arxiv: 2605.08167 · v1 · submitted 2026-05-04 · 💻 cs.CV · cs.AI

Recognition: no theorem link

Digital Image Forgery Detection Using Transfer Learning

Fatma Betul Buyuk , Gozde Karatas Baydogmus , Ali Buldu , Ayaulym Tulendiyeva , Zhuldyz Baizhumanova

Authors on Pith no claims yet

Pith reviewed 2026-05-12 00:57 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords digital image forgery detectiontransfer learningconvolutional neural networkscompression featuresCASIA v2.0image forensicsYouden Index

0 comments

The pith

Transfer learning with compression difference features improves detection of image forgeries on the CASIA v2.0 dataset.

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

The paper introduces a transfer learning framework that feeds pretrained CNNs a hybrid input of ordinary RGB images plus compression difference maps to make tampering traces easier to spot. It applies this setup across six common architectures and uses an automatic threshold based on the Youden Index to balance correct detections against mistakes. Experiments on the CASIA v2.0 collection of authentic and edited images show DenseNet121 reaching top accuracy and AUC while ResNet50 gives the most stable overall scores. A reader would care because photo editing tools are now widespread and reliable automated checks matter for verifying visual evidence in news, legal, and security contexts.

Core claim

The integration of RGB images with compression difference-based features (FDIFF) as input to multiple pretrained CNN architectures, combined with Youden Index adaptive thresholding, improves the visibility of manipulation artifacts and yields more robust classification performance than standard approaches on the CASIA v2.0 dataset.

What carries the argument

The hybrid input representation that combines RGB images with compression difference-based features (FDIFF) to explicitly highlight subtle manipulation artifacts.

If this is right

DenseNet121 records the highest accuracy and area under the ROC curve among the tested models.
ResNet50 produces the most balanced predictions with the highest Matthews correlation coefficient.
Forensic applications require metrics beyond accuracy because minimizing false negatives is critical.
The overall design supports use in real-world digital image forgery detection scenarios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same hybrid feature approach could be tested on other public forgery datasets to check whether the performance gains generalize beyond CASIA v2.0.
The Youden Index thresholding step might transfer to other binary classification problems where false positive and false negative costs differ.
Extending the input representation to include additional low-level signals such as noise residuals could further strengthen detection of sophisticated edits.

Load-bearing premise

The forgeries present in the CASIA v2.0 dataset are representative of the manipulation artifacts found in real-world images taken under uncontrolled conditions.

What would settle it

Applying the same framework to a new collection of images edited with different software or under varied lighting and compression settings and checking whether the reported gains in artifact visibility and balanced classification metrics disappear.

Figures

Figures reproduced from arXiv: 2605.08167 by Ali Buldu, Ayaulym Tulendiyeva, Fatma Betul Buyuk, Gozde Karatas Baydogmus, Zhuldyz Baizhumanova.

**Figure 2.** Figure 2: Illustration of transfer learning, where knowledge learned from large [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Different transfer learning scenarios based on dataset size and simi [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Proposed transfer learning-based classification architecture incorpo [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: ROC curves illustrating the classification performance of each evalu [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: Confusion matrices for each evaluated CNN model. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

read the original abstract

The increasing availability of advanced image editing tools has led to a significant rise in manipulated digital content, posing serious challenges for digital forensics and information security. This study presents a transfer learning-based framework for digital image forgery detection that integrates compression-aware feature enhancement with deep convolutional neural network (CNN) architectures. The proposed approach introduces a hybrid input representation that combines RGB images with compression difference-based features (FDIFF), explicitly highlighting subtle manipulation artifacts that are often difficult to detect. In addition, a model-specific adaptive threshold optimization strategy based on the Youden Index is employed to improve classification reliability by achieving a better balance between true positive and false positive rates. Experiments conducted on the CASIA v2.0 dataset using multiple pretrained CNN architectures, including DenseNet121, VGG16, ResNet50, EfficientNetB0, MobileNet, and InceptionV3, demonstrate the effectiveness and robustness of the proposed framework. The models are evaluated using comprehensive performance metrics such as accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), and area under the ROC curve (AUC). The results show that DenseNet121 achieves the highest accuracy and AUC, while ResNet50 provides the most balanced and reliable predictions with the highest MCC. The findings emphasize that relying solely on accuracy is insufficient for forensic applications, where minimizing false negatives is critical. Overall, the proposed framework improves the visibility of manipulation artifacts and enhances classification robustness, making it suitable for real-world digital image forgery detection scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Routine transfer learning on CASIA v2.0 with FDIFF features and Youden-tuned thresholds, but the thresholds are fit on the test set itself.

read the letter

The paper takes several standard pretrained CNNs, feeds them a hybrid RGB plus compression-difference input, and picks per-model decision thresholds by maximizing the Youden index. That combination is the only real engineering step beyond what has already appeared in the forgery-detection literature. They run the setup on DenseNet121, ResNet50, VGG16 and the rest, then report accuracy, AUC, MCC and the usual metrics. ResNet50 comes out looking balanced on MCC, which is a reasonable point to make for forensic work where false negatives cost more than accuracy suggests. The experiments are at least consistent across six architectures on a public dataset, so the numbers themselves are not obviously fabricated. Beyond that, nothing new in architecture or theory is claimed or shown. The soft spot is the threshold step. Optimizing the Youden index on the same test split used for the final scores is post-hoc fitting; the reported performance is therefore optimistic and would drop on fresh data. The abstract and results give no training schedule, no cross-validation details, and no ablation that isolates whether the FDIFF channel actually moves the needle. The claim that the framework is suitable for real-world scenarios rests entirely on CASIA v2.0, whose splicing and copy-move forgeries plus fixed JPEG pipeline do not cover GAN content, social-media recompression chains, or unseen sensors. A reader who needs a quick baseline pipeline for this narrow task can pull the reported numbers and the hybrid input idea. Anyone expecting generalizable forensic performance or reproducible training details will be disappointed. The work is coherent on its own terms and shows honest empirical effort, so it clears the bar for peer review, but the referee should require the threshold procedure to be moved to a validation set and some form of cross-dataset check before the real-world language is kept.

Referee Report

4 major / 2 minor

Summary. The paper proposes a transfer learning framework for digital image forgery detection that combines RGB images with compression-difference features (FDIFF) as hybrid input to pretrained CNNs (DenseNet121, VGG16, ResNet50, EfficientNetB0, MobileNet, InceptionV3). Model-specific decision thresholds are set by maximizing the Youden Index, and the approach is evaluated on the CASIA v2.0 dataset using accuracy, precision, recall, F1-score, MCC, and AUC. The authors report that DenseNet121 yields the highest accuracy/AUC while ResNet50 is most balanced by MCC, and conclude that the framework improves artifact visibility and is suitable for real-world forgery detection.

Significance. If the central performance claims can be substantiated without post-hoc fitting and with broader validation, the hybrid FDIFF representation could offer a lightweight way to emphasize compression artifacts in forensic pipelines. The evaluation across six architectures and the emphasis on MCC (rather than accuracy alone) are positive aspects. However, the current results are confined to one dataset with limited forgery types, so the claimed real-world suitability remains unsubstantiated.

major comments (4)

[Methods / Evaluation] The Youden-Index threshold optimization is performed on the same test data used to report final accuracy, AUC, and MCC scores (as described in the adaptive threshold optimization strategy and evaluation on CASIA v2.0). This introduces circularity that directly inflates the reported classification reliability and undermines the claim of improved robustness.
[Experiments] No ablation is presented that compares the hybrid RGB+FDIFF input against standard RGB-only inputs on the same models and splits. Without this, it is impossible to attribute the reported gains to the FDIFF component rather than to the choice of backbone or threshold tuning.
[Conclusion] All quantitative results (including the highest accuracy/AUC for DenseNet121 and highest MCC for ResNet50) come exclusively from CASIA v2.0. The conclusion that the framework is 'suitable for real-world digital image forgery detection scenarios' therefore rests on an untested assumption that CASIA forgeries (splicing/copy-move with specific post-processing) represent the artifact distributions encountered under uncontrolled conditions.
[Experiments] The manuscript provides no details on training protocol (learning rate, optimizer, number of epochs, data augmentation, or train/validation/test split ratios) or cross-validation scheme. This absence makes the performance numbers non-reproducible and prevents assessment of whether the results are stable across different partitions.

minor comments (2)

[Abstract] The abstract states that the models are 'evaluated using comprehensive performance metrics' but does not mention any training hyperparameters or validation procedure; adding a brief sentence on these would improve clarity.
[Related Work] The paper would benefit from explicit comparison to prior FDIFF or compression-aware forgery detectors to clarify the incremental contribution of the hybrid input.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which highlight important methodological and reporting issues. We will revise the manuscript accordingly to improve reproducibility, remove bias in evaluation, provide necessary ablations, and moderate our claims.

read point-by-point responses

Referee: [Methods / Evaluation] The Youden-Index threshold optimization is performed on the same test data used to report final accuracy, AUC, and MCC scores (as described in the adaptive threshold optimization strategy and evaluation on CASIA v2.0). This introduces circularity that directly inflates the reported classification reliability and undermines the claim of improved robustness.

Authors: We agree that performing Youden Index optimization on the test set creates circularity and potential over-optimism. In the revised version, we will split the data into train/validation/test sets, optimize the model-specific thresholds exclusively on the validation set, and report all final metrics on the held-out test set using those fixed thresholds. revision: yes
Referee: [Experiments] No ablation is presented that compares the hybrid RGB+FDIFF input against standard RGB-only inputs on the same models and splits. Without this, it is impossible to attribute the reported gains to the FDIFF component rather than to the choice of backbone or threshold tuning.

Authors: We acknowledge that the absence of this ablation prevents clear attribution of gains to the FDIFF representation. We will add a direct comparison of hybrid RGB+FDIFF versus RGB-only inputs for all six backbones under identical training and evaluation conditions, including the same data splits and threshold strategy. revision: yes
Referee: [Conclusion] All quantitative results (including the highest accuracy/AUC for DenseNet121 and highest MCC for ResNet50) come exclusively from CASIA v2.0. The conclusion that the framework is 'suitable for real-world digital image forgery detection scenarios' therefore rests on an untested assumption that CASIA forgeries (splicing/copy-move with specific post-processing) represent the artifact distributions encountered under uncontrolled conditions.

Authors: We accept that the real-world suitability claim is overstated given the single-dataset evaluation. In the revision, we will revise the conclusion and discussion sections to state that the framework demonstrates strong performance on CASIA v2.0 and to explicitly note the limitations regarding generalization to other forgery types and uncontrolled conditions, while suggesting directions for future multi-dataset validation. revision: yes
Referee: [Experiments] The manuscript provides no details on training protocol (learning rate, optimizer, number of epochs, data augmentation, or train/validation/test split ratios) or cross-validation scheme. This absence makes the performance numbers non-reproducible and prevents assessment of whether the results are stable across different partitions.

Authors: We regret the omission of these critical details. The revised manuscript will include a dedicated experimental setup subsection specifying the optimizer (e.g., Adam), learning rate, batch size, number of epochs, data augmentation strategies, exact train/validation/test split ratios, and any cross-validation procedure used. revision: yes

Circularity Check

1 steps flagged

Youden-optimized threshold fitted on test set makes reported metrics a post-hoc fit rather than independent evaluation

specific steps

fitted input called prediction [Abstract and framework description (threshold optimization paragraph)]
"a model-specific adaptive threshold optimization strategy based on the Youden Index is employed to improve classification reliability by achieving a better balance between true positive and false positive rates."

The Youden Index is computed and maximized on the identical test partition whose accuracy, AUC, and MCC values are subsequently reported. The chosen threshold is therefore a fitted parameter of the evaluation set itself; the published numbers are the result of that fit rather than an independent forecast.

full rationale

The paper's central performance claims (accuracy, AUC, MCC etc.) rest on a hybrid RGB+FDIFF + CNN pipeline whose decision threshold is selected by maximizing the Youden Index directly on the CASIA v2.0 test split. Because the same data used for final reporting is also used to tune the threshold, the quoted metrics are statistically forced by construction and do not constitute an out-of-sample prediction. No other load-bearing step reduces to self-definition or self-citation; the CASIA-only evaluation is a generalization weakness, not a circular derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The work rests on standard pretrained CNN weights, the public CASIA v2.0 corpus, and the assumption that compression artifacts are reliable forgery cues. No new mathematical axioms or invented physical entities are introduced.

free parameters (1)

Youden-optimized decision threshold
Per-model threshold selected on the evaluation set to maximize Youden Index; this is a fitted hyperparameter.

pith-pipeline@v0.9.0 · 5591 in / 1096 out tokens · 35653 ms · 2026-05-12T00:57:25.530351+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 1 internal anchor

[1]

Multiple Image Splicing Dataset (MISD): A Dataset for Multiple Splicing,

K. D. Kadam, S. Ahirrao, and K. Kotecha, “Multiple Image Splicing Dataset (MISD): A Dataset for Multiple Splicing,”Data, vol. 6, no. 10, p. 102, 2021, doi: 10.3390/data6100102

work page doi:10.3390/data6100102 2021
[2]

A survey on deep learning-based image forgery detection,

F. Z. Mehrjardi, A. M. Latif, M. S. Zarchi, and R. Sheikhpour, “A survey on deep learning-based image forgery detection,”Pattern Recognition, vol. 144, p. 109778, 2023, doi: 10.1016/j.patcog.2023.109778

work page doi:10.1016/j.patcog.2023.109778 2023
[3]

Deep Learning Based Algorithm (ConvLSTM) for Copy Move Forgery Detection,

M. A. Elaskily, M. H. Alkinani, A. Sedik, and M. Dessouky, “Deep Learning Based Algorithm (ConvLSTM) for Copy Move Forgery Detection,”Journal of Intelligent & Fuzzy Systems, vol. 40, no. 3, pp. 4385–4405, 2021, doi: 10.3233/JIFS-201192

work page doi:10.3233/jifs-201192 2021
[4]

Digital Image Forgery Detection Approaches: A Review,

A. Mohassin and K. Farida, “Digital Image Forgery Detection Approaches: A Review,” inApplications of Artificial Intelligence in Engineering, Ad- vances in Intelligent Systems and Computing, vol. 1138, Springer, 2021, pp. 863–882, doi: 10.1007/978-981-33-4604-8_70

work page doi:10.1007/978-981-33-4604-8_70 2021
[5]

Image Splicing Forgery Detection Techniques: A Review,

K. B. Meena and V. Tyagi, “Image Splicing Forgery Detection Techniques: A Review,” inAdvances in Computing and Data Sciences, Communications inComputerandInformationScience(CCIS),vol.1441, Springer, 2021, pp. 364–388, doi: 10.1007/978-3-030-88244-0_35

work page doi:10.1007/978-3-030-88244-0_35 2021
[6]

Image forgery detection by transforming local descriptors into deep-derived features,

M. A. Anwar, S. F. Tahir, L. G. Fahad, and K. Kifayat, “Image forgery detection by transforming local descriptors into deep-derived features,”Applied Soft Computing, vol. 147, p. 110730, 2023, doi: 10.1016/j.asoc.2023.110730

work page doi:10.1016/j.asoc.2023.110730 2023
[7]

Image forgery detection: A survey of recent deep-learning techniques,

M. Zanardelli, D. Cozzolino, and L. Verdoliva, “Image forgery detection: A survey of recent deep-learning techniques,”Multimedia Tools and Applica- tions, vol. 82, pp. 12921–12966, 2023, doi: 10.1007/s11042-022-13797-w

work page doi:10.1007/s11042-022-13797-w 2023
[8]

In-Air Hand Ges- ture Signature Using Transfer Learning and Its Forgery Attack,

W. H. Khoh, Y. H. Pang, A. B. J. Teoh, and S. Y. Ooi, “In-Air Hand Ges- ture Signature Using Transfer Learning and Its Forgery Attack,”Applied Soft Computing, vol. 108, p. 107453, 2021, doi: 10.1016/j.asoc.2021.107453

work page doi:10.1016/j.asoc.2021.107453 2021
[9]

Deep residual learning for image recognition

K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Im- age Recognition,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90. 21

work page doi:10.1109/cvpr.2016.90 2016
[10]

Very Deep Convolutional Networks for Large-Scale Image Recognition,

K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” in Proc. Int. Conf. Learn. Representations (ICLR), 2015

work page 2015
[11]

Going Deeper with Convolutions,

C. Szegedy et al., “Going Deeper with Convolutions,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2015, pp. 1–9

work page 2015
[12]

Hybrid LSTM and Encoder-Decoder Architecture for Detection of Image Forgeries,

J. H. Bappy et al., “Hybrid LSTM and Encoder-Decoder Architecture for Detection of Image Forgeries,”IEEE Trans. Image Processing, vol. 28, no. 7, pp. 3119–3131, 2019

work page 2019
[13]

EfficientNet: Rethinking Model Scaling for Convolu- tional Neural Networks,

M. Tan and Q. Le, “EfficientNet: Rethinking Model Scaling for Convolu- tional Neural Networks,” in Proc. Int. Conf. Mach. Learn. (ICML), 2019

work page 2019
[14]

Densely Connected Convolutional Networks

G. Huang et al., “Densely Connected Convolutional Networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, doi: 10.1109/CVPR.2017.243

work page doi:10.1109/cvpr.2017.243 2017
[15]

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

A. G. Howard et al., “MobileNets: Efficient Convolutional Neural Net- works for Mobile Vision Applications,” arXiv:1704.04861, 2017, doi: 10.48550/arXiv.1704.04861

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1704.04861 2017
[16]

A comprehensive survey on cognitive cyber security analysis using machine learn- ing approaches,

A. H. Khalil et al., “Enhancing Digital Image Forgery Detection Us- ing Transfer Learning,”IEEE Access, vol. 11, 2023, doi: 10.1109/AC- CESS.2023.3307357

work page doi:10.1109/ac- 2023
[17]

Image Forgery Detection System using VGG16 UNET Model,

R. R. Choudhary, S. Paliwal, and G. Meena, “Image Forgery Detection System using VGG16 UNET Model,”Procedia Computer Science, vol. 235, pp. 735–744, 2024, doi: 10.1016/j.procs.2024.04.070

work page doi:10.1016/j.procs.2024.04.070 2024
[18]

ASCA-SqueezeNet: Hybrid Deep Learning Model for Digital Image Forgery Detection,

G. Nirmalapriya et al., “ASCA-SqueezeNet: Hybrid Deep Learning Model for Digital Image Forgery Detection,”Computers & Security, vol. 134, p. 103273, 2023, doi: 10.1016/j.cose.2023.103273

work page doi:10.1016/j.cose.2023.103273 2023
[19]

Single-image splicing localization through autoencoder-based anomaly detection,

D. Cozzolino and L. Verdoliva, “Single-image splicing localization through autoencoder-based anomaly detection,”IEEE Trans. Informa- tion Forensics and Security, vol. 12, no. 7, pp. 1566–1577, 2017, doi: 10.1109/TIFS.2017.2652079

work page doi:10.1109/tifs.2017.2652079 2017
[20]

A deep learning approach to universal image manipulation detection using a new convolutional layer,

B. Bayar and M. C. Stamm, “A deep learning approach to universal image manipulation detection using a new convolutional layer,” inProc. ACM Workshop on Information Hiding and Multimedia Security (IH&MMSec), 2016, pp. 5–10, doi: 10.1145/2909827.2930786

work page doi:10.1145/2909827.2930786 2016
[21]

Learning rich features for image manipulation detection,

P. Zhou et al., “Learning rich features for image manipulation detection,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 1053–1061, doi: 10.1109/CVPR.2018.00116

work page doi:10.1109/cvpr.2018.00116 2018
[22]

A deep learning approach to detection of splicing and copy-move forgeries,

Y. Rao and J. Ni, “A deep learning approach to detection of splicing and copy-move forgeries,” inProc. IEEE Int. Workshop Inf. Forensics Security (WIFS), 2016, pp. 1–6, doi: 10.1109/WIFS.2016.7823911. 22

work page doi:10.1109/wifs.2016.7823911 2016
[23]

Exploiting spatial structure for localizing manipulated image regions,

J. H. Bappy et al., “Exploiting spatial structure for localizing manipulated image regions,” inProc. IEEE Int. Conf. Comput. Vis. (ICCV), 2017, pp. 4980–4989, doi: 10.1109/ICCV.2017.532

work page doi:10.1109/iccv.2017.532 2017
[24]

ManTra-Net: Manipulation Tracing Network for Detection and Localization of Image Forgeries,

Y. Wu, W. Abd-Almageed, and P. Natarajan, “ManTra-Net: Manipulation Tracing Network for Detection and Localization of Image Forgeries,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR),2019, pp.9543– 9552, doi: 10.1109/CVPR.2019.00978

work page doi:10.1109/cvpr.2019.00978 2019
[25]

Image forgery localization via block- grained analysis of JPEG artifacts,

T. Bianchi and A. Piva, “Image forgery localization via block- grained analysis of JPEG artifacts,”IEEE Transactions on Informa- tion Forensics and Security, vol. 7, no. 3, pp. 1003–1017, 2012, doi: 10.1109/TIFS.2012.2187516

work page doi:10.1109/tifs.2012.2187516 2012
[26]

Toward Deep-Learning-Based Methods in Image Forgery Detection: A Survey,

N. T. Pham and C. S. Park, “Toward Deep-Learning-Based Methods in Image Forgery Detection: A Survey,”IEEE Access, vol. 11, pp. 11224– 11237, 2023, doi: 10.1109/ACCESS.2023.3241837

work page doi:10.1109/access.2023.3241837 2023
[27]

Image forgery detection,

H. Farid, “Image forgery detection,”IEEE Signal Processing Magazine, vol. 26, no. 2, pp. 16–25, 2009

work page 2009
[28]

Exposing DeepFake Videos By Detecting Face Warping Artifacts,

Y. Li and S. Lyu, “Exposing DeepFake Videos By Detecting Face Warping Artifacts,”IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

work page 2019
[29]

CASIA 2.0 Image Tampering Detection Dataset,

D. Gupta, “CASIA 2.0 Image Tampering Detection Dataset,” Kaggle,

work page
[30]

Available: https://www.kaggle.com/datasets/divg07/casia- 20-image-tampering-detection-dataset

[Online]. Available: https://www.kaggle.com/datasets/divg07/casia- 20-image-tampering-detection-dataset. [Accessed: Apr. 2, 2026]

work page 2026
[31]

Index for rating diagnostic tests,

W. J. Youden, “Index for rating diagnostic tests,”Cancer, vol. 3, no. 1, pp. 32–35, 1950. 23

work page 1950