The DOI matches its Crossref record; 49 of the 52 references it cites were found in a registry or an index; 9 of the 16 sources its claims cite were checked for support.
Forensic Analysis of Synthetically Generated Western Blot Images
Open accessAbstract
The widespread diffusion of synthetically generated content is a serious threat that needs urgent countermeasures. As a matter of fact, the generation of synthetic content is not restricted to multimedia data like videos, photographs or audio sequences, but covers a significantly vast area that can include biological images as well, such as western blot and microscopic images. In this paper, we focus on the detection of synthetically generated western blot images. These images are largely explored in the biomedical literature and it has been already shown they can be easily counterfeited with few hopes to spot manipulations by visual inspection or by using standard forensics detectors. To overcome the absence of publicly available data for this task, we create a new dataset comprising more than 14K original western blot images and 24K synthetic western blot images, generated using four different state-of-the-art generation methods.We investigate different strategies to detect synthetic western blots, exploring binary classification methods as well as one-class detectors. In both scenarios, we never exploit synthetic western blot images at training stage. The achieved results show that synthetically generated western blot images can be spot with good accuracy, even though the exploited detectors are not optimized over synthetic versions of these scientific images. We also test the robustness of the developed detectors against post-processing operations commonly performed on scientific images, showing that we can be robust to JPEG compression and that some generative models are easily recognizable, despite the application of editing might alter the artifacts they leave.
Overview
CoChat Summary
TL;DR
The paper creates a dataset of more than 14K real and 24K synthetic western blot images generated by Pix2pix, CycleGAN, StyleGAN2-ADA, and DDPM, then tests binary and one-class detectors that never train on synthetic western blots. It finds synthetic western blots can be detected with high AUC and balanced accuracy, with one-class detectors robust to JPEG compression but less robust to resizing.
Key findings
- The best binary detector, trained on both ProGAN and StyleGAN2 natural images, achieved AUCs of 0.959 for Pix2pix, 0.921 for CycleGAN, 0.915 for StyleGAN2-ADA, and 0.961 for DDPM; balanced accuracies were 92.58%, 84.55%, 88.93%, and 94.69%, respectively.
- The best one-class Isolation Forest detector achieved AUCs of 0.923 for Pix2pix, 0.961 for CycleGAN, 0.904 for StyleGAN2-ADA, and 0.999 for DDPM; balanced accuracies were 88.93%, 88.99%, 84.69%, and 92.65%, respectively. The one-class approach outperformed the binary one on CycleGAN and DDPM datasets, on DDPM only in AUC.
- One-class detection was robust to JPEG compression, with performance loss contained around 12% except for Pix2pix synthetic images; DDPM synthetic images were detected with high accuracy independently of the post-processing applied. Resizing remained challenging for most synthetic images.
- Feature analysis showed that for Pix2pix, CycleGAN, and DDPM images, single metrics (energy, homogeneity, and correlation, respectively) could differentiate classes; StyleGAN2-ADA images presented stronger artifacts along the vertical direction.
Methods
- The authors created a dataset from 284 original western blot images, resized so the minimum dimension was 256 pixels, yielding 14,200 real 256×256 patches; synthetic data comprised 24,000 256×256 images, with 6,000 each generated by Pix2pix, CycleGAN, StyleGAN2-ADA, and DDPM. Blot-masks were built via Otsu thresholding and Watershed segmentation, then selected by visual inspection.
- Binary detection used modified ResNet50 detectors from prior work, trained on natural images (ProGAN, StyleGAN2, or both) and never on western blots. One-class detection used high-pass filtering, gray-level co-occurrence matrices at distances 4, 8, 16, and 32 along horizontal and vertical directions, and five texture features (contrast, homogeneity, dissimilarity, energy, correlation), producing 40 features per image; OCSVM and Isolation Forest were trained on half of the real western blot patches (142 images, 7,100 patches) with 2-fold cross-validation.
- Performance was measured by AUC and balanced accuracy per generative model, and robustness was tested against upscaling (1.25, 1.5), down-upscaling (0.5, 0.75, 0.9), and JPEG compression (quality factors 80, 90, 100).
Limitations
- The paper states that additional research is needed to enhance robustness against standard processing operations and unseen generative models, and that robustness to resizing operations remains a challenging issue for spotting the majority of synthetic images.
- The paper states that the binary detector showed consistent performance loss in almost all post-processing scenarios, especially on Pix2pix and DDPM images, while the one-class classifier was significantly more robust against JPEG-based post-processing but not resizing, except for DDPM images.
- Evident from the text: the original dataset is based on only 284 source images, and synthetic data are generated by four specific models, so generalization to other generators and real-world processing chains is untested.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
7 Valid · 2 Not valid · 4 Not checked · 65 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“If the western blot image under analysis has been synthetically generated, either in its entirety or locally only in specific pixel regions, there would be essentially no hope to spot such traces by visual inspection [10].”
What the sources say
- Partly supportedcites [10] Qi, 2021, Emerging concern of scientific fraud: Deep learning and image manipulation, bioRxiv sourceExperimental — accuracy not yet measured
The indexed Abstract reports that researchers identified synthetic western blots with 0.52 accuracy, close to guessing, supporting the point that entirely generated blots are difficult to detect visually. It does not examine blots generated only in specific pixel regions, so the claim's extension to local generation is unsupported. (LLM, conf=0.85)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“As a matter of fact, forensics techniques aiming at spotting local image tampering have a hard time in detecting manipulations applied to scientific images. This is often due to their reduced pixel resolution and the numerous amount of processing operations applied to create realistic forgeries [13].”
What the sources say
- Partly supportedcites [13] sourceExperimental — accuracy not yet measured
The Results (p. 7) conclude that existing natural-image forensic methods are not robust on biomedical images, supporting the claimed detection difficulty. Section 4 (pp. 5–6) discusses different challenges, including annotations, similar-looking images and plain textures; it does not attribute the difficulty to reduced pixel resolution or numerous processing operations used to make realistic forgeries. (LLM, conf=0.87)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Indeed, during some preliminary experiments, the authors of [10] verified that standard image generation techniques based on GANs [14], [15] can synthesize almost indistinguishable western blots with respect to the real ones, even at the experts' eyes.”
What the sources say
- Supportedcites [10] Qi, 2021, Emerging concern of scientific fraud: Deep learning and image manipulation, bioRxiv source
The supplied passages (labelled Abstract, no page given) report that the authors trained GAN-based image-translation models to generate western blots and asked researchers familiar with western blotting to distinguish synthetic images from real ones. Overall accuracy was 0.52, close to chance, supporting the claim that the images were almost indistinguishable to those researchers. (LLM, conf=0.93)
- Not checked — no full text of the sourcecites [14] source
- Partly supportedcites [15] sourceExperimental — accuracy not yet measured
The Related Work section (p. 2) describes GANs as methods for generating realistic images, and the Experiments section (p. 6) reports realistic outputs on scene-image datasets. The paper does not report synthesizing western blots or testing whether experts can distinguish them from real blots; its human evaluation compares outputs from different synthesis methods using Mechanical Turk workers. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“This result confirms the experiments performed in the original paper [16]: the bigger the training dataset, the better the generalization capability of the detector.”
What the sources say
- Partly supportedcites [16] sourceExperimental — accuracy not yet measured
The Experimental Results (p. 4) say that training-data diversity increases robustness, and the authors report results for a detector trained on 23 StyleGAN2 models. They do not establish the claimed relationship between the number of training images and generalization: the compared training sets also differ in their source models (Experimental Results, pp. 3–4). (LLM, conf=0.76)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The achieved results show that synthetically generated western blot images can be spot with good accuracy, even though the exploited detectors are not optimized over synthetic versions of these scientific images.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on 2 assumptions not shown here.
37 more claims
Create a free account to see all 78 claims and to verify the 1 reference still open.
- It may state two separate claims as one. 28
- Probably not a substantial claim. 12
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.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ C. One-class detection results
“Pix2pix, CycleGAN and DDPM images can be differentiated by one single metric, f e , f h and f ρ , respectively”
§ C. One-class detection results
“| CycleGAN | 0 . 586 | 0 . 770 | 0 . 637 | 0 . 816 | 0 . 955 | | StyleGAN2-ADA | 0 . 731 | 0 . 578 | 0 . 772 | 0 . 677 | 0 . 894 | | DDPM | 0 . 788 | 0 . 976 | 0 . 853 | 0 . 955 | 0 . 743 |”
Citations
Cites
A Threshold Selection Method from Gray-Level Histograms
Otsu, N. (1979) · IEEE Transactions on Systems, Man, and Cybernetics · doi:10.1109/tsmc.1979.4310076 · Cited by 43,398 · cited 1×
“For each image, we exploit Otsu’s image thresholding [26] andWatershed segmentation [27] algorithms to automatically obtain possible blot-masks associatedwith the image, thenwe pick the best mask by visual inspection.”
MethodUnpaired image-to-image translation using cycle-consistent adversarial networks
Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.244 · Cited by 23,602 · cited 13×
“The main feature of CycleGAN is its ‘‘cycle-consistency’’ property which translates an input image to an output meaningful synthetic image belonging to a different category [17].”
MethodInfluentialImage-to-Image Translation with Conditional Adversarial Networks
Isola, P., Zhu, J.-Y., Zhou, T., & Efros, A. A. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.632 · Cited by 22,930 · cited 4×
“Among the various methods presented in the literature, we focus on Pix2pix [14] and CycleGAN [17] models, being two of the best performing and most widespread generation methods.”
BackgroundMethodInfluentialAdvances in Neural Information Processing Systems 19
(2007) · The MIT Press eBooks · doi:10.52202/085713 · Cited by 15,539
A Style-Based Generator Architecture for Generative Adversarial Networks
Karras, T., Laine, S., & Aila, T. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00453 · Cited by 11,714
Denoising Diffusion Probabilistic Models
Ho, J., Jain, A., & Abbeel, P. (2020) · arXiv · doi:10.48550/arxiv.2006.11239 · Cited by 8,712
Liu, F. T., Ting, K. M., & Zhou, Z.-H. (2008) · 2008 Eighth IEEE International Conference on Data Mining · doi:10.1109/icdm.2008.17 · Cited by 6,601 · cited 1×
“sifier, investigating both the well-known andwidely exploited One Class Support Vector Machine (OCSVM) [35] that we consider as a baseline reference and the more recent Isolation Forest (IF) [36], which has been proposed as efficient strategy for anomaly detection.”
BackgroundAnalyzing and Improving the Image Quality of StyleGAN
Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., & Aila, T. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00813 · Cited by 6,332
Estimating the Support of a High-Dimensional Distribution
Schölkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J., & Williamson, R. C. (2001) · Neural Computation · doi:10.1162/089976601750264965 · Cited by 6,168 · cited 9×
“Tables 3 and 4 show the best achieved AUC on each selected feature by exploiting OCSVM and IF, respectively.”
BackgroundInfluentialWatersheds in digital spaces: an efficient algorithm based on immersion simulations
Vincent, L., & Soille, P. (1991) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/34.87344 · Cited by 5,649 · cited 1×
“For each image, we exploit Otsu’s image thresholding [26] andWatershed segmentation [27] algorithms to automatically obtain possible blot-masks associatedwith the image, thenwe pick the best mask by visual inspection.”
Method
Research works printed but not linked: 4
- 2022, Science Integrity Digest: Western Blots
- 2022, Flickr-Faces-HQ Dataset (FFHQ)
- 2022, Metfaces Dataset
- 2022, Scikit-Learn: Novelty and Outlier Detection
Cited by
A review of deep learning‐based approaches for deepfake content detection
(2024) · doi:10.1111/exsy.13570 · Cited by 116
“Currently, there are still very few studies carried out in the sense of using diffusion models to generate deepfake [142, 143] and, in the same sense, few studies involving the detection of deepfake also generated by diffusion models.”
BackgroundA product form design method integrating Kansei engineering and diffusion model
Yang, C., Liu, F., & Ye, J. (2023) · Advanced Engineering Informatics · doi:10.1016/j.aei.2023.102058 · Cited by 97
Detecting Gan-Generated Images by Orthogonal Training of Multiple CNNs
Mandelli, S., Bonettini, N., Bestagini, P., & Tubaro, S. (2022) · Proceedings - International Conference on Image Processing · doi:10.1109/icip46576.2022.9897310 · Cited by 115
“For this reason, the latest research trend is to develop methods that can generalize well, detecting images generated with unseen techniques [11, 14].”
BackgroundSelf-citationA Review of Deep Learning-based Approaches for Deepfake Content Detection
Passos, L. A., Jodas, D., Costa, K. A. P., Júnior, L. A. S., Rodrigues, D., Ser, J. D., Camacho*, D., & Papa, J. (2023) · doi:10.22541/au.169735672.27713914/v1 · Cited by 116
How to stop AI deepfakes from sinking society — and science
Jones, N. (2023) · Nature · doi:10.1038/d41586-023-02990-y · Cited by 18
Metadata
- Authors
- Mandelli, S., Cozzolino, D., Cannas, E. D., Cardenuto, J. P., Moreira, D., Bestagini, P., Scheirer, W. J., Rocha, A., Verdoliva, L., Tubaro, S., & Delp, E. J.
- Affiliations
- Politecnico di Milano (IT)
- University of Naples Federico II (IT)
- Artificial Intelligence Laboratory, RECOD.ai, Institute of Computing, University of Campinas, Campinas, Brazil
- University of Notre Dame (US)
- Universidade Estadual de Campinas (UNICAMP) (BR)
- Purdue University West Lafayette (US)
- ORCID
- Venue
- IEEE Access
- Published
- 2022
- Volume
- 10
- Pages
- 59919-59932
- ISSN
- 2169-3536 (ISSN-L)
- OpenAlex
- W4285176942
- Cited by
- 26
- References
- 52
- Citations by year
- 2026: 2 · 2025: 7 · 2024: 12 · 2023: 3 · 2022: 2
- FWCI
- 1.96
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- IEEE Access (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- Virtual Community of Pathological Anatomy (University of Castilla La Mancha) (submittedVersion, other-oa)
- Loyola eCommons (Loyola University of Chicago) (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- arXiv (Cornell University)
- Greater South Information System
- Funders
- Defense Advanced Research Projects Agency (DARPA) and the Air Force Research Laboratory (FA8750-20-2-1004)
- PREMIER project, funded by the Italian Ministry of Education, University, and Research within the PRIN 2017 program
- NVIDIA corporation
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
- Physical Sciences › Mathematics › Statistics and Probability › Benford’s Law and Fraud Detection
- Keywords
- synthetic content detection, Western blot images, image forensics, generative models, binary classification, JPEG compression robustness
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 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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: Forensic Analysis of Synthetically Generated Western Blot Images. https://app.cochat.ai/papers/10.1109/access.2022.3179116
- The paper
- Mandelli, S., Cozzolino, D., Cannas, E. D., Cardenuto, J. P., Moreira, D., Bestagini, P., Scheirer, W. J., Rocha, A., Verdoliva, L., Tubaro, S., & Delp, E. J. (2022). Forensic Analysis of Synthetically Generated Western Blot Images. IEEE Access. https://doi.org/10.1109/access.2022.3179116
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.