A Comprehensive Survey of Deepfake Video Detection: Methods, Datasets, Challenges, and Future Directions

This survey has reviewed deepfake video detection methods from 2018 to 2024, covering five major paradigms: spa-tial CNN methods, temporal recurrent and hybrid methods, attention-guided and transformer methods, frequency-domain methods, and biological cue methods. Key findings are as follows. Detection accuracy on the FaceForensics++ benchmark has largely plateaued above 95% for frame-level methods and above 96% for temporal models, suggesting that FF++ is ap-proaching saturation as a challenging benchmark. EfficientNet-B4 and the proposed Attention-CNN-LSTM achieve the high-est FF++ accuracies (96.1% and 96.8%) while maintaining reasonable compression robustness. Cross-dataset generalisation remains the most significant unsolved problem. Models trained on FF++ lose 20–30 per-centage points on Celeb-DF or DFDC without fine-tuning. Biological cue methods (rPPG, lip sync) and identity-aware methods show the best cross-dataset generalisation because they exploit constraints independent of the specific generation pipeline. The emergence of diffusion-model-based deepfakes since 2022 has invalidated the frequency-domain assumptions of many existing detectors and motivates the development of generation-agnostic detection approaches. Future progress will most likely come from domain adaptation and universal fea-ture learning, multimodal audio-visual detection, proactive watermarking infrastructure, and large-scale self-supervised pretraining—directions that move beyond fitting a classifier to a single controlled dataset toward building detection systems that remain effective as generation technology continues to evolve.

[1]     P. Korshunov and S. Marcel, “Deepfakes: A new threat to face recog-nition? Assessment and detection,” arXiv:1812.08685, 2018. [2]     I. Goodfellow et al., “Generative adversarial networks,” Proc. NeurIPS, vol. 27, 2014. [3]     D. Guera and E. J. Delp, “Deepfake video detection using recurrent neural networks,” Proc. IEEE AVSS, 2018. [4]     D. Afchar, V. Nozick, J. Yamagishi, and I. Echizen, “MesoNet: A compact facial video forgery detection network,” Proc. IEEE WIFS, pp. 1–7, 2018. [5]     A. Rossler et al., “FaceForensics++: Learning to detect manipulated facial images,” Proc. IEEE/CVF ICCV, pp. 1–11, 2019. [6]     F. Chollet, “Xception: Deep learning with depthwise separable convo-lutions,” Proc. IEEE CVPR, pp. 1251–1258, 2017. [7]     Y. Li, M. Chang, and S. Lyu, “In ictu oculi: Exposing AI generated fake face videos by detecting eye blinking,” Proc. IEEE WIFS, 2018. [8]     Y. Li and S. Lyu, “Exposing deepfake videos by detecting face warping artifacts,” Proc. IEEE CVPR Workshops, 2019. [9]     X. Yang, Y. Li, and S. Lyu, “Exposing deep fakes using inconsistent head poses,” Proc. IEEE ICASSP, pp. 8261–8265, 2019. [10]    Y. Li et al., “Celeb-DF: A large-scale challenging dataset for deepfake forensics,” Proc. IEEE/CVF CVPR, pp. 3207–3216, 2020.