The DOI matches its Crossref record; 263 of the 287 references it cites were found in a registry or an index, and 24 are not checked yet; 88 of the 165 sources its claims cite were checked for support.
Deepfake detection across image, video, and audio: a comprehensive survey with empirical evaluation of generalization and robustness
Open accessPublished version of a preprint
Abstract
Deepfakes (DFs) have emerged as a significant threat in recent years, being exploited for malicious purposes such as impersonation, misinformation dissemination, and artistic style imitation, thereby raising critical ethical and security concerns. This survey presents a comprehensive analysis of passive DF detection methods across image, video, and audio modalities, addressing critical gaps in existing literature. Unlike previous surveys which examine modalities in isolation, we explore inter-modality relationships and shared challenges. We systematically categorize detection approaches based on their underlying methodologies: forensic-based, data-driven, fingerprint-based, and hybrid techniques for visual modalities, and handcrafted versus learnable features for audio. We also extend our analysis beyond mere detection accuracy to include essential performance dimensions for real-world deployment, including generalization and robustness. Additionally, this survey provides an in-depth empirical evaluations of 50 state-of-the-art detection methods across 10 popular datasets, assessing their performance in three critical dimensions: (1) within-domain accuracy, (2) cross-domain generalization, and (3) robustness against adversarial attacks. Our experiments reveal a persistent generalization gap, with performance degradations of 15-20% in cross-domain scenarios, and vulnerability to white-box adversarial attacks exceeding 80% success rates. We also analyze the advantages and limitations of existing datasets, benchmarks, and evaluation metrics for passive DF detection. Finally, we propose future research directions to address these unexplored and emerging issues in the field of passive DF detection. This survey serves as a comprehensive resource for researchers and practitioners, providing insights into the current landscape, methodological approaches, and promising future directions in this rapidly evolving field.
Verified 2026-10-05 · 24 references not checked yet
Overview
CoChat Summary
## TL;DR
This survey reviews passive deepfake detection across image, video, and audio, emphasizing how methods relate across modalities and how they perform under real-world requirements like generalization and robustness. Its empirical evaluation of 50 methods across 10 datasets finds a persistent 15–20% cross-domain performance drop and white-box adversarial success rates above 80% against undefended models.
## Key findings
- The authors empirically evaluate 50 methods across 10 datasets and report that within-domain performance is often near-perfect, but cross-domain performance typically drops by 15–20%.
- In their verification study of reported results, several methods perform substantially worse than originally claimed; for example, UCF AUC drops from 82.4% to 70.61% on Celeb-DF, STIL drops from 89.80% to 66.89% on DFDC, and Conformer’s EER rises from 3.28% to 94.82% on ASVSpoof21.
- For generalization, the survey reports that frequency-domain methods consistently outperform spatial-only approaches, and notes that adaptive learning methods achieve stronger cross-domain results than simple augmentation, but with more complexity.
- For robustness, the paper reports that white-box attacks reach 81.06–95.62% ASR before defense, while adaptive adversarial training reduces WA ASR to 5.25% in one case and pure adversarial training reduces WA ASR from 81.06% to 37.03%.
## Methods
- The paper is a systematic survey of passive deepfake detection methods across image, video, and audio modalities, with categories including forensic-based, data-driven, fingerprint-based, and hybrid approaches.
- The literature review uses IEEE Xplore, ACM Digital Library, Google Scholar, and arXiv, and includes papers published in peer-reviewed journals or conferences from 2020 to 2025.
- The empirical section evaluates methods under three protocols: within-domain, cross-domain generalization, and adversarial robustness; it uses 10 widely adopted datasets and metrics including ACC, AUC, EER, and ASR.
## Limitations
- The paper states that current benchmark datasets have severe class imbalance, insufficient challenging scenarios such as multiple faces and occlusions, and no adversarially perturbed deepfake samples for testing.
- It explicitly notes that existing surveys lacked independent empirical verification, which this paper addresses; this is a limitation of prior work, not of the survey’s own design.
- Evident from the text: the empirical evaluation is limited to the datasets and methods selected in the paper, and the paper itself notes that some original papers did not fully report datasets or metrics, which constrained verification in some cases.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
37 Valid · 23 Not valid · 32 Not checked · 115 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“While current survey papers [5, 7, 9, 15] provide extensive taxonomies and theoretical discussions, they lack empirical evaluations that would enable researchers to understand the relative strengths and limitations of different approaches.”
What the sources say
- Partly supportedcites [5] [5] Kaur, A., Noori Hoshyar, A., Saikrishna, V., Firmin, S., Xia, F.: Deepfake video detection: challenges and opportunities. Artificial Intelligence Review 57 (6), 147 (2024) - sourceExperimental — accuracy not yet measured
Sections 1.1 and 1.3 describe taxonomies and comparisons of detection methods, while the Conclusion states that this survey generated or analysed no datasets. Section 1.2 also identifies gaps in real-world testing and uncertainty about detector performance. The paper does not establish that survey papers generally lack empirical evaluations or that this prevents researchers from understanding approaches’ relative strengths and limitations; Section 1.1 says an earlier survey compares methodologies and their pros and cons. (LLM, conf=0.82)
- Partly supportedcites [7] [7] Wang, X., Guo, H., Hu, S., Chang, M., Lyu, S.: Gan-generated faces detection: A survey and new perspectives. arxiv 2022. arXiv preprint arXiv:2202.07145 - sourceExperimental — accuracy not yet measured
The Introduction and §3 describe a taxonomy of GAN-face detection methods, and §4 says a sufficiently large-scale benchmark dataset for empirical evaluation is lacking (page numbers unavailable). But §3 also points to Table 1’s performance comparison, which reports test datasets and accuracy figures for multiple methods. The claim is supported if it means a common benchmark evaluation, but overstates the absence of empirical comparisons altogether. (LLM, conf=0.86)
- Supportedcites [9] [9] Malik, A., Kuribayashi, M., Abdullahi, S.M., Khan, A.N.: Deepfake detection for human face images and videos: A survey. Ieee Access 10 , 18757-18775 (2022) - source
The Introduction describes this paper as a review of datasets and detection techniques, and the DeepFakes Forensics-Based Techniques section says it groups and summarizes existing approaches. Although it reports results from studies it reviews, these passages describe no empirical head-to-head evaluation conducted by the survey itself. It therefore supports the claim as an example of the kind of survey cited, not as evidence that every current survey lacks such an evaluation. (LLM, conf=0.84)
- Disputedcites [15] [15] Pei, G., Zhang, J., Hu, M., Zhang, Z., Wang, C., Wu, Y., Zhai, G., Yang, J., Shen, C., Tao, D.: Deepfake generation and detection: A benchmark and survey. arXiv preprint arXiv:2403.17881 (2024) - sourceExperimental — accuracy not yet measured
The paper provides taxonomies and technical discussion, but it does not lack empirical evaluation. Section 4, “Benchmark Results” (page unavailable), evaluates representative methods on widely used datasets and compares their performance using results from the original papers. (LLM, conf=0.90)
Checked by a reasoning model — may be wrong.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“While surveys report impressive within-domain accuracies, they fail to examine how these same detectors perform when facing new data types, unseen manipulation techniques, or emerging generator architectures [5, 7, 9, 15].”
What the sources say
- Partly supportedcites [5] [5] Kaur, A., Noori Hoshyar, A., Saikrishna, V., Firmin, S., Xia, F.: Deepfake video detection: challenges and opportunities. Artificial Intelligence Review 57 (6), 147 (2024) - sourceExperimental — accuracy not yet measured
The Motivation section says only a few surveys examine real-world performance, including testing against new deepfake technologies, and that most focus on images rather than video. The Open issues section discusses poor generalisation across datasets and generative models. But the paper does not establish that existing surveys uniformly fail to examine these conditions; it says some do address cross-dataset challenges (Previous surveys). (LLM, conf=0.85)
- Partly supportedcites [7] [7] Wang, X., Guo, H., Hu, S., Chang, M., Lyu, S.: Gan-generated faces detection: A survey and new perspectives. arxiv 2022. arXiv preprint arXiv:2202.07145 - sourceExperimental — accuracy not yet measured
Section 4 (page unavailable) says a large-scale benchmark for GAN-face detection is lacking, and §6.1 identifies detection of faces from newer GAN models as an ongoing challenge. But the survey also discusses cross-model and cross-data evaluations, so it does not support the categorical claim that surveys fail to examine such conditions; it does not establish the claim about unseen manipulation techniques. (LLM, conf=0.74)
- Disputedcites [9] [9] Malik, A., Kuribayashi, M., Abdullahi, S.M., Khan, A.N.: Deepfake detection for human face images and videos: A survey. Ieee Access 10 , 18757-18775 (2022) - sourceExperimental — accuracy not yet measured
The survey does examine performance on unseen manipulations: its DNN-based techniques section reports a detector tested on unseen DeepFakes and gives its performance on an in-the-wild video dataset. Its Challenges for DeepFake Detection section also discusses unknown manipulation types, contrary to the claim that surveys fail to examine this issue. The cited passages do not establish whether it evaluates emerging generator architectures. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
- Partly supportedcites [15] [15] Pei, G., Zhang, J., Hu, M., Zhang, Z., Wang, C., Wu, Y., Zhai, G., Yang, J., Shen, C., Tao, D.: Deepfake generation and detection: A benchmark and survey. arXiv preprint arXiv:2403.17881 (2024) - sourceExperimental — accuracy not yet measured
The Introduction says earlier surveys lack discussion of new technologies, especially diffusion-based image and video generation. It does not establish that surveys fail to examine detector performance on new data types or unseen manipulations; indeed, Section 4.3 of this survey reports cross-dataset detector evaluations. The claim is supported only as to the gap concerning emerging generation methods. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“For BA, disjoint frequency ensembles [206] lower ASR from 74 . 05% to 27 . 54%”
What the sources say
- Disputedcites [206] [206] Hooda, A., Mangaokar, N., Feng, R., Fawaz, K., Jha, S., Prakash, A.: D4: Detection of adversarial diffusion deepfakes using disjoint ensembles. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 3812-3822 (2024) - sourceExperimental — accuracy not yet measured
Table 1 (listed under §7.1.3; page unavailable) reports that Boundary attack ASR rises from 6% for D4 (SIZE=1) to 8% for the disjoint D4 (SIZE=4) ensemble. If “BA” means Boundary attack, the reported direction and both figures contradict the claim; the table does not report 74.05% or 27.54%. (LLM, conf=0.84)
Checked by a reasoning model — may be wrong.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“Research from the broader computer vision field suggests that there exists an inherent trade-off between a model's ability to generalize to unseen data and its robustness against adversarial perturbations [219-222].”
What the sources say
- Supportedcites [219] [219] Zhang, H., Yu, Y., Jiao, J., Xing, E., El Ghaoui, L., Jordan, M.: Theoretically principled trade-off between robustness and accuracy. In: International Conference on Machine Learning, pp. 7472-7482 (2019). PMLR - source
The Introduction (p. 1) defines the problem in terms of test accuracy on natural and adversarial examples, and Section 2.4 (p. 4) illustrates a trade-off between natural and robust error. This supports the citing sentence’s paraphrase of a trade-off between generalization to unseen data and adversarial robustness, in the paper’s setting of restricted perturbations. (LLM, conf=0.91)
- Disputedcites [220] [220] Stutz, D., Hein, M., Schiele, B.: Disentangling adversarial robustness and generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6976-6987 (2019) - sourceExperimental — accuracy not yet measured
Section 3.5 (p. 7) explicitly contrasts the paper's findings with earlier claims of an inherent trade-off and says its experiments indicate the contrary. Section 3.6 (p. 7) concludes that robustness and generalization are not necessarily conflicting goals, although particular training strategies can exhibit a trade-off. (LLM, conf=0.91)
Checked by a reasoning model — may be wrong.
- Not checked — no full text of the sourcecites [221] [221] Li, Y., Xu, C.: Trade-off between robustness and accuracy of vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7558-7568 (2023) - source
- Partly supportedcites [222] [222] Gowda, S., Zonooz, B., Arani, E.: Conserve-update-revise to cure generalization and robustness trade-off in adversarial training. In: The Twelfth International Conference on Learning Representations (2024) - sourceExperimental — accuracy not yet measured
The Introduction (p. 1) says adversarial training often compromises performance on clean images, creating a trade-off between standard and adversarial generalization. It does not establish that the trade-off is inherent or unavoidable; the paper reports that CURE improves the balance between the two (Figure 1, p. 2). (LLM, conf=0.87)
hedgedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“None of the existing datasets incorporate adversarially perturbed DFs designed to evade detection, which is a critical real-world threat [17, 21].”
What the sources say
- Not matchedcites [17] [17] Corvi, R., Cozzolino, D., Poggi, G., Nagano, K., Verdoliva, L.: Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 973-982 (2023) -
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Partly supportedcites [21] [21] Hou, Y., Guo, Q., Huang, Y., Xie, X., Ma, L., Zhao, J.: Evading deepfake detectors via adversarial statistical consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12271-12280 (2023) - sourceExperimental — accuracy not yet measured
The Introduction says transferable adversarial examples pose a practical threat to DeepFake detection. Section 6.1 describes the datasets used in the experiments, but neither it nor the other passages establishes that no existing dataset contains adversarially perturbed deepfakes. The categorical dataset claim is therefore unsupported by this paper. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
- Not matched
106 more claims · 17 with their cited sources checked
Create a free account to see all 207 claims and to verify the 2 references still open.
- It may state two separate claims as one. 65
- Probably not a substantial claim. 56
- The quote only partly states this claim. 1
Citations
Cites
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472
Knowledge Distillation: A Survey
Gou, J., Yu, B., Maybank, S. J., & Tao, D. (2021) · International Journal of Computer Vision · doi:10.1007/s11263-021-01453-z · Cited by 3,931
Transformers in Vision: A Survey
Khan, S., Naseer, M., Hayat, M., Zamir, S. W., Khan, F. S., & Shah, M. (2022) · ACM Computing Surveys · doi:10.1145/3505244 · Cited by 3,595
Transformers in Vision: A Survey
Khan, S., Naseer, M., Hayat, M., Zamir, S. W., Khan, F. S., & Shah, M. (2021) · arXiv · doi:10.48550/arxiv.2101.01169 · Cited by 3,594
FaceForensics++: Learning to Detect Manipulated Facial Images
Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., & Niessner, M. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00009 · Cited by 3,314
Region Filling and Object Removal by Exemplar-Based Image Inpainting
Criminisi, A., Perez, P., & Toyama, K. (2004) · IEEE Transactions on Image Processing · doi:10.1109/tip.2004.833105 · Cited by 3,258
Rich Models for Steganalysis of Digital Images
Fridrich, J., & Kodovsky, J. (2012) · IEEE Transactions on Information Forensics and Security · doi:10.1109/tifs.2012.2190402 · Cited by 2,144
A survey on federated learning
Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., & Gao, Y. (2021) · Knowledge-Based Systems · doi:10.1016/j.knosys.2021.106775 · Cited by 1,899
Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics
Li, Y., Yang, X., Sun, P., Qi, H., & Lyu, S. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00327 · Cited by 1,846
Mamba: Linear-Time Sequence Modeling with Selective State Spaces
Gu, A., & Dao, T. (2023) · arXiv · doi:10.48550/arxiv.2312.00752 · Cited by 1,579
Cited by
Bui, N., Nguyen, H., Phan, K., Ngo, N., Tran, A., Nguyen, T., Nguyen, H., Doan, T., Nguyen, H., & Le, T. (2026) · Computer Vision and Image Understanding · doi:10.1016/j.cviu.2026.104867
Metadata
- Authors
- Nguyen-Le, H.-H., Tran, V.-T., Nguyen, D.-T., & Le-Khac, N.-A.
- Affiliations
- University College Dublin (IE)
- Trinity College Dublin (IE)
- Ho Chi Minh City University of Science (VN)
- ORCID
- Nguyen-Le, Hong-Hanh
- Venue
- Artificial Intelligence Review
- Published
- 2026-06-12
- Volume
- 59
- Issue
- 10
- ISSN
- 0269-2821 (ISSN-L)
- 1573-7462
- OpenAlex
- W4415902342
- Cited by
- 28
- References
- 245
- Citations by year
- 2026: 18
- FWCI
- 28.14
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor, 68 d)
- Locations
- Artificial Intelligence Review (publishedVersion, cc-by)
- Research Square (submittedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion, cc-by-nc-sa)
- arXiv (Cornell University)
- Funders
- University College Dublin
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Physical Sciences › Computer Science › Signal Processing › Biometric Identification and Security
- Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
- Keywords
- deepfake detection, deepfakes, image editing, video manipulation, sound manipulation, domain generalization, adversarial robustness, white-box attacks, hybrid detection, evaluation metrics, misinformation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Deepfake detection across image, video, and audio: a comprehensive survey with empirical evaluation of generalization and robustness. https://app.cochat.ai/papers/10.1007/s10462-026-11608-4
- The paper
- Nguyen-Le, H.-H., Tran, V.-T., Nguyen, D.-T., & Le-Khac, N.-A. (2026). Deepfake detection across image, video, and audio: a comprehensive survey with empirical evaluation of generalization and robustness. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11608-4
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.