Pith. sign in

REVIEW 3 major objections 6 minor 17 references

Detecting Facial Image Manipulations with Multi-Layer CNN Models

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a six-layer convolutional network, MesoMultiNet+, distinguishes genuine faces from deepfake and face-swap manipulations with 76% accuracy on a balanced dataset, and that the improvement comes from deeper convolution…

desk verdict Routine MesoNet extension whose headline 76% accuracy is undermined by a post-hoc dataset swap and no documented held-out split. read the letter →

arxiv 2412.06643 v1 pith:UEHWOMU2 submitted 2024-12-09 cs.CV cs.AI

Alejandro Marco Montejano , Angela Sanchez Perez , Javier Barrachina , David Ortiz-Perez , Manuel Benavent-Lledo , Jose Garcia-Rodriguez This is my paper · ORCID
classification cs.CVcs.AI
keywords deepfakedetectionfacialimagemanipulationconvolutionalneuralnetworkMesoNetmulti-classclassificationstablediffusionimagesfaceswaptransferlearning
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

The paper sets out to build lightweight convolutional neural networks that can tell genuine facial images apart from manipulated ones, extending the compact MesoNet architecture to handle multiple attack types at low computational cost. It reports that a six-layer multi-class model, MesoMultiNet+, reaches 76% accuracy on a balanced dataset separating real faces from deepfake and face-swap images, compared with 43% for the four-layer version and 68% for the six-layer version trained without AI-generated images. The authors argue that adding convolutional depth, variable filter sizes, facial-landmark alignment, and transfer learning from a pretrained binary model are what make this performance possible. A sympathetic reader would take the central claim to be that a modestly deeper CNN, trained with synthetic text-to-image outputs in the mix, can serve as an effective and resource-efficient detector of facial manipulation.

What carries the argument

The load-bearing object is MesoMultiNet+, a six-layer convolutional architecture inspired by MesoNet. It uses alternating convolutional, batch-normalization, and max-pooling blocks with increasing filter counts and mixed kernel sizes, followed by dense layers with dropout and a softmax output. Training relies on categorical cross-entropy with the Adam optimizer, transfer learning from a pretrained binary MesoNet model, and a preprocessing step that aligns faces using a 68-point facial landmark detector before classification.

What would settle it

Run the trained MesoMultiNet+ model on an unseen deepfake dataset, such as the original DeepFake images from the MesoNet dataset or an independent benchmark, without any retraining; if accuracy falls toward chance levels (50% in binary or 33% in three-class settings), the reported 76% reflects dataset-specific artifacts rather than genuine generalization.

Watch

Extended reading notes

Core claim

The central discovery claimed in the paper is that extending the MesoNet architecture from four to six convolutional layers, together with replacing the deepfake training class with Stable Diffusion images, lifts multi-class facial manipulation detection from 43% to 76% accuracy on a balanced three-way dataset. In the final configuration, MesoMultiNet+ achieves per-class F1 scores of 0.78 for deepfake and face-swap and 0.74 for genuine images, with precision and recall balanced across all three classes. The authors also report that the binary MesoNet+ model reaches 90% accuracy on a balanced dataset of deepfakes and genuine images, and they attribute this improvement to the added layers and to careful preprocessing. The paper frames this as surpassing traditional approaches, where the comparison baseline is the original MesoNet binary model and its four-layer multi-class adaptation.

Load-bearing premise

The results are computed on the same balanced dataset used for training, where the deepfake class was replaced with Stable Diffusion images; the assumption is that accuracy on this dataset reflects the model's ability to detect manipulations in the wild.

Editorial extensions

If this is right

  • A compact six-layer CNN can separate genuine faces from deepfake and face-swap images at 76% accuracy on a balanced set, making multi-class manipulation screening feasible on low-resource devices.
  • The jump from 43% accuracy with four layers to 68% with six layers suggests that convolutional depth adds useful discriminative power for fine-grained manipulation artifacts.
  • Including Stable Diffusion images in the training set raised accuracy from 68% to 76%, indicating that exposure to text-to-image synthetic outputs sharpens the model's ability to separate manipulated from genuine faces.
  • Transfer learning from a pretrained binary model to the multi-class model accelerates training and improves feature reuse, according to the paper's reported results.

Reading between the lines

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

  • The headline 76% figure does not measure detection of classic deepfakes: the deepfake class in that experiment consists of Stable Diffusion images, so the number reflects separation among genuine, face-swap, and diffusion-generated faces rather than detection of the original DeepFake artifacts.
  • A natural next experiment, not reported in the paper, would be to evaluate the same six-layer architecture on an independent deepfake benchmark without retraining; that would show whether the accuracy transfers beyond the specific balanced dataset.
  • The claim of surpassing traditional approaches is supported only against the paper's own MesoNet baselines; a comparison with other post-MesoNet detection methods is absent, so the relative standing of the model in the broader field is not yet established.
  • The binary 90% accuracy on the balanced dataset is also dataset-specific; real-world value would depend on the prevalence of manipulations and the variety of unseen attack types encountered in deployment.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper proposes lightweight CNN architectures for detecting manipulated facial images, extending the MesoNet approach. The authors introduce a binary model (MesoNet+) and two multi-class models (MesoMultiNet and MesoMultiNet+) with four and six convolutional layers, respectively. They evaluate the binary model on the original MesoNet dataset and Celeb-DF, and the multi-class models on a balanced dataset of bonafide, DeepFake (later replaced by Stable Diffusion images from DiffusionDB), and FaceSwap images. The headline result is 76% accuracy for the six-layer model on the modified dataset, compared to 43% and 68% for the two earlier multi-class variants. The paper also documents the original MesoNet's generalization failure on Celeb-DF (AUC drop from 0.96 to 0.56).

Significance. If the 76% accuracy were established under a valid evaluation protocol, the proposed models would offer a practically useful lightweight solution for distinguishing real faces from certain AI-generated manipulations and face swaps, with potential deployment in resource-constrained settings. The paper's strength lies in its transparent reporting of negative results: the drop in MesoNet's AUC on Celeb-DF and the initial 43% accuracy of the four-layer multi-class model are honestly presented. The explicit disclosure of the dataset modification (replacing DeepFake images with Stable Diffusion images) is also a point in the authors' favor. However, the central quantitative claim is not currently supported: the final experiment changes the classification task, no held-out test set is documented, and no external baseline is compared, so the reported 76% does not substantiate the abstract's claim of surpassing traditional approaches.

major comments (3)
  1. [Section IV-A, Tables IV-V and Section IV-F, Table VIII] In Section IV-A you state that, because FaceSwap and DeepFake images were difficult to differentiate, the DeepFake images from the previous set were replaced by images generated with Stable Diffusion from the DiffusionDB dataset. Consequently, the 'DeepFake' class in Tables V, VII, VIII, and IX is not a deepfake class at all; it is a Stable Diffusion class. The 76% accuracy reported in Section IV-F and Table VIII is therefore for a three-class problem of bonafide vs. FaceSwap vs. Stable Diffusion images, not for the task claimed in the abstract (distinguishing manipulated images from genuine ones). The 'DeepFake' row in Table VIII is mislabeled, and the improvement from 68% to 76% in Table IX is partly a result of this dataset change rather than of the architectural changes alone. Please either relabel the class and qualify the claims, or rerun the experiment with actual DeepFake images, or provide a clear justification for why Stable Diffusion images are a proxy for deepfakes.
  2. [Section IV-A and Section IV-F] No train/validation/test split is described for the multi-classification experiments. Section IV-A says the balanced dataset is used to train and evaluate the multi-classifier system, and Table V lists only total image counts per class; Section IV-F then reports the 76% accuracy apparently on the same data. With roughly 1,500 images per class and a six-layer CNN, the reported metrics may reflect in-sample performance rather than generalization to unseen images. Please specify the exact data partitioning (e.g., a fixed train/validation/test split or cross-validation), state which partition produced the numbers in Tables VI-VIII, and verify that no test images were used for model selection or early stopping. Without this, the 76% figure cannot be interpreted as an out-of-sample result.
  3. [Section IV-G and abstract] The abstract claims the proposed models achieve accuracy surpassing traditional approaches, but Table IX compares only the three proposed architectures (MesoMultiNet, MesoMultiNet+, and MesoMultiNet+ with Stable Diffusion). No external baseline is evaluated under the same protocol: the original MesoNet (or MesoInception-4) is not retrained or evaluated on this dataset, and no published state-of-the-art forensics method is included. The progression 43% to 68% to 76% also conflates architectural changes with dataset changes, since the dataset composition changes between Table IV and Table V. A fair comparative evaluation against at least the original MesoNet on the identical data split is needed to support the superiority claim.
minor comments (6)
  1. [Section IV-B] The ROC curve is defined with the false negative rate FPR; this should be the false positive rate.
  2. [Section IV-D, final paragraph] The phrase 'can be seen in Table I' likely refers to Table VI (or VII) for multi-class metrics; Table I is the binary dataset split.
  3. [Throughout] The model is named inconsistently: MesoMultiNet and MultiMesoNet are used interchangeably (e.g., Section IV-G vs. Section III-B); please standardize.
  4. [Section II-D] Reference [11] is cited for MesoNet architecture details, but [11] is a ControlFace paper; the correct reference for MesoNet is [6]. The reference numbering appears misaligned.
  5. [Section IV-B] The text says 'The original accuracy, which stood at 0.96, decreases drastically to 0.56' and then refers to an AUC-ROC value of 0.56; these are two different metrics and should not be conflated.
  6. [Various] There are several typos: 'DifussionDB' (Section IV-A), 'deconvolution layers' (Sections IV-B/IV-C; the added layers are convolutional), and 'We would like to thanks' (Acknowledgments).

Circularity Check

2 steps flagged · score 6.0 of 10

The 76% result is an in-sample accuracy on a dataset whose difficult DeepFake class was swapped for Stable Diffusion, so the headline claim reduces to the chosen training input.

  1. fitted input called prediction [Section IV-A (dataset replacement) and Section IV-F/Table VIII]
    "When comparing FaceSwap images with DeepFake images, we found that these two classes are complex to differentiate, so we made a change to the dataset. ... The DeepFake images from the previous set are replaced by images generated with Stable Diffusion from the DifussionDB dataset [17]."

    The reported 76% accuracy is obtained only after the hard DeepFake class was replaced by Stable Diffusion images (Table V), and Table VIII labels those Stable Diffusion images 'DeepFake.' The jump from 68% (Table VII) to 76% is thus driven by changing the input class distribution, not by a model that better detects DeepFake manipulations. The claimed 'DeepFake' detection result is a renamed evaluation of an easier, substituted class.

  2. fitted input called prediction [Section IV-A and Section IV-F, Tables IV-V and Table VIII]
    "The database used to train and evaluate the multi-classifier system has been balanced, ensuring an equal distribution of samples between the different classes, this can be seen in Table IV. ... In the latter proposed multi-classification architecture, results of 76% accuracy are obtained."

    No train/validation/test split is documented for the multi-class system; only total class counts are given for the set 'used to train and evaluate.' The 76% figure is therefore measured on the same distribution that trained the model, i.e., in-sample accuracy. Without a held-out partition, the number is a restatement of training-set fit rather than an out-of-sample prediction of manipulation detection.

full rationale

The paper's central quantitative claim—76% accuracy 'distinguishing manipulated images from genuine ones, surpassing traditional approaches'—is not a derived prediction but an in-sample measurement on a dataset that was explicitly modified after the authors found the original DeepFake class hard to distinguish. Section IV-A states the balanced set is 'used to train and evaluate' and gives no held-out split; Section IV-F then reports 76% on that same set. The replacement of DeepFake images with Stable Diffusion images (Table V) and the relabeling of that class as 'DeepFake' in Table VIII means the 76% is partly a consequence of dataset selection. There is no self-citation chain or imported uniqueness theorem; the circularity is in the evaluation protocol: the reported accuracy reduces to training-set fit and to a swapped, easier input class. External benchmarks or a documented held-out split could have made the claim non-circular, but none is provided.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unvalidated choices: hand-picked architecture hyperparameters, a pretrained landmark detector for alignment, a balanced dataset of roughly 1,500 images per class, and a mid-study replacement of the deepfake class with Stable Diffusion images. Because these choices are not ablated or independently verified, they carry much of the explanatory burden for the reported accuracy.

free parameters (6)
  • Number of convolutional layers = 4 (MesoMultiNet) and 6 (MesoMultiNet+)
    Chosen by hand to improve feature extraction; no systematic search or ablation is reported.
  • Number of filters per layer = 8, 16, 32 (approximate)
    Filter counts follow the original MesoNet and are increased in the deeper variant; no analysis of alternatives is provided.
  • Convolutional kernel sizes = 3x3 and 5x5
    Selected based on prior recommendations; no experimental comparison of other kernel sizes is reported.
  • Learning rate (Adam) = Not specified
    The paper states Adam with 'a specified learning rate' but never reports the value, making exact reproduction impossible.
  • Epochs and batch size = Not specified
    Training duration and batch size are not reported, so the training procedure is underspecified.
  • Dataset composition per class = Approximately 1,450 to 1,550 images per class
    The dataset is balanced by hand, and the DeepFake class is later replaced with Stable Diffusion images; this choice drives the final 76% accuracy result.
assumptions (4)
  • domain assumption MesoNet's pretrained weights provide a useful starting point for multi-class manipulation detection.
    The method transfers weights from binary MesoNet to multi-class models without ablation, assuming they capture manipulation-relevant features.
  • domain assumption Facial landmark alignment with a 68-point detector improves detection.
    Alignment is applied to all images, but no experiment compares aligned versus unaligned input; the benefit is assumed from prior practice.
  • ad hoc to paper Replacing the DeepFake class with Stable Diffusion images is a valid way to evaluate generalization to AI-generated manipulations.
    This dataset change was made after observing that DeepFake and FaceSwap were hard to distinguish, and the improved accuracy is measured on the modified distribution, not on an independent benchmark.
  • domain assumption A balanced dataset of about 1,500 images per class is sufficient for training and evaluating a generalizable multi-class detector.
    No cross-validation, identity-disjoint splits, or external test set are provided; the paper states limited dataset size as a limitation but proceeds without addressing it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detecting Facial Image Manipulations with Multi-Layer CNN Models." pith.science (2026). https://pith.science/paper/UEHWOMU2

@misc{pith2026241206643,
  author       = {Pith},
  title        = {Pith review of: Detecting Facial Image Manipulations with Multi-Layer CNN Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UEHWOMU2}},
  note         = {Machine review of arXiv:2412.06643}
}
read the original abstract

The rapid evolution of digital image manipulation techniques poses significant challenges for content verification, with models such as stable diffusion and mid-journey producing highly realistic, yet synthetic, images that can deceive human perception. This research develops and evaluates convolutional neural networks (CNNs) specifically tailored for the detection of these manipulated images. The study implements a comparative analysis of three progressively complex CNN architectures, assessing their ability to classify and localize manipulations across various facial image modifications. Regularization and optimization techniques were systematically incorporated to improve feature extraction and performance. The results indicate that the proposed models achieve an accuracy of up to 76\% in distinguishing manipulated images from genuine ones, surpassing traditional approaches. This research not only highlights the potential of CNNs in enhancing the robustness of digital media verification tools, but also provides insights into effective architectural adaptations and training strategies for low-computation environments. Future work will build on these findings by extending the architectures to handle more diverse manipulation techniques and integrating multi-modal data for improved detection capabilities.

Figures

Figures reproduced from arXiv: 2412.06643 by the authors.

Figure 1
Figure 1. Comparison of original and aligned images for Deep [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. True positive rate and false negative rate in the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. True positive rate and false negative rate in the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (3 more)
Figure 7
Figure 7. Figure 7: MesoMultiNet architecture confusion matrix. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: MesoMultiNet+ architecture confusion matrix. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: MesoMultiNet+ architecture confusion matrix. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 8 canonical work pages

  1. [1]

    On the detection of digital face manipulation,

    H. Dang, F. Liu, J. Stehouwer, X. Liu, and A. K. Jain, “On the detection of digital face manipulation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 5781–5790

    work page 2020

  2. [2]

    Fused classification for differential face morphing detection,

    I. Medvedev, J. A. Pimenta, and N. Gonc ¸alves, “Fused classification for differential face morphing detection,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 1043– 1050

    work page 2024

  3. [3]

    Research on splicing image detection algorithms based on natural image statistical characteristics,

    A. Xiang, J. Zhang, Q. Yang, L. Wang, and Y . Cheng, “Research on splicing image detection algorithms based on natural image statistical characteristics,” arXiv preprint arXiv:2404.16296 , 2024

    arXiv 2024

  4. [4]

    Faceforensics++: Learning to detect manipulated facial images,

    A. Rossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Niess- ner, “Faceforensics++: Learning to detect manipulated facial images,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019

    work page 2019

  5. [5]

    Singh, G

    G. Singh, G. Bhardwaj, S. V . Singh, and V . Garg, Biometric Iden- tification System: Security and Privacy Concern . Cham: Springer International Publishing, 2021, pp. 245–264

    work page 2021

  6. [6]

    Mesonet: a compact facial video forgery detection network,

    D. Afchar, V . Nozick, J. Yamagishi, and I. Echizen, “Mesonet: a compact facial video forgery detection network,” in 2018 IEEE International Workshop on Information Forensics and Security (WIFS), 2018, pp. 1–7

    2018

  7. [7]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 6840–6851

    2020

  8. [8]

    How generative adversarial networks and their variants work: An overview,

    Y . Hong, U. Hwang, J. Yoo, and S. Yoon, “How generative adversarial networks and their variants work: An overview,” ACM Computing Surveys (CSUR), vol. 52, no. 1, pp. 1–43, 2019

    2019

Show all 17 references
  1. [9]

    Diffface: Diffusion-based face swapping with facial guidance,

    K. Kim et al. , “Diffface: Diffusion-based face swapping with facial guidance,” arXiv preprint arXiv:2212.13344 , 2022

    2022 arXiv

  2. [10]

    Detecting and grounding multi-modal media manipulation,

    R. Shao, T. Wu, and Z. Liu, “Detecting and grounding multi-modal media manipulation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 6904–6913

    2023

  3. [11]

    Controlface: Feature disentangling for controllable face swapping,

    X. Zhang, W. Zhou, K. Liu, H. Tang, Z. Zhang, W. Zhang, and N. Yu, “Controlface: Feature disentangling for controllable face swapping,” Journal of Imaging , vol. 10, no. 1, p. 0, Jan 2024, pMID: 38249006; PMCID: PMC11154356

    2024

  4. [12]

    Learning structure and strength of cnn filters for small sample size training,

    R. Keshari, M. Vatsa, R. Singh, and A. Noore, “Learning structure and strength of cnn filters for small sample size training,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 9349–9358

    2018

  5. [13]

    Using cnn for facial expression recognition: a study of the effects of kernel size and number of filters on accuracy,

    A. Agrawal and N. Mittal, “Using cnn for facial expression recognition: a study of the effects of kernel size and number of filters on accuracy,” The Visual Computer , vol. 36, no. 2, pp. 405–412, 2020

    2020

  6. [14]

    Shape predictor 68-point facial landmark detector,

    D. King, “Shape predictor 68-point facial landmark detector,” 2009

    2009

  7. [15]

    Celeb-df: A large- scale challenging dataset for deepfake forensics,

    Y . Li, X. Yang, P. Sun, H. Qi, and S. Lyu, “Celeb-df: A large- scale challenging dataset for deepfake forensics,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 3207–3216

    2020

  8. [16]

    The deepfake detection challenge (dfdc) dataset,

    B. Dolhansky, J. Bitton, B. Pflaum, J. Lu, R. Howes, M. Wang, and C. C. Ferrer, “The deepfake detection challenge (dfdc) dataset,” arXiv preprint arXiv:2006.07397, 2020

    2006 arXiv

  9. [17]

    Diffusiondb: A large-scale prompt gallery dataset for text-to- image generative models,

    Z. J. Wang, E. Montoya, D. Munechika, H. Yang, B. Hoover, and D. H. Chau, “Diffusiondb: A large-scale prompt gallery dataset for text-to- image generative models,” arXiv preprint arXiv:2210.14896 , 2022

    2022 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.