Enhanced Preprocessing Stage For Feature Extraction of Deepfake Detection Based on Deep Learning Methods

Biometric identities are at risk of deepfake facial manipulation. Over the past few years, deepfake detection has been the subject of extensive research, most which centered on the application of machine learning strategies. However, the recognition of deepfake is still one of the most challenging problems in computer vision today. Deep learning has recently gained popularity due to its potential to solve a wide range of real-world problems, including detecting deepfakes. Our research focused on improving feature extraction methods by using deep learning and comparing preprocessing methods to show how they affect detection performance. A preprocessing approach was proposed to improve the regions of interest (ROIs) used to feed Moodle feature extraction. The proposed algorithmic approach involves finding one key frame extracted from a video by using the oriented fast and rotated brief algorithm, detecting the face oval region using the DLIB library as an ROI and performing quality improvement via contrast-limited adaptive histogram equalization (CLAHE)/adaptive histogram equalization (AHE) and comparing them. On the FaceForensics ++ dataset, CLAHE demonstrated superior performance compared with AHE with InceptionV3 as a feature extractor. The final classification result had an accuracy ratio of 89%, and the area under the curve was measured to be 77%.