Authors: Marko Brodarič, Vitomir Štruc, Peter Peer
Abstract: Due to the recent advances in generative deep learning, numerous techniques have been proposed in the literature that allow for the creation of so-called deepfakes, i.e., forged facial images commonly used for malicious purposes. These developments have triggered a need for effective deepfake detectors, capable of identifying forged and manipulated imagery as robustly as possible. While a considerable number of detection techniques has been proposed over the years, generalization across a wide spectrum of deepfake-generation techniques still remains an open problem. In this paper, we study a representative set of deepfake generation methods and analyze their performance in a cross-dataset setting with the goal of better understanding the reasons behind the observed generalization performance. To this end, we conduct a comprehensive analysis on the FaceForensics++ dataset and adopt Gradient-weighted Class Activation Mappings (Grad-CAM) to provide insights into the behavior of the evaluated detectors. Since a new class of deepfake generation techniques based on diffusion models recently appeared in the literature, we introduce a new subset of the FaceForensics++ dataset with diffusion-based deepfake and include it in our analysis. The results of our experiments show that most detectors overfit to the specific image artifacts induced by a given deepfake-generation model and mostly focus on local image areas where such artifacts can be expected. Conversely, good generalization appears to be correlated with class activations that cover a broad spatial area and hence capture different image artifacts that appear in various part of the facial region.
[PDF]