DeepFake is a concern for celebrities and everyone because it is simple to create. DeepFake images, especially high-quality ones, are difficult to detect using people, local descriptors, and current approaches. On the other hand, video manipulation detection is more accessible than an image, which many state-of-the-art systems offer. Moreover, the detection of video manipulation depends entirely on its detection through images. Many worked on DeepFake detection in images, but they had complex mathematical calculations in preprocessing steps, and many limitations, including that the face must be in front, the eyes have to be open, and the mouth should be open with the appearance of teeth, etc. Also, the accuracy of their counterfeit detection in all previous studies was less than what this paper achieved, especially with the benchmark Flickr faces high-quality dataset (FFHQ). This study proposed, a new, simple, but powerful method called image Re-representation by combining the local binary pattern of multiple-channel (IR-CLBP-MC) color space as an image re-representation technique improved DeepFake detection accuracy. The IRCLBP- MC is produced using the fundamental concept of the multiple-channel of the local binary pattern (MCLBP), an extension of the original LBP. The primary distinction is that in our method, the LBP decimal value is calculated in each local patch channel, merging them to re-represent the image and producing a new image with three color channels. A pretrained convolutional neural network (CNN) was utilized to extract the deep textural features from twelve sets of a dataset of IR-CLBP-MC images made from different color spaces: RGB, XYZ, HLS, HSV, YCbCr, and LAB. Other than that, the experimental results by applying the overlap and non-overlap techniques showed that the first technique was better with the IR-CLBP-MC, and the YCbCr image color space is the most accurate when used with the model and for both datasets. Extensive experimentation is done, and the high accuracy obtained are 99.4% in the FFHQ and 99.8% in the CelebFaces Attributes dataset (Celeb-A).
This work is licensed under a