Enhanced Detection of Deepfakes through Facial Expression Processing Using Deep Learning Networks

Deepfakes imply huge potential for privacy, security, and trust. It consists of synthetic media in which a person's likeness in videos or images is replaced through deep learning techniques. This paper discusses the detection and analysis of deepfakes by employing state-of-the-art facial processing methods. The approach begins with detecting the faces and then applying face alignment, crop, resizing, and normalization to guarantee consistency. The dataset used consisted of real videos of Barack Obama and fake videos. Feature extraction and augmentation augmented this dataset to give it at least a better chance of training. Correlation-based feature Selection further did features around the eyes, which are very critical, and then classification was done using BiLSTM, GRU, and RNN models. Each model was trained for 50 epochs, returning accuracies of 97%, 98%, and 97%, respectively. It has proved to be quite efficient in deepfake detection, as the accuracy of the GRU model was the best, at 98%. The experiment results are represented in detail through training accuracy, loss graphs, confusion matrices, and classification reports for each model.