Faceswap Deepfakes Detection using Novel Multi-directional Hexadecimal Feature Descriptor

With the growing number of sophisticated deep learning algorithms and fake video generation applications, it is now possible to create highly realistic deepfake videos. Faceswap is the most commonly employed deepfakes approach, which is challenging to detect due to variations in the facial skin tone, illumination conditions, presence of accessories like glasses on the face, compression artifacts, etc. Existing local texture descriptors have achieved better performance on face recognition applications; however, they compute only the limited directional information while ignoring the magnitude details. This motivated us to develop a robust local texture descriptor to extract more directional and magnitude details from the adjacent pixels to effectively represent the video frames. For this purpose, we proposed a robust multi-directional hexadecimal feature descriptor (MDHFD) by combining the local hexadecimal pattern (LHeXDP) and Local Adjacent Neighborhood Magnitude Pattern (LANMP). LHeXDP calculates the orientation-based pattern by computing 1
^st
- and 2
^nd
-order derivatives at 0°, 45°, 90°, and 135° angles from each center pixel. LANMP computes the magnitude information from each central pixel to its adjacent pixels in horizontal, vertical, diagonal, and diagonal-back directions. Histograms of both the LHeXDP and LANMP are fused to compute a multi-directional feature vector, which is used to train a support vector machine to classify between the original or faceswap deepfakes video. We measured the performance of our system on the challenging faceswap subset of a diverse and large-scale Face Forensic++ and World Leaders datasets. Experimental results illustrate that the proposed method outperforms state-of-art methods for the detection of faceswap deepfakes.