The DOI matches its Crossref record; 4 of the 4 references it cites were found in a registry or an index.
A Domain Shift Evaluation Protocol for Fake Satellite Image Detection: CNN Superiority and the Enhanced Model Paradox
Abstract
AI-generated images have achieved an unparalleled realism that has proven hard to be detected by human naked eye. Specialized software to detect such content is crucial in the field of geospatial intelligence, because there is a great potential for fake satellite images to change the perception of crucial infrastructure, strategic decisions and other information that is collected via satellite imagery. For this reason, we propose a domain shift evaluation protocol for assessing fake satellite image detection models under realistic conditions. Our protocol evaluates the impact of studying models that were trained on one set of synthetic imagery against completely different generation methods (e.g different random seeds, or different image generator such as VAEs, diffusion models, etc). This effectively, helps simulating operational scenarios where novel synthesis (out of distribution) techniques may be used for deception or cyber-manipulation. Using this protocol, we analyze multiple architectures including CNNs, Vision Transformers, and CLIP models. Our main findings are that (1) CNNs are better at generalizing on unseen stimuli (domain shift robustness), with all CNN models achieving 99.9% accuracy compared to 96.7%-99.8% for Vision Transformers, ≥ (2) CLIP models, reach perfect test accuracy in domain dataset, but they show much larger performance degradation (0.44–0.68% detection rates) on the new domains, and (3) the enhanced model paradox, where training CLIP models with 64% more diverse data leads to 36% worse performance due to over-specialization on the combined training distribution rather than learning generalizable detection principles. Both CLIP models show concerning overconfidence (95–98%) in incorrect predictions. These results establish CNNs as the preferred architecture for satellite image authentication while highlighting critical evaluation methodology limitations that our proposed protocol addresses.
Overview
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
High-Resolution Image Synthesis with Latent Diffusion Models
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01042 · Cited by 20,302
Emerging Properties in Self-Supervised Vision Transformers
Caron, M., Touvron, H., Misra, I., Jegou, H., Mairal, J., Bojanowski, P., & Joulin, A. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00951 · Cited by 5,638
Media Forensics and DeepFakes: An Overview
Verdoliva, L. (2020) · IEEE Journal of Selected Topics in Signal Processing · doi:10.1109/jstsp.2020.3002101 · Cited by 979
Understanding Robustness of Transformers for Image Classification
Bhojanapalli, S., Chakrabarti, A., Glasner, D., Li, D., Unterthiner, T., & Veit, A. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.01007 · Cited by 440 · cited 1×
“ViTs have minimal built-in priors and usually need much more data to match generalization performance of CNNs [23], [24].”
Research works printed but not linked: 20
- Chai, Deepfake detection in the wild
- Tolga, 2020, Deepfake detection using xceptionnet, Deep Learning for Media Forensics
- 10.1109/CVPR42600.2020.00327 · DOI · Paper page
- Li, Faceforensics++: Learning to detect manipulated facial images
- Rössler, Faceforensics++: Detecting manipulated facial videos with deep learning
- 10.1109/CVPR42600.2020.00791 · DOI · Paper page
- Shamshad, 2023, Fake or real? detecting synthetic satellite imagery, Remote Sensing
- 10.1561/116.00000072 · DOI
- Radford, Learning transferable visual models from natural language supervision
- Guo, On calibration of modern neural networks
- Wang, Generalizing across domains: A survey on domain generalization
- 10.1109/TPAMI.2022.3195549 · DOI · Paper page
- Gulrajani, In search of lost domain generalization
- Tan, Efficientnet: Rethinking model scaling for convolutional neural networks
- Dosovitskiy, An image is worth 16x16 words: Transformers for image recognition at scale
- 10.1016/j.ipm.2009.03.002 · DOI · Paper page
- Shu, Clipood: Generalizing clip to out-of-distributions, 31 716
- Fang, Clip’s magic ingredient: Data diversity in pre-training, 456
- Raghu, 2021, Do vision transformers see like convolutional neural networks?, Advances in Neural Information Processing Systems (NeurIPS)
- 10.1109/CVPR52688.2022.01167 · DOI · Paper page
Metadata
- Authors
- Anand, V., & Rodriguez, I. F.
- Venue
- 2026 IEEE 5th International Conference on AI in Cybersecurity (ICAIC)
- Publisher
- IEEE
- Published
- 2026-02-18
- Pages
- 1-8
- Cited by
- 0
- References
- 4
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.
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: A Domain Shift Evaluation Protocol for Fake Satellite Image Detection: CNN Superiority and the Enhanced Model Paradox. https://app.cochat.ai/papers/10.1109/icaic67076.2026.11395782
- The paper
- Anand, V., & Rodriguez, I. F. (2026). A Domain Shift Evaluation Protocol for Fake Satellite Image Detection: CNN Superiority and the Enhanced Model Paradox. 2026 IEEE 5th International Conference on AI in Cybersecurity (ICAIC). https://doi.org/10.1109/icaic67076.2026.11395782
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.