CoChat

Fighting Deepfakes by Detecting GAN DCT Anomalies

Open access
Giudice, Oliver, Guarnera, Luca, Battiato, Sebastiano · 2021 · Journal of Imaging

The DOI matches its Crossref record; 58 of the 58 references it cites were found in a registry or an index.

Abstract

To properly contrast the Deepfake phenomenon the need to design new Deepfake detection algorithms arises; the misuse of this formidable A.I. technology brings serious consequences in the private life of every involved person. State-of-the-art proliferates with solutions using deep neural networks to detect a fake multimedia content but unfortunately these algorithms appear to be neither generalizable nor explainable. However, traces left by Generative Adversarial Network (GAN) engines during the creation of the Deepfakes can be detected by analyzing ad-hoc frequencies. For this reason, in this paper we propose a new pipeline able to detect the so-called GAN Specific Frequencies (GSF) representing a unique fingerprint of the different generative architectures. By employing Discrete Cosine Transform (DCT), anomalous frequencies were detected. The β statistics inferred by the AC coefficients distribution have been the key to recognize GAN-engine generated data. Robustness tests were also carried out in order to demonstrate the effectiveness of the technique using different attacks on images such as JPEG Compression, mirroring, rotation, scaling, addition of random sized rectangles. Experiments demonstrated that the method is innovative, exceeds the state of the art and also give many insights in terms of explainability.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.3390/jimaging7080128

Cites

58 of 58 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 Example of real (a) and deepfake datasets (b) used in our experiments. The CelebA dataset was used to generate human face images with the StarGAN, AttGAN and GDWCT architectures. The FFHQ dataset was used to generate human face images with the StyleGAN and StyleGAN2 architectures.
Figure 1 Example of real (a) and deepfake datasets (b) used in our experiments. The CelebA dataset was used to generate human face images with the StarGAN, AttGAN and GDWCT architectures. The FFHQ dataset was used to generate human face images with the StyleGAN and StyleGAN2 architectures.CC BY · PubMed Central
Figure 2 Example of image generated by StyleGAN properly filtered to highlight patterns resulting from the generative process.
Figure 2 Example of image generated by StyleGAN properly filtered to highlight patterns resulting from the generative process.CC BY · PubMed Central
Figure 3 GSF that identify the generative architectures. (a) Zig-zag order after DCT transform. (b) DCT 8×8 frequencies.
Figure 3 GSF that identify the generative architectures. (a) Zig-zag order after DCT transform. (b) DCT 8×8 frequencies.CC BY · PubMed Central
Figure 4 Plot of β statistics of each involved dataset. The average β value for each i-th coefficient is reported. (a) Shows the average β trend of all datasets (real and deepfake); (b) Shows the average β trend of StyleGAN and StyleGAN2 compared to the real image dataset used for their creation (FFHQ); (c) Shows the average β trend of StarGAN, AttGAN and GDWCT compared to the real image dataset used for their creation (CelebA). For each plot, the abscissa axis represents the 64 coefficients of the 8×8 block, while the ordinate axis are the respective inferred β values (in our case the average of the β values computed for all images of the respective datasets).
Figure 4 Plot of β statistics of each involved dataset. The average β value for each i-th coefficient is reported. (a) Shows the average β trend of all datasets (real and deepfake); (b) Shows the average β trend of StyleGAN and StyleGAN2 compared to the real image dataset used for their creation (FFHQ); (c) Shows the average β trend of StarGAN, AttGAN and GDWCT compared to the real image dataset used for their creation (CelebA). For each plot, the abscissa axis represents the 64 coefficients of the 8×8 block, while the ordinate axis are the respective inferred β values (in our case the average of the β values computed for all images of the respective datasets).CC BY · PubMed Central
Figure 5 CTF-DCT approach: (a) Dataset used for our experiments; (b) Discrete Cosine Transform (DCT) of a given image at each 8 × 8 blocks; (c) GAN Specific Frequencies (GSF) that identify involved GAN architectures.
Figure 5 CTF-DCT approach: (a) Dataset used for our experiments; (b) Discrete Cosine Transform (DCT) of a given image at each 8 × 8 blocks; (c) GAN Specific Frequencies (GSF) that identify involved GAN architectures.CC BY · PubMed Central
Figure 6 Average Accuracy results (%) obtained for the binary classification task employing only the GSF. 700 images were employed for testing, 200 images for training, 5-fold cross validated, classes are balanced.
Figure 6 Average Accuracy results (%) obtained for the binary classification task employing only the GSF. 700 images were employed for testing, 200 images for training, 5-fold cross validated, classes are balanced.CC BY · PubMed Central
Figure 7 GSF and classification accuracy results (%) obtained for each binary classification task.
Figure 7 GSF and classification accuracy results (%) obtained for each binary classification task.CC BY · PubMed Central
Figure 8 Abnormal frequencies inspection. (a) Image example from the StarGAN dataset; (b) Fourier Spectra of the input image (a); (c) Abnormal frequency shown by means of GSF amplification.
Figure 8 Abnormal frequencies inspection. (a) Image example from the StarGAN dataset; (b) Fourier Spectra of the input image (a); (c) Abnormal frequency shown by means of GSF amplification.CC BY · PubMed Central
Figure 9 Examples of ATTGAN, GDWCT, STARGAN, STYLEGAN, STYLEGAN2 images in which we applied different attacks: Random Square, Gaussian Blur, Rotation, Mirror, Scaling and JPEG Compression. They were also applied in the real dataset (CelebA and FFHQ).
Figure 9 Examples of ATTGAN, GDWCT, STARGAN, STYLEGAN, STYLEGAN2 images in which we applied different attacks: Random Square, Gaussian Blur, Rotation, Mirror, Scaling and JPEG Compression. They were also applied in the real dataset (CelebA and FFHQ).CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

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.

The self-check needs the paper’s full text.

Citations

Cites

Cited by

See all 118

Metadata

Authors
Giudice, O., Guarnera, L., & Battiato, S.
Affiliations
University of Catania (IT)
Corresponding author
Giudice, Oliver
Venue
Journal of Imaging
Publisher
MDPI AG
Published
2021-07-30
Volume
7
Issue
8
Pages
128
ISSN
2313-433X (ISSN-L)
OpenAlex
W3122852338
Cited by
118
References
58
Citations by year
2026: 17 · 2025: 32 · 2024: 29 · 2023: 20 · 2022: 16 · 2021: 4
FWCI
6.20
Type
journal article
Language
en
Open access
Locations
  • Journal of Imaging (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, other-oa)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
Keywords
deepfake detection, generative adversarial networks, discrete cosine transform, explainable AI, robustness to image attacks, image forgery detection

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.

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.