Human deepfake-face detection varies by generator and training
Original title: From GAN-Based Models to GPT-5: A Taxonomy-Guided Study of Human Performance in Detecting Deepfake Face Images
Among 524 participants, model-specific training generally produced larger gains, and accuracy held five months later.
A study proposes the Deepfake Face Artifact Taxonomy and tests human detection of face images from five generators: three GAN-based, one diffusion-based, and GPT-5. Accuracy depended strongly on the model, and newer generators were not necessarily harder to detect. Training improved performance in most conditions. A five-month follow-up with 342 participants found no significant drop in accuracy or rise in perceived stress.
- Model-specific materials generally beat general-purpose training numerically.
- The online experiment had 524 participants and the follow-up had 342.
- Improvement after training varied by generative model.
From the original article
1 min read at arxiv.orgGenerative models for deepfakes have rapidly evolved from GAN-based architectures to diffusion models and multimodal large language models (MLLMs). However, empirical research on human performance in detecting deepfake face images has not kept pace.
By Zhiqing Guo · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.