets) for extracting spatial features, and long short-term memory (LSTM) networks. This hybrid architecture effectively combines LSTM-CNN's ability to recognize dynamic facial expressions and movements across frames with ResNet's proficiency in capturing complex facial patterns and contextual information. Furthermore, transfer learning techniques, including pre-training on a diverse dataset and fine-tuning on deepfake-specific data, are utilized to enhance model generalization.