Deepfake Detection Based on LSTM Networks with Facial Geometric Features

The increasing sophistication of Deepfake technology necessitates robust detection methods. This paper proposes a Deepfake detection approach utilizing Long ShortTerm Memory (LSTM) networks to analyze temporal variations in facial geometric features. By extracting precise facial landmark coordinates over video frames, we capture subtle dynamic inconsistencies characteristic of manipulated content. These landmark sequences are transformed into feature vectors, which are then fed into an LSTM network designed to model temporal patterns and distinguish between genuine and forged videos. The efficacy of our method is demonstrated through experiments on publicly available datasets. On the UADFV dataset, our approach achieves an accuracy of 90.45%, and on the FaceForensics++ (FF++) dataset, it reaches 93.50% accuracy, highlighting its potential for effective Deepfake detection.