"Researchers Identify a Resilient Trait of Deepfakes That Could Aid Long-Term Detection"
The computer vision and security research community has been trying to define an essential characteristic of deepfake videos or signals that may be resistant to advancements in popular facial synthesis technologies. Such technologies include autoencoder-based deepfake packages like DeepFaceLab and FaceSwap, as well as Generative Adversarial Networks (GANs) that recreate, simulate, or alter human faces. Many signs, such as lack of blinking, have been rendered obsolete by advances in deepfakes. In addition, the potential use of digital provenance techniques, including blockchain approaches and digital watermarking of possible source photos, would either necessitate extensive and costly changes to the existing body of available source images on the Internet or would require a significant cooperative effort among nations to develop systems of supervision and authentication. Therefore, it would be beneficial if a truly fundamental and resilient trait could be determined in image and video content that features altered, invented, or identity-swapped human faces. Researchers are seeking a trait that could be inferred directly from falsified videos, without the need for large-scale verification, cryptographic asset hashing, context-checking, plausibility evaluation, artifact-centric detection routines, or other time-consuming approaches to deepfake detection. A collaborative study between researchers in China and Australia developed a method for comparing the spatial integrity and temporal continuity of real videos to those containing deepfake content, and discovered that any deepfake interference disrupts the image's regularity, however imperceptibly. They incorporated the research's functionality into a plug-and-play module called Pseudo-fake Generator (P-fake Generator), which transforms real videos into faux-deepfake videos by perturbing them in the same way that the actual deepfake process does, but without performing any deepfake operations. Tests show that the module can be added to all existing deepfake detection systems for practically no cost, and that it significantly improves their performance. The discovery may aid in addressing another stumbling block in deepfake detection research: a lack of authentic and up-to-date datasets. Because deepfake generation is a complex and time-consuming process, the community has created a number of deepfake datasets over the last five years, many of which are outdated. This article continues to discuss the researchers' study on detecting deepfakes by creating spatio-temporal regularity disruption.