Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 40 of the 42 references it cites were found in a registry or an index; 2 of the 9 sources its claims cite were checked for support.

Deep fake detection and classification using error-level analysis and deep learning

Open access

Abstract

Due to the wide availability of easy-to-access content on social media, along with the advanced tools and inexpensive computing infrastructure, has made it very easy for people to produce deep fakes that can cause to spread disinformation and hoaxes. This rapid advancement can cause panic and chaos as anyone can easily create propaganda using these technologies. Hence, a robust system to differentiate between real and fake content has become crucial in this age of social media. This paper proposes an automated method to classify deep fake images by employing Deep Learning and Machine Learning based methodologies. Traditional Machine Learning (ML) based systems employing handcrafted feature extraction fail to capture more complex patterns that are poorly understood or easily represented using simple features. These systems cannot generalize well to unseen data. Moreover, these systems are sensitive to noise or variations in the data, which can reduce their performance. Hence, these problems can limit their usefulness in real-world applications where the data constantly evolves. The proposed framework initially performs an Error Level Analysis of the image to determine if the image has been modified. This image is then supplied to Convolutional Neural Networks for deep feature extraction. The resultant feature vectors are then classified via Support Vector Machines and K-Nearest Neighbors by performing hyper-parameter optimization. The proposed method achieved the highest accuracy of 89.5% via Residual Network and K-Nearest Neighbor. The results prove the efficiency and robustness of the proposed technique; hence, it can be used to detect deep fake images and reduce the potential threat of slander and propaganda.

Overview

CoChat Summary

## TL;DR

The paper proposes a deepfake image detection pipeline that combines error level analysis (ELA), pretrained CNN feature extraction, and SVM/KNN classification. On the Yonsei University real/fake face dataset, it reports a best accuracy of 89.5%.

## Key findings

  • The proposed framework preprocesses images with ELA, then extracts deep features using pretrained CNNs (GoogLeNet, ResNet18, SqueezeNet), and classifies them with SVM and KNN.
  • The best reported performance was 89.5% accuracy using ResNet18 features with KNN in one place, and 89.5% accuracy using ResNet18’s feature vector with SVM in the conclusion; the paper is internally inconsistent about which classifier achieved the top score.
  • For the other feature extractors, the paper reports 81% accuracy for GoogLeNet + KNN, 80.9% for GoogLeNet + SVM, 69.4% for SqueezeNet + SVM, and 68.8% for SqueezeNet + KNN.
  • The authors say the method outperformed several prior deepfake detection approaches and claim it can detect deepfake images in real time.

## Methods

  • The pipeline has three steps: (i) resize images and generate ELA to detect pixel-level alterations, (ii) extract deep features using pretrained CNNs, and (iii) classify features with SVM and KNN after hyperparameter optimization.
  • The dataset is a publicly accessible Yonsei University Computational Intelligence and Photography Lab real/fake face database, containing labeled real and fake face images with metadata such as age, gender, ethnicity, and manipulation type; manipulated images are split into easy, mid, and hard categories.
  • Performance was evaluated with accuracy, precision, recall, and F1-score, using confusion matrices and parameter tuning for both classifiers.

## Limitations

  • The paper explicitly says the proposed method is developed using image-based data and that future work should test video-based deepfake datasets.
  • The text also notes that the authors aim to obtain a real-life deepfake dataset in future work, implying limited ecological coverage in the current study.
  • Evident from the text: the reported results are somewhat hard to interpret because the paper gives inconsistent attributions of the top 89.5% result to different classifier pairings (ResNet18+KNN vs. ResNet18+SVM).

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

Integrity signals

Registry

Crossref · 10.1038/s41598-023-34629-3

Cites

40 of 40 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Workflow diagram of the proposed method.
Figure 1 Workflow diagram of the proposed method.CC BY · PubMed Central
Figure 2 Result of ELA on dataset images.
Figure 2 Result of ELA on dataset images.CC BY · PubMed Central
Figure 3 General CNN architecture.
Figure 3 General CNN architecture.CC BY · PubMed Central
Figure 4 ResNet18 architecture44.
Figure 4 ResNet18 architecture44.CC BY · PubMed Central
Figure 5 GoogLeNet architecture46.
Figure 5 GoogLeNet architecture46.CC BY · PubMed Central
Figure 6 KNN.
Figure 6 KNN.CC BY · PubMed Central
Figure 7 Possible SVM hyperplanes30.
Figure 7 Possible SVM hyperplanes30.CC BY · PubMed Central
Figure 8 Image samples from the dataset showing real and edited images.
Figure 8 Image samples from the dataset showing real and edited images.CC BY · PubMed Central
Figure 9 Results obtained from ResNet18's confusion matrix.
Figure 9 Results obtained from ResNet18's confusion matrix.CC BY · PubMed Central
Figure 10 ResNet18's confusion matrix via (a) SVM, (b) KNN.
Figure 10 ResNet18's confusion matrix via (a) SVM, (b) KNN.CC BY · PubMed Central
Figure 11 GoogLeNet’s results in terms of ACC, PRE, REC and F1-Score.
Figure 11 GoogLeNet’s results in terms of ACC, PRE, REC and F1-Score.CC BY · PubMed Central
Figure 12 Confusion matrix obtained from GoogLeNet.
Figure 12 Confusion matrix obtained from GoogLeNet.CC BY · PubMed Central
Figure 13 Results obtained from SqueezeNet's confusion matrices.
Figure 13 Results obtained from SqueezeNet's confusion matrices.CC BY · PubMed Central
Figure 14 Confusion matrix obtained from SqueezeNet.
Figure 14 Confusion matrix obtained from SqueezeNet.CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 2 Not valid · 4 Not checked · 13 Own results and interpretations

6 more claims

Create a free account to see all 19 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 4
  • It may state two separate claims as one. 4
  • The quote only partly states this claim. 1

Create a free account to see these claims.

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.

2 of 2

Cited works that were retracted

Works this paper cites that were later retracted, from the retraction notices publishers deposit. Major when the notice came out before this paper was published.

1 of 1

Citations

Cites

1 retracted

Show all 42 references

Research works printed but not linked: 6

  1. Boylan, J. F. Will Deep-Fake Technology Destroy Democracy (The New York Times, 2018).
  2. Harwell, D. Scarlett Johansson on fake AI-generated sex videos:‘Nothing can stop someone from cutting and pasting my image’. J. Washigton Post 31, 12 (2018).
  3. Schroepfer, M. J. F. Creating a data set and a challenge for deepfakes. Artif. Intell. 5, 263 (2019).
  4. Agarwal, S., Farid, H., Gu, Y., He, M., Nagano, K. & Li, H. Protecting world leaders against deep fakes. in CVPR Workshops. Vol. 1. 38 (2019).
  5. Salman, F. M. & Abu-Naser, S. S. Classification of real and fake human faces using deep learning. IJAER 6(3), 1–14 (2022).
  6. He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA. 770–778 (2016). · DOI · Paper page

Cited by

See all 155

Metadata

Authors
Rafique, R., Gantassi, R., Amin, R., Frnda, J., Mustapha, A., & Alshehri, A. H.
Affiliations
  • University of Engineering and Technology Taxila (PK)
  • Chonnam National University (KR)
  • VSB - Technical University of Ostrava (CZ)
  • Tun Hussein Onn University of Malaysia (MY)
  • Shaqra University (SA)
Corresponding author
Amin, Rashid
Venue
Scientific Reports
Published
2023-05-08
Volume
13
Issue
1
Pages
7422-7422
ISSN
2045-2322 (ISSN-L)
OpenAlex
W4375955734
Cited by
155
References
42
Citations by year
2026: 31 · 2025: 68 · 2024: 49 · 2023: 7
FWCI
14.04
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • DSpace VŠB-TUO (VŠB-TUO) (submittedVersion, cc-by)
  • Greater South Information System
Funders
  • Ministerstvo Školství, Mládeže a Tělovýchovy
  • Ostravská Univerzita v Ostravě
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, error level analysis, image modification detection, convolutional neural networks, ResNet, support vector machine, k-nearest neighbors, disinformation, social media hoaxes, hyperparameter optimization

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 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

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: Deep fake detection and classification using error-level analysis and deep learning. https://app.cochat.ai/papers/10.1038/s41598-023-34629-3
The paper
Rafique, R., Gantassi, R., Amin, R., Frnda, J., Mustapha, A., & Alshehri, A. H. (2023). Deep fake detection and classification using error-level analysis and deep learning. Scientific Reports. https://doi.org/10.1038/s41598-023-34629-3

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.