Answers are grounded in this paper and cite their sources.
Deepfakes are highly realistic synthetic images, videos and audio produced with deep learning, and they have moved from a research curiosity to a serious threat to privacy, elections and financial security. This paper presents a comparative review of deepfake generation and detection, synthesised from three recent studies: a critical review of state-of-the-art detection approaches (Soundarya and Gururaj, 2026), a benchmark and survey of generation and detection (Pei et al., 2026), and the literature they cover. We organise the field into generation (GANs, VAEs, diffusion models, NeRF), detection (feature-based, deep-learning, frequency, temporal, multimodal and audio methods), benchmark datasets, and reported performance. We find that detectors reach over 98% accuracy on in-dataset tests such as FaceForensics++, but cross-dataset AUC often falls to the 65–85% range on harder sets such as DFDC. Generalisation to unseen forgeries, robustness to compression, lack of a unified evaluation protocol and the rapid rise of diffusion-based generation remain the central open problems
No takes yet. Share an insight, caveat, or question.
Shinde et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: