New deepfake survey finds top detectors fail on unseen generators

A comprehensive survey paper accepted in ACM Computing Surveys, led by Florinel-Alin Croitoru and nine other researchers, provides an updated landscape of deepfake generation and detection in the generative AI era. The work covers all deepfake media types—image, video, audio, and multimodal—constructing detailed taxonomies of both generation and detection methods. The authors also gather major datasets used in deepfake detection research and provide updated rankings of the best-performing detectors on the most popular benchmarks.

Crucially, the team developed a novel multimodal benchmark to evaluate deepfake detectors on out-of-distribution content. The results are sobering: state-of-the-art detectors consistently fail when faced with deepfakes generated by unseen generators (i.e., models not included in their training data). This finding underscores a fundamental weakness in current detection approaches—they lack robustness to novel synthesis techniques. The paper's project page and benchmark are publicly available, offering a resource for future research aimed at building more generalizable detection systems.