CoChat

Deepfake Media Forensics: Status and Future Challenges

Open access

The DOI matches its Crossref record; 205 of the 205 references it cites were found in a registry or an index; 10 of the 14 sources its claims cite were checked for support.

Abstract

The rise of AI-generated synthetic media, or deepfakes, has introduced unprecedented opportunities and challenges across various fields, including entertainment, cybersecurity, and digital communication. Using advanced frameworks such as Generative Adversarial Networks (GANs) and Diffusion Models (DMs), deepfakes are capable of producing highly realistic yet fabricated content, while these advancements enable creative and innovative applications, they also pose severe ethical, social, and security risks due to their potential misuse. The proliferation of deepfakes has triggered phenomena like "Impostor Bias", a growing skepticism toward the authenticity of multimedia content, further complicating trust in digital interactions. This paper is mainly based on the description of a research project called FF4ALL (FF4ALL-Detection of Deep Fake Media and Life-Long Media Authentication) for the detection and authentication of deepfakes, focusing on areas such as forensic attribution, passive and active authentication, and detection in real-world scenarios. By exploring both the strengths and limitations of current methodologies, we highlight critical research gaps and propose directions for future advancements to ensure media integrity and trustworthiness in an era increasingly dominated by synthetic media.

Overview

CoChat Summary

TL;DR

This review/position paper surveys deepfake media forensics within the FF4ALL project, covering generation, detection, attribution, passive/active authentication, continual learning, explainability, and ethics. It argues that detection must adapt to evolving generative models, real-world compression and adversarial conditions, and calls for scalable, interpretable, and ethically governed tools.

Key findings

  • The paper synthesizes detection approaches including undirected deep learning, visual artifacts, biological artifacts, and texture/spatio-temporal consistency; attribution aims to identify architecture and even model instance, with one cited method using ResNET-18 plus metric learning to distinguish 100 different StyleGAN2-ADA instances.
  • It reports dataset scales: WildFake has over 3.6 million images (2.6 M fake and 1 M real); Artifact has over 2.4 million images (1.5 million synthetic from 25 methods); DIFF has over 500,000 images from 13 diffusion models; DFDC has nearly 120,000 videos (100,000 fake); ASVspoof 2019 contains 78 speakers (33 male, 45 female).
  • Continual learning results include DFWF achieving an Average Equal Error Rate reduction of 81.63% compared to fine-tuning on ASVspoof 2019, and a Transformer-based continual learning method achieving 90–95% AUC on unseen datasets using ASVspoof datasets and over 2 million synthetic samples.
  • Real-world challenges include social media compression degrading forensic traces, adversarial and laundering attacks evading detectors, active authentication via cryptographic signatures/BLS and JPEG Trust, and ethical/privacy risks alongside “Impostor Bias.”

Methods

  • Design: review/position paper based on the FF4ALL project, extending previous work and providing a structured comparison of deepfake detection approaches, datasets, continual learning, explainability, attribution, passive/active authentication, and realistic scenarios.
  • Data or sample: no new primary dataset or experiment is reported; the paper draws on cited state-of-the-art datasets and methods across image, audio, video, and multimodal modalities.
  • Main analysis: narrative synthesis and tables (e.g., Table 1 on project topics, Table 3 comparing detection methods), identifying strengths, weaknesses, research gaps, and future directions.

Limitations

  • The paper states that scalability, real-time applicability, and robustness in dynamic environments remain persistent challenges; many state-of-the-art techniques require high computational resources and extensive labeled datasets, suffer latency, and are vulnerable to adversarial/laundering attacks. It also notes the lack of universally accepted benchmarks, sensitivity to media compression, insufficient interpretability for legal/forensic use, and ethical/privacy concerns.
  • The provided text makes evident that several sections are empty or truncated (e.g., Introduction, Conclusions, and some subsections), so not all details can be assessed from this version.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/jimaging11030073

Cites

205 of 205 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Generic diagram of the deepfake creation process.
Figure 1 Generic diagram of the deepfake creation process.CC BY · PubMed Central
Figure 2 (a) The GAN framework consists of a Generator (G) that creates synthetic data x¯ from random noise z, aiming to learn the data distribution pdata, and a Discriminator (D) that distinguishes between real and generated data. Both models are trained simultaneously in an adversarial manner; (b) The Diffusion Model [8] uses a fixed Markov chain to add Gaussian noise to data, approximating the posterior distribution q(xt|xt−1)fort=1,…,T. The goal is to learn the reverse process pθ(xt−1|xt) to generate data by reversing the noise-adding chain, where x1,…,xT are latent variables with the same dimensionality as x0.
Figure 2 (a) The GAN framework consists of a Generator (G) that creates synthetic data x¯ from random noise z, aiming to learn the data distribution pdata, and a Discriminator (D) that distinguishes between real and generated data. Both models are trained simultaneously in an adversarial manner; (b) The Diffusion Model [8] uses a fixed Markov chain to add Gaussian noise to data, approximating the posterior distribution q(xt|xt−1)fort=1,…,T. The goal is to learn the reverse process pθ(xt−1|xt) to generate data by reversing the noise-adding chain, where x1,…,xT are latent variables with the same dimensionality as x0.CC BY · PubMed Central
Figure 3 Taxonomy of Deepfake Detection Methods. Detection techniques are categorized into four primary groups: general network-based approaches, methods focusing on visual artifacts, approaches targeting biological inconsistencies, and techniques leveraging texture and spatio-temporal consistency. Each category addresses specific characteristics of manipulated content.
Figure 3 Taxonomy of Deepfake Detection Methods. Detection techniques are categorized into four primary groups: general network-based approaches, methods focusing on visual artifacts, approaches targeting biological inconsistencies, and techniques leveraging texture and spatio-temporal consistency. Each category addresses specific characteristics of manipulated content.CC BY · PubMed Central
Figure 4 Example of deepfake detection features. The full face (a) exhibits visual artifacts, such as unnatural pixel formations around the facial features (e.g., glasses and skin edges). These are indicative of irregularities introduced by generative algorithms or compression distortions. The zoomed-in region (b) around the mouth shows potential texture inconsistencies, such as unnatural blending of lip textures and a lack of smooth transitions typical of natural skin and lip patterns.
Figure 4 Example of deepfake detection features. The full face (a) exhibits visual artifacts, such as unnatural pixel formations around the facial features (e.g., glasses and skin edges). These are indicative of irregularities introduced by generative algorithms or compression distortions. The zoomed-in region (b) around the mouth shows potential texture inconsistencies, such as unnatural blending of lip textures and a lack of smooth transitions typical of natural skin and lip patterns.CC BY · PubMed Central
Figure 5 A conceptual pipeline for deepfake detection and model recognition, illustrating the process of identifying whether an image is real or synthetic, determining the generative architecture, and tracing the specific model instance used for creation.
Figure 5 A conceptual pipeline for deepfake detection and model recognition, illustrating the process of identifying whether an image is real or synthetic, determining the generative architecture, and tracing the specific model instance used for creation.CC BY · PubMed Central
Figure 6 The TrueFace Dataset [158], comprising 80k fake images generated with StyleGAN models and 70k real images, of which 60k have been also shared on three distinct social networks.
Figure 6 The TrueFace Dataset [158], comprising 80k fake images generated with StyleGAN models and 70k real images, of which 60k have been also shared on three distinct social networks.CC BY · PubMed Central

Verification

Claims the paper makes

9 Valid · 0 Not valid · 3 Not checked · 13 Own results and interpretations

20 more claims · 4 with their cited sources checked

Create a free account to see all 25 claims and every source verdict.

See all claims — free account

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Items 201–205 of 205

Research works printed but not linked: 9

  1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative Adversarial Nets. Adv. Neural Inf. Process. Syst., 27.
  2. Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., and Ferrer, C.C. (2020). The deepfake detection challenge (DFDC) dataset. arXiv.
  3. Barrington, S., Bohacek, M., and Farid, H. (2024). DeepSpeak Dataset v1.0. arXiv.
  4. Kang, S., Shi, Z., and Zhang, X. (2023, January 23–29). Poisoning Generative Replay in Continual Learning to Promote Forgetting. Proceedings of the International Conference on Machine Learning, PMLR, Honolulu, HI, USA.
  5. Nguyen Le, T.D., Teh, K.K., and Dat Tran, H. (2024). Continuous Learning of Transformer-based Audio Deepfake Detection. arXiv.
  6. Zhang, X., Yi, J., and Tao, J. (2024). EVDA: Evolving Deepfake Audio Detection Continual Learning Benchmark. arXiv.
  7. National Research Council, Policy and Global Affairs, Division on Engineering and Physical Sciences, Committee on Science, Technology, and Law, Committee on Applied and Theoretical Statistics, and Committee on Identifying the Needs of the Forensic Sciences Community (2009). Strengthening Forensic Science in the United States: A path forward, National Academies Press.
  8. Koopman, M., Rodriguez, A.M., and Geradts, Z. (2018, January 29–31). Detection of DeepFake Video Manipulation. Proceedings of the Irish Machine Vision and Image Processing Conference (IMVIP), Belfast, Northern Ireland.
  9. Tsipras, D., Santurkar, S., Engstrom, L., and Turner, A. (2018). Robustness May Be at Odds with Accuracy. arXiv.

Cited by

See all 113

Metadata

Authors
Amerini, I., Barni, M., Battiato, S., Bestagini, P., Boato, G., Bruni, V., Caldelli, R., De Natale, F., De Nicola, R., Guarnera, L., Mandelli, S., Majid, T., Marcialis, G. L., Micheletto, M., Montibeller, A., Orrù, G., Ortis, A., Perazzo, P., Puglisi, G., Purnekar, N., Salvi, D., Tubaro, S., Villari, M., & Vitulano, D.
Affiliations
  • Sapienza University of Rome (IT)
  • University of Siena (IT)
  • University of Catania (IT)
  • Politecnico di Milano (IT)
  • University of Trento (IT)
  • Consorzio Nazionale Interuniversitario per le Telecomunicazioni (IT)
  • Department of Information Engineering and Computer Science, University of Trento, 38123 Trento, Italy
  • IMT School for Advanced Studies, 55100 Lucca, Italy
  • Department of Computer, Control and Management Engineering, Sapienza University of Rome, 00185 Roma, Italy
  • University of Cagliari (IT)
  • University of Pisa (IT)
  • University of Messina (IT)
Corresponding author
Guarnera, Luca
Venue
Journal of Imaging
Publisher
MDPI AG
Published
2025-02-28
Volume
11
Issue
3
Pages
73
ISSN
2313-433X (ISSN-L)
OpenAlex
W4408054937
Cited by
113
References
197
Citations by year
2026: 68 · 2025: 45
FWCI
39.26
Type
journal article
Language
en
Open access
Locations
  • Journal of Imaging (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Use Siena air (University of Siena) (submittedVersion, cc-by)
  • CINECA IRIS Institutial research information system (University of Pisa) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • UNICA IRIS Institutional Research Information System (University of Cagliari) (submittedVersion, other-oa)
  • IRIS Research product catalog (Sapienza University of Rome) (submittedVersion, cc-by)
  • Iris (University of Trento) (submittedVersion, cc-by)
Funders
  • SERICS (PE00000014)
  • European Union—NextGenerationEU (PE00000014)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
Keywords
deepfake detection, synthetic media, generative adversarial networks, diffusion models, media authentication, passive authentication, continuous authentication, media integrity

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 2026-09-26 (prompt extract-v2).

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

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.