A Unified CNN-Vision Transformer Framework for Deepfake Detection
Abstract
Deepfakes in the present era have developed to such an extent that they almost precisely replicate the visuals and even the behaviour of real humans. When used with ill intent, they pose a severe threat to authentic media and digital security. To detect deepfakes, the proposed study suggests using a hybrid model, a deep learning architecture that combines CNNs and LSTM networks. The LSTM unit in a long short-term memory network tracks temporal dependencies between adjacent frames to reveal motion inconsistencies, blink issues, or facial expression flaws. At the same time, the other portion of our model, named the convolutional neural network (CNN), gets spatial features such as facial texture, edges, or artefacts from each frame. The Face Forensics dataset available on Kaggle, a mix of real and faked videos, is used for the model's training and testing phases. The dataset was pre-processed with four techniques: face detection, frame extraction, resizing, and normalization. A detailed analysis of the findings revealed that, compared to standard CNN methods, the new hybrid CNN-LSTM architecture performed better, with an accuracy of 94.8%, a precision of 95.2%, a recall of 94.1%, and an F1-score of 94.6%. The data show that incorporating both temporal and spatial features into deepfake detection improves detection effectiveness; therefore, the proposed model is well-suited for real-world use cases such as digital forensic analysis.