Anisotropic multiresolution analyses for deepfake detection

Generative Adversarial Networks (GANs) can be misused to fabricate elaborate lies. The threat posed by GANs has sparked the need to discern between genuine and fabricated content. We argue that since GANs primarily utilize isotropic convolutions to generate their output, they leave clear traces, their fingerprint, in the coefficient distribution on sub-bands extracted by anisotropic multiresolution transforms. We employ the fully separable wavelet transform and anisotropic multiwavelets to obtain anisotropic features to feed to lightweight convolutional neural network classifiers. The proposed approach is capable of considerably improving the state-of-the-art in detecting fully GAN-generated images. It is particularly resilient to common perturbations, such as compression, noise or blur. We find that anisotropic transforms, when combined with XceptionNet, also significantly enhance the state-of-the-art in detecting partially manipulated images. • We are the first to use anisotropic transformations for deepfake detection. • A lightweight model fed by the proposed transformations considerably improves state-of-the-art baselines. • Anisotropic transformations are robust against several image degradations, especially against image compression. • We have evaluated our model on four datasets: CelebA, LSUN bedroom, FFHQ, and FaceForensics++.