Pith. sign in

REVIEW 4 major objections 4 minor 18 references

An Independent Discriminant Network Towards Identification of Counterfeit Images and Videos

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Face-crop CNN flags GAN-made images and video frames as counterfeit at 90.08 percent accuracy.

desk verdict Table 1 contradicts the paper's own 128-image test set, so the 90.08% accuracy claim is unsupported by the manuscript's numbers. read the letter →

arxiv 2506.05377 v1 pith:IKGJ4TVD submitted 2025-05-30 cs.CV

Shayantani Kar , B. Shresth Bhimrajka , Aditya Kumar , Sahil Gupta , Sourav Ghosh , Subhamita Mukherjee , Shauvik Paul This is my paper
classification cs.CV
keywords ConvolutionalNeuralNetworkGenerativeAdversarialNetworksInceptionResNetV2Net50EfficientB0DeepFakeDetectionImageForensics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Generative adversarial networks can now create images and video frames that humans cannot reliably tell apart from photographs, and this paper argues that an automated forensic classifier can close much of that gap. The proposed independent discriminant network extracts faces from images and video frames using Dlib's MMOD detector, then feeds the cropped faces to a modified InceptionResNetV2 convolutional network that outputs the probability that each face is real. Trained on the Nvidia and Kaggle GAN-generated face datasets, the model is reported to reach 90.08% accuracy, outperforming ResNet50 and EfficientNetB0 in the authors' comparison. The same model is packaged into a mobile application and a cloud backend so that non-experts can upload an image or video and receive a real-or-counterfeit verdict for each detected face.

What carries the argument

The load-bearing component is InceptionResNetV2, a 164-layer convolutional network that fuses Inception modules, which apply filters of several sizes at once, with residual connections that let gradients flow through the deep stack; the paper adapts it by adding a dense sigmoid layer that maps the learned features to the probability that an input face is real. Around this model sits a preprocessing pipeline: Dlib's MMOD-based CNN detects faces, each video contributes ten sampled frames, and faces are cropped to the native 340x340 video resolution before resizing to 256x256 for training. Residual connections reduce the training-time degradation problem of deep networks, while the multi-scale Inception filters let the model capture both fine pixel-level artifacts and larger structural traces left by GAN generators.

What would settle it

Run the trained InceptionResNetV2 weights on a held-out set of face crops from GAN families and video manipulations not present in the Kaggle and Nvidia datasets; if accuracy falls well below 90%, the headline claim is dataset-bound. A simpler internal check: recompute accuracy from the paper's own confusion matrix (true positives 120, true negatives 83, false positives 87, false negatives 45), which covers 256 samples and yields roughly 79% rather than 90.08%, so the source of that discrepancy should be resolved before relying on the number.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single binary classifier, built on InceptionResNetV2 with a dense sigmoid output, can independently discriminate genuine photographs from GAN-generated images and video frames once both are reduced to face crops. On the authors' test evaluation, the classifier reaches 90.08% accuracy on both the Nvidia and Kaggle datasets, with precision 0.9375, recall 0.7273, and F-score 0.8191 on the Kaggle set; this is better than ResNet50's 82.08% and EfficientNetB0's 83.01% in the same comparison. The paper further claims that video detection follows from classifying randomly sampled frames from a video, and that the resulting system is deployable as a mobile application backed by an API hosted on Cloud Run.

Load-bearing premise

The load-bearing premise is that the face crops extracted from the Kaggle and Nvidia datasets, after face detection and cropping (Sections 3.1.1 and 3.1.2), are representative of the GAN-made counterfeit images and videos the model claims to detect in the wild, and that the test set behind the 90.08% figure is trustworthy; the data-availability statement at the end of the paper, saying no datasets were generated or analysed, leaves that premise unsupported.

Editorial extensions

If this is right

  • Forensic examiners could use the model as a fast screening step, since only ten sampled frames per video are needed to produce a per-frame real-or-counterfeit decision.
  • A mobile front end with a Flask API lets journalists, law-enforcement officers, and ordinary users check an image or video without the upload being stored, preserving privacy.
  • Packaging the model as a Docker container on Cloud Run gives automatic scaling to 100 concurrent containers, so the detection service can handle real-time requests without dedicated infrastructure.
  • Because the classifier works directly on face crops, it offers an image-only alternative to temporal and audio-based deepfake detectors and can be combined with them for a higher-confidence verdict.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The strongest untested implication is generalization: the 90.08% figure comes from only two GAN datasets, so a natural next experiment is to run the same frozen weights on face crops from newer generators such as StyleGAN or diffusion models; if accuracy drops sharply, the method is dataset-bound rather than a general forensic detector.
  • Because the model sees only faces detected by Dlib, it would miss manipulations of backgrounds, bodies, or objects; a full-image path alongside the face path would be a straightforward testable extension.
  • Sampling only ten frames per video trades recall for speed, so aggregating predictions over more frames or across a temporal window could catch a single manipulated frame that random sampling might skip.
  • The paper does not isolate which ingredient drives the result, so an ablation that swaps InceptionResNetV2 for another feature extractor while keeping the face-crop pipeline fixed would show whether the architecture or the preprocessing deserves the credit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The manuscript proposes a CNN-based 'independent discriminant network' to distinguish real images/videos from GAN-generated counterfeits. It compares three backbones (ResNet50, EfficientNetB0, InceptionResNetV2) on face crops from Kaggle and Nvidia GAN datasets, claims InceptionResNetV2 achieves 90.08% accuracy, and describes a Flask/React Native application deployed on Google Cloud Run. The central quantitative claim rests on a single confusion-matrix table (Table 1) and two accuracy plots.

Significance. The stated contribution—a fine-tuned InceptionResNetV2 binary discriminator for GAN-generated face crops plus a mobile/cloud application—would be a modest but potentially useful forensic screening tool if the 90.08% accuracy were reliable. The application design is a reasonable deployment sketch. However, the paper's significance is currently limited by the absence of verifiable experimental support: Table 1 is internally inconsistent, the datasets are not described in sufficient detail, no code or data are provided, and the Data Availability Statement contradicts the experimental section. The paper does not ship reproducible artifacts or parameter-free derivations that could compensate for these gaps.

major comments (4)
  1. [Section 4.4, Table 1] The only quantitative evidence for the central 90.08% accuracy claim is internally inconsistent. Section 3.1.4 and Section 4.4 state that performance is evaluated on a random sample of 128 test images, but the InceptionResNetV2 row (TP=120, FP=87, TN=83, FN=45) sums to 335, not 128. Using the reported counts, accuracy would be (120+83)/335 = 60.6%, not 90.08%, and the claimed precision 0.9375 with TP=120 would require FP=8, not 87. The ResNet50 and EfficientNetB0 rows also sum to 256 and have TP+FN = 131 and 132, respectively, exceeding the 128-image test set; their precision/recall values are likewise incompatible with the stated counts. Since no raw prediction logs or corrected tables are provided, the F-scores, precision, recall, and accuracy in Table 1 cannot be trusted, and the paper's headline result is unsupported by its own numbers.
  2. [Data Availability Statement] There is a direct contradiction between the experimental description and the Data Availability Statement. Sections 3.1.4, 4.1-4.4 describe training on Kaggle and Nvidia GAN datasets, Dlib face detection, and evaluation on 128 test images, yet the Data Availability Statement asserts 'no datasets were generated or analysed during the current study.' This contradiction means the provenance of Table 1 and Figures 9-12 is unclear, and it prevents any independent verification of the experiments. Please provide the datasets, code, and corrected experimental logs, or explicitly retract the quantitative claims.
  3. [Section 4.3, Figures 11-12] The paper claims the InceptionResNetV2 model achieves the best accuracy on both Nvidia and Kaggle datasets, but no Nvidia-specific results are reported. Both Figure 11 and Figure 12 have the identical caption 'Accuracy obtained using InceptionResNetV2 in Kaggle dataset', and Table 1 presents only Kaggle results. Because the two datasets are said to come from different GAN generators, the evaluation on Nvidia must be reported separately (with sample sizes and confusion matrices) before the claim 'best accuracy on both Nvidia and Kaggle dataset' is supportable. Without a cross-generator or external benchmark, the general claim that the model detects GAN-generated images and videos in the wild is not established.
  4. [Section 3.1.4 and 4.4] The experimental protocol is under-specified. The manuscript does not report dataset sizes (number of videos/images, number of real versus fake crops), class balance, train/validation/test split ratios, hyperparameters, optimizer, learning rate, number of epochs, or how the 128-image test sample was drawn. It also does not provide code or trained weights. These omissions would prevent replication even if Table 1 were arithmetically consistent; combined with the Data Availability contradiction, the experimental section does not provide an auditable evaluation.
minor comments (4)
  1. [Section 3.1.3.3.3] The output-layer description is ambiguous: the text says the final level follows a SoftMax activation, but also describes a single dense neuron output as a sigmoid probability. For binary classification these are different output heads; the authors should state which one was used and how the decision threshold was set.
  2. [References] Several references are incomplete or inconsistent: reference [4] lacks a page range and year, reference [16] ends with '201.' instead of a year, reference [22] is marked '(to appear)', and references [28]-[35] are not cited in the body and appear unrelated to image forensics. The bibliography should be cleaned and limited to works actually used.
  3. [Figures 7-13] The figures reporting accuracy and data analysis lack axis labels, units, and value ranges; for example, Figures 11 and 12 have identical captions even though the text says they show different datasets. Please add clear captions and readable plots with numeric values.
  4. [Section 3.1.4] There is a typo: 'evaluated thorough sampling' should be 'evaluated through sampling'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the 90.08% accuracy is an empirical measurement, though Table 1 and the Data Availability Statement make it numerically unsupported.

full rationale

The paper contains no derivation chain whose conclusion is identified with its premise; it is an empirical comparison of three pretrained CNNs. InceptionResNetV2, ResNet50, and EfficientNetB0 are fine-tuned on GAN-generated and real face crops from the Kaggle/Nvidia datasets, and the 90.08% figure is presented as a measured accuracy on a random 128-image test sample (Sections 3.1.4 and 4.3). Measuring held-out accuracy on the same distribution used for training is standard evaluation, not circular prediction. There are no self-citations, no imported uniqueness theorems, no ansatz adopted from the authors' prior work, and no renamed known result. The paper does have serious evidentiary problems, but they are not circularity: Table 1's confusion-matrix counts sum to 335 for InceptionResNetV2 (120+87+83+45) and 256 for the other two models, not the stated 128 test images; the reported precision, recall, and accuracy do not follow from the counts (e.g., TP=120 with precision 0.9375 implies FP=8, giving 165 test items); and the Data Availability Statement says 'no datasets were generated or analysed during the current study,' directly contradicting the experimental section. These issues undermine the numerical support for the 90.08% claim but do not reduce the claim to its inputs by definition or by self-citation.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No equations or derivations are presented; the central claim is an empirical accuracy number. The claim rests on standard transfer-learning assumptions and on the representativeness of two unnamed datasets. The paper provides no explicit fitted constants, so free parameters are limited to undisclosed trained weights and hyperparameters used to produce the reported accuracy.

free parameters (2)
  • Trained InceptionResNetV2 weights (fine-tuned on Kaggle and Nvidia face crops) = not disclosed
    The reported 90.08% accuracy is a property of these fitted weights; without sharing them or the exact training procedure, the number cannot be reproduced or verified.
  • Binary decision threshold = 0.5
    Chosen by convention, not tuned; affects the precision and recall trade-off reported in Table 1.
assumptions (3)
  • domain assumption Face crops contain enough GAN-specific artifacts for reliable classification
    The pipeline detects faces with Dlib MMOD, crops them, and discards the rest of the image or video context (Sections 3.1.2 and 3.2.2). If the forensic signal lives outside the face crop, the method fails.
  • domain assumption The Kaggle and Nvidia datasets are representative of GAN-generated counterfeit images and videos in general
    Section 4 introduces these datasets as the only training and test sources; the abstract and conclusion generalize to all counterfeit images and videos without cross-dataset evaluation.
  • domain assumption ImageNet-pretrained InceptionResNetV2 features transfer to the forensic real-versus-GAN binary classification task
    The model is described as pretrained on over one million images and then used with a dense sigmoid head (Section 3.1.3.3.3); transfer efficacy is assumed, not validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Independent Discriminant Network Towards Identification of Counterfeit Images and Videos." pith.science (2026). https://pith.science/paper/IKGJ4TVD

@misc{pith2026250605377,
  author       = {Pith},
  title        = {Pith review of: An Independent Discriminant Network Towards Identification of Counterfeit Images and Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IKGJ4TVD}},
  note         = {Machine review of arXiv:2506.05377}
}
read the original abstract

Rapid spread of false images and videos on online platforms is an emerging problem. Anyone may add, delete, clone or modify people and entities from an image using various editing software which are readily available. This generates false and misleading proof to hide the crime. Now-a-days, these false and counterfeit images and videos are flooding on the internet. These spread false information. Many methods are available in literature for detecting those counterfeit contents but new methods of counterfeiting are also evolving. Generative Adversarial Networks (GAN) are observed to be one effective method as it modifies the context and definition of images producing plausible results via image-to-image translation. This work uses an independent discriminant network that can identify GAN generated image or video. A discriminant network has been created using a convolutional neural network based on InceptionResNetV2. The article also proposes a platform where users can detect forged images and videos. This proposed work has the potential to help the forensics domain to detect counterfeit videos and hidden criminal evidence towards the identification of criminal activities.

Figures

Figures reproduced from arXiv: 2506.05377 by the authors.

Figure 5
Figure 5. Workflow of the Backend System The main modules and functionalities of the application are; [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 7
Figure 7. Analysis of the Metadata [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Analysis of the distribution of the data [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figures from the paper (6 more)
Figure 6
Figure 6. Figure 6: Workflow of the Backend System The backend for this application is deployed on Google Cloud Platform using their Platform as a Service (PaaS) Cloud Run. This service is chosen because of low-cost option which runs the backend only when a request is assessed. The backen…
Figure 12
Figure 12. Figure 12: Accuracy obtained using InceptionResNetV2 in Kaggle dataset It is observed in [PITH_FULL_IMAGE:figures/full_fig_p006_12.png]
Figure 10
Figure 10. Figure 10: Accuracy obtained using EfficientNetB0 in Kaggle dataset It is observed in [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: Accuracy obtained using InceptionResNetV2 in Kaggle dataset [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]
Figure 13
Figure 13. Figure 13: Comparative analysis of the models 4.5 User Interface Design of the Application [PITH_FULL_IMAGE:figures/full_fig_p007_13.png]
Figure 14
Figure 14. Figure 14: (a) and 14(b) shows the user interface that has been developed using React Native and Expo. The proposed frontend architecture has been designed taking into account user-friendliness and clarity of output. Here are few examples of the interface [PITH_FULL_IMAGE:figur…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

  1. [1]

    It prepares and cleans the raw data to make it useable by the model

    Data visualization and Analysis Data is cleaned, visualized and analysed in this phase. It prepares and cleans the raw data to make it useable by the model. It performs face detection in video and crop out the required section as training images. MMOD (Max Margin Object Detection) is used for this step. Parameters for facial detection are, a. 2x scale for...

    work page 1900

  2. [10]

    Image forgery detection through residual-based local descriptors and block-matching,

    D. Cozzolino, D. Gragnaniello, and L.Verdoliva, “Image forgery detection through residual-based local descriptors and block-matching,” in IEEE Conference on Image Processing(ICIP), October 2014

    work page 2014

  3. [17]

    Peng, C., Liu, Y., Yuan, X. et al. Research of image recognition method based on enhanced inception-ResNet-V2. Multimed Tools Appl 81, 34345–34365 (2022)

    work page 2022

  4. [18]

    Detecting GAN-Generated Imagery Using Saturation Cues,

    S. McCloskey and M. Albright, "Detecting GAN-Generated Imagery Using Saturation Cues," 2019 IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, 2019, pp. 4584-4588, doi: 10.1109/ICIP.2019.8803661

    work page arXiv 2019

  5. [20]

    Detection of gan-generated counterfeit images over social networks

    F. Marra, D. Gragnaniello, D. Cozzolino and L. Verdoliva, "Detection of gan-generated counterfeit images over social networks", 2018 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR),

    work page 2018

  6. [23]

    Can Forensic Detectors Identify GAN Generated Images?,

    H. Li, H. Chen, B. Li and S. Tan, "Can Forensic Detectors Identify GAN Generated Images?," 2018 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Honolulu, HI, USA, 2018, pp. 722-727, doi: 10.23919/APSIPA.2018.8659461

    work page arXiv 2018

  7. [24]

    GAN-Generated Image Detection With Self-Attention Mechanism Against GAN Generator Defect,

    Z. Mi, X. Jiang, T. Sun and K. Xu, "GAN-Generated Image Detection With Self-Attention Mechanism Against GAN Generator Defect," in IEEE Journal of Selected Topics in Signal Processing, vol. 14, no. 5, pp. 969-981, Aug. 2020, doi: 10.1109/JSTSP.2020.2994523

    work page arXiv 2020

  8. [25]

    Incremental learning for the detection and classification of GAN-generated images,

    F. Marra, C. Saltori, G. Boato and L. Verdoliva, "Incremental learning for the detection and classification of GAN-generated images," 2019 IEEE International Workshop on Information Forensics and Security (WIFS), Delft, Netherlands, 2019, pp. 1-6, doi: 10.1109/WIFS47025.2019.9035099

    work page arXiv 2019

Show all 18 references
  1. [26]

    On the use of Benford's law to discriminate GAN-generated images,

    N. Bonettini, P. Bestagini, S. Milani and S. Tubaro, "On the use of Benford's law to discriminate GAN-generated images," 2020 25th International Conference on Pattern Recognition (ICPR), Milan, Italy, 2021, pp. 5495-5502, doi: 10.1109/ICPR48806.2021.9412944

    2020

  2. [27]

    Detecting Generated Image Based on a Coupled Network with Two-Step Pairwise Learning,

    Y. -X. Zhuang and C. -C. Hsu, "Detecting Generated Image Based on a Coupled Network with Two-Step Pairwise Learning," 2019 IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, 2019, pp. 3212-3216, doi: 10.1109/ICIP.2019.8803464

    2019

  3. [28]

    Precision parameter estimation in Proton Exchange Membrane Fuel Cells using depth information enhanced Differential Evolution

    P. Jangir, A.E. Ezugwu, Arpita et al. “Precision parameter estimation in Proton Exchange Membrane Fuel Cells using depth information enhanced Differential Evolution”, Sci Rep 14, 29591 (2024)

    2024

  4. [29]

    A hybrid mutational Northern Goshawk and elite opposition learning artificial rabbits optimizer for PEMFC parameter estimation

    P. Jangir, A.E. Ezugwu, K. Saleem et al. “A hybrid mutational Northern Goshawk and elite opposition learning artificial rabbits optimizer for PEMFC parameter estimation”. Sci Rep 14, 28657 (2024)

    2024

  5. [30]

    A levy chaotic horizontal vertical crossover based artificial hummingbird algorithm for precise PEMFC parameter estimation

    P. Jangir, A.E. Ezugwu, K. Saleem et al. “A levy chaotic horizontal vertical crossover based artificial hummingbird algorithm for precise PEMFC parameter estimation”. Sci Rep 14, 29597 (2024)

    2024

  6. [31]

    A cooperative strategy-based differential evolution algorithm for robust PEM fuel cell parameter estimation

    P. Jangir, Arpita, S.K. Agrawal et al., “A cooperative strategy-based differential evolution algorithm for robust PEM fuel cell parameter estimation”. Ionics. 31, 2024, 703-741. 10.1007/s11581-024-05963-x

    2024 doi

  7. [32]

    Compressive behavior of elliptical concrete-filled steel tubular short columns using numerical investigation and machine learning techniques

    H.S. Mohamed, T., Qiong, H.F. Isleem et al. “Compressive behavior of elliptical concrete-filled steel tubular short columns using numerical investigation and machine learning techniques” Sci Rep 14, 27007 (2024)

    2024

  8. [33]

    Utilization finite element and machine learning methods to investigation the axial compressive behavior of elliptical FRP-confined concrete columns

    Y. Chen, F. Haytham, D. N. Isleem et al., "Utilization finite element and machine learning methods to investigation the axial compressive behavior of elliptical FRP-confined concrete columns", Structures, Volume 70, 2024, 107681, ISSN 2352-0124,

    2024

  9. [34]

    Food Security Based Marine Life Ecosystem for Polar Region Conditioning: Remote Sensing Analysis with Machine Learning Model

    G. Srikanth, D. Nimma, RVS Lalitha et al. “Food Security Based Marine Life Ecosystem for Polar Region Conditioning: Remote Sensing Analysis with Machine Learning Model”, Remote Sens Earth Syst Sci (2024)

    2024

  10. [35]

    The quick crisscross sine cosine algorithm for optimal FACTS placement in uncertain wind integrated scenario based power systems

    Sunilkumar, P. Agrawal, P. Jangir, et al., "The quick crisscross sine cosine algorithm for optimal FACTS placement in uncertain wind integrated scenario based power systems", Results in Engineering, Volume 25, 2025, 103703, ISSN 2590-1230 Shayantani Kar has pursued her B.Tech....

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.