Deepfakes image detection using a SVM algorithm based on texture features analysis

This study proposes a system focused on the analysis of statistical characteristics of texture in the detection of deepfakes employing the Discrete Cosine Transform (DCT) to represent the images in the spectral domain, where the differences between authentic and generated images are remarkable. Then, using the DCT spectrum, the co-occurrence matrix is calculated to extract Haralick texture features that are input data for a SVM in binary classification between the authentic and deepfake images. The proposed system was evaluated in detecting deepfake images with different truncation factors, for psi=0.5 (small variability) and for psi close to 1 (high variability), obtaining accuracies of 99.98% for both. The designed framework demonstrated superior performance and low computational cost in comparison against existing systems in detecting deepfake images with wide range variability, where their efficiency is significantly better for deepfakes with small variability.