deepfake-detectors-pursue-new-ground-latent-diffusion-models-and-gans

Opinion Of late, the deepfake detection research community, which has since late 2017 been occupied almost exclusively with the autoencoder-based framework that premiered at that time to such public awe (and dismay), has begun to take a forensic interest in less stagnant architectures, including latent diffusion models such as DALL-E 2 and Stable Diffusion, as well as the output of Generative Adversarial Networks (GANs). For instance, in June, UC Berkeley published the results of its research into the development of a detector for the output of the then-dominant DALL-E 2. What seems to be driving this growing interest is the sudden evolutionary jump in the capability and availability of latent diffusion models in 2022, with the closed-source and limited-access release of DALL-E 2 in spring, followed in late summer by the sensational open sourcing of Stable Diffusion by stability.ai. GANs have also been long-studied in this context, though less intensively, since it is very difficult to use them for convincing and elaborate video-based recreations of people; at least, compared to the by-now venerable autoencoder packages such as FaceSwap and DeepFaceLab – and the latter's live-streaming cousin, DeepFaceLive. In either case, the galvanizing factor appears to be the prospect of a subsequent developmental sprint for video synthesis. The start of October – and 2022's major conference season – was characterized by an avalanche of sudden and unexpected solutions to various longstanding video synthesis bugbears: no sooner had Facebook released samples of its own text-to-video platform, than Google Research quickly drowned out that initial acclaim by announcing its new Imagen-to-Video T2V architecture, capable of outputting high resolution footage (albeit only via a 7-layer network of upscalers).