A Lightweight and Effective Image Tampering Localization Network with Vision Mamba

Gang Cao; Jiaoyun Liu; Kun Guo; Xianglin Huang; Zijie Lou

REVIEW 2 major objections 6 minor 1 cited by

A Lightweight and Effective Image Tampering Localization Network with Vision Mamba

T0 review · 2 major / 6 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 ForMa, a vision-Mamba network for tampering localization, reports the best average F1 and IoU across ten datasets while using the least computation.

desk verdict Useful lightweight Mamba-based tampering localizer, but the SOTA claim rests on a baseline comparison that doesn't control training data. read the letter →

arxiv 2502.09941 v1 pith:KOP2LEHX submitted 2025-02-14 cs.CV cs.CR

Kun Guo , Gang Cao , Zijie Lou , Xianglin Huang , Jiaoyun Liu This is my paper · ORCID

classification cs.CVcs.CR

keywords imagetamperinglocalizationforensicsvisionMambastatespacemodelpixelshufflenoise-assisteddecodinglightweightnetwork

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

This paper proposes ForMa, a tampering-localization network built on the vision Mamba state-space backbone, and claims it locates edited regions more accurately than existing CNN- and Transformer-based methods while using less computation. It argues that a selective-scan state-space encoder can supply global context at linear complexity, replacing the quadratic attention of Transformers in forensic models. ForMa adds a decoder that upsamples with parameter-free pixel shuffle and a noise-assisted decoding strategy that injects Noiseprint++, SRM, and Bayar features at the decoder rather than the encoder. Across ten cross-domain test sets, the paper reports an average F1 of 64.1% and IoU of 56.2%, ahead of the best Transformer baseline TruFor (59.8%/51.1%) and the best CNN baseline CAT-Net (37.6%/32.0%), with only 42G FLOPs at 512x512. These results lead the authors to conclude that a lightweight forensic model can generalize across splicing, copy-move, deepfake, and diffusion-generated forgeries and survive social-network post-processing.

What carries the argument

The central mechanism is the VSS encoder built on the selective state-space model S6 from Mamba, realized as SS2D: image patches are unfolded into sequences along four scanning directions, processed by parallel S6 systems, and merged back into spatial feature maps, giving global dependency modeling with linear complexity. The decoder's key operation is pixel shuffle, a parameter-free layer that rearranges channel dimensions into spatial resolution, replacing bilinear interpolation and reducing FLOPs. Around these, ForMa uses multi-scale feature maps from four VSS stages and a noise-assisted decoding strategy that concatenates forensic residuals from Noiseprint++, SRM filters, and Bayar convolution in the decoder, fusing them with encoder features before pixel-wise prediction.

What would settle it

Retrain TruFor, IML-ViT, SparseViT, and CAT-Net on the same CAT-Net training set under the same 512x512 protocol, evaluate all models on the same ten datasets, and recompute average F1 and IoU; if TruFor's average F1 matches or passes 64.1%, the claimed state-of-the-art advantage does not hold.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a visual state-space encoder built from Mamba's selective scanning can serve as the backbone for blind image tampering localization and outperform both convolutional and attention-based architectures in average accuracy on ten standard datasets. ForMa reaches an average F1 of 64.1% and IoU of 56.2%, compared with 59.8%/51.1% for TruFor and 37.6%/32.0% for CAT-Net, while requiring 42G FLOPs at 512x512 input, the lowest among the compared models. The explanation offered is that the SS2D module captures multi-scale global dependencies at linear cost, the shuffle-based decoder preserves spatial detail without interpolation parameters, and the noise-assisted decoding strategy supplies complementary manipulation traces exactly where the final prediction is made. The paper therefore frames ForMa not as an incremental CNN/Transformer hybrid but as a new backbone choice for forensic localization.

Load-bearing premise

The comparison claim rests on the assumption that the reported scores for CAT-Net, TruFor, IML-ViT, and SparseViT were obtained under the same training data and evaluation conditions as ForMa; the paper says ForMa was trained on the CAT-Net dataset but does not say the baselines were retrained there.

Editorial extensions

If this is right

A state-space encoder gives global context at linear cost, so the same design can scale to higher-resolution forensic inputs without the quadratic blow-up of attention.
ForMa's gains on FF++ and CoCoGlide imply that a model trained only on classic splicing and copy-move data can transfer to deepfake and diffusion-generated manipulations.
Injecting noise features at the decoder rather than the encoder improves accuracy while cutting FLOPs by roughly 190G in the Korus ablation, suggesting a general design rule for forensic decoders.
Pixel-shuffle upsampling removes interpolation parameters from the decoder, and other dense-prediction forensic tasks could adopt the same parameter-free upsampling.
A single lightweight model could serve as a general-purpose blind tampering localizer, since it reports the best average F1 across all ten test sets at the lowest compute.

Reading between the lines

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

A direct per-dataset ablation of the noise extractor would show whether decoder-stage noise injection is most valuable on deepfake and diffusion data, as the large FF++ jump suggests.
Replacing the VMamba backbone with a CNN or Transformer while keeping the identical shuffle decoder and noise strategy would isolate how much of the gain comes from the state-space encoder versus the new decoder.
Because initialization uses ImageNet-pretrained VMamba weights, testing the same architecture from random initialization would separate pretraining benefits from the selective-scan mechanism.

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.

Desk Editor's Note

read the letter

The thing to know: ForMa is a clean, cheap Mamba-based tampering localization network that appears to beat TruFor on average F1 while using roughly a fifth of the FLOPs. I wouldn't trust that headline yet, because the baseline comparison doesn't clearly establish that the other methods were trained on the same data.

What's actually new: the paper is the first to put a visual state-space backbone (VMamba) on blind tampering localization, and it pairs that with a pixel-shuffle decoder and decoder-stage fusion of noise features rather than early fusion. The ablations support each piece—noise fusion, pixel shuffle, and the combined loss all move F1 on the tested sets. The compute story is credible: 42G FLOPs at 512×512 and 37M parameters, below SparseViT's 46G and far below TruFor's 231G. Code is public, which makes the work reproducible.

The soft spot is load-bearing. Section III-A says, "Consistent with CAT-Net, TruFor, IML-ViT and SparseViT, our ForMa is trained on the CAT-Net dataset." That is not a statement that the baselines were retrained on CAT-Net. The TruFor numbers in Table I (e.g., DSO 91.0, FF++ 69.2) look like they came from the original TruFor paper, which trained on a different, larger dataset. If Table I is a mix of original-paper results, then the average-F1 advantage is confounded by training distribution, not just architecture. Per dataset, the pattern is also less flattering than the average suggests: ForMa gets 38.7 on DSO against TruFor's 91.0, and 30.4 on Korus against 37.7. The impressive gains on FF++ and MISD carry the average. So "state-of-the-art generalization ability" is overclaimed unless the authors clarify that all baselines were evaluated under identical training conditions.

Minor issues: no error bars or variance, and averaging F1 over datasets of very different sizes is coarse, though they cite a prior paper for the protocol. None of this undercuts the engineering contribution.

This paper is for someone who cares about efficient forensics on normal hardware. I'd send it to peer review, but the referee should require a controlled comparison: retrain the baselines on the same CAT-Net split, or report both original and retrained numbers, and give per-dataset variance. The architecture is worth a look; the headline just needs proof.

Referee Report

2 major / 6 minor

Summary. The paper introduces ForMa, a lightweight image tampering localization network based on a Vision Mamba (VMamba) encoder, a shuffle-based decoder with pixel-shuffle upsampling, and a noise-assisted decoding strategy that combines Noiseprint++, SRM, and Bayar convolution features. The authors report a state-of-the-art average F1 of 64.1% and IoU of 56.2% over 10 tampering localization datasets, while claiming the lowest computational cost (42G FLOPs at 512x512, 37M parameters) among compared CNN, Transformer, and Mamba methods. The manuscript also presents ablations and robustness experiments against post-processing.

Significance. If the reported results are verified under a fair comparison protocol, ForMa would be a meaningful contribution to the tampering-localization literature by demonstrating that a state-space-model backbone can achieve competitive accuracy with substantially lower complexity than Transformer-based methods. The paper provides code and evaluates on a broad set of 10 datasets, which is valuable. However, the validity of the central SOTA claim currently rests on the assumption that baselines were evaluated under identical training conditions, which is not explicitly documented.

major comments (2)

[Section III-A, Table I] The training protocol for the baselines is unclear. The sentence "Consistent with CAT-Net, TruFor, IML-ViT and SparseViT, our ForMa is trained on the CAT-Net dataset" does not confirm that all baseline models were retrained on the same CAT-Net training set or that their reported numbers were generated under identical conditions (e.g., same image preprocessing, thresholding, and evaluation protocol). If the baseline numbers are taken from the original publications, differences in training data (for instance, TruFor used a larger custom training set) confound the claimed average-F1 advantage of +4.3% over TruFor. The authors must provide a precise description of the training and evaluation protocol for each baseline, or retrain all baselines in the same setting, to support the state-of-the-art claim.
[Section II-B, Table II] The computational complexity comparison is incomplete without a detailed account of the noise extractor. The noise-assisted decoding strategy uses Noiseprint++, SRM, and Bayar convolution, but the manuscript does not specify the architecture, parameter count, or FLOPs of these components or whether they are frozen or fine-tuned. If Noiseprint++ is a deep network and is executed during inference, then the reported 42G FLOPs and 37M parameters likely undercount the true complexity, making the "lowest computational complexity" claim unfair. The authors should clarify whether the FLOPs/parameters in Table II include the entire inference pipeline, and provide architecture details for the noise extractor.

minor comments (6)

[Section II-C] The phrase "an linear layer" is a grammatical error; it should be "a linear layer."
[Fig. 2 caption] The word "layar" in the caption should be "layer."
[Eq. (3)] The use of "∀i" after the definition of F is confusing because F is not indexed by i; please rewrite the equation to make clear that the four Fi are concatenated in the Concat operation.
[Section III-D] The claim that ForMa "maintains a performance advantage" over TruFor is true only on average across the four datasets; on DSO, TruFor achieves higher F1 in several settings (e.g., Facebook row: TruFor 65.4 vs. ForMa 39.4). Please qualify the statement accordingly.
[Section III-A] The paper does not specify whether the average F1 and IoU is an unweighted macro-average over the 10 datasets or a micro-average; please provide the exact formula or the specific equation from reference [26] that was followed.
[Table III] The ablation row "Noise fed into Encoder" is not described in the text; please explain how this variant was constructed and how it differs from the proposed decoder-side noise injection.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ForMa's SOTA claim rests on external benchmarks and ablations, not on self-referential derivation.

full rationale

This is an empirical engineering paper whose central claims—best average F1/IoU over 10 datasets and lowest FLOPs—are established by training ForMa on the CAT-Net dataset and evaluating it on external test sets, with ablation studies isolating each design component. None of the reported equations (e.g., the S6 state-space equations from Mamba, the pixel-shuffle decoder formulation, or the noise-feature fusion) assumes the target result. The backbone is a pretrained VMamba-tiny, and the noise extractors (Noiseprint++, SRM, Bayar) are standard external components; no parameter is fitted to the benchmark numbers and then renamed as a prediction. The only self-citations ([2] and [26]) concern a loss-function recipe and an averaging convention for mean F1/IoU, both of which are methodological choices rather than evidence for the SOTA claim. The skeptical concern that baselines such as TruFor may not have been retrained on CAT-Net is a legitimate question of experimental fairness and benchmark protocol, not circularity, because the paper's reported numbers are not forced by construction from its assumptions. Therefore the derivation chain is self-contained with respect to the claimed contribution, and the appropriate circularity score is 0.

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

The central empirical claim rests on trained network weights, pretrained backbones, and the CAT-Net training distribution; no new physical or conceptual entities are introduced.

free parameters (4)

Embedding dimension C = 96
Default feature dimension chosen by design, not derived; affects all decoder linear layers and model capacity.
Expansion factors r_i = [1, 2, 4, 8]
Upsampling ratios for the four multi-scale decoder branches; chosen by hand, control the pixel shuffle operation.
VSS block depths L_i = {2, 2, 9, 2}
Number of VSS blocks per stage, taken from VMamba-tiny configuration; not derived for this task.
Evaluation threshold = 0.5
F1/IoU are computed with a default 0.5 threshold, following prior work; a different threshold would change all scores.

assumptions (4)

domain assumption VMamba-tiny backbone pretrained on ImageNet yields visual features suitable for forensic tampering localization.
The network is initialized with ImageNet pretrained weights; the transferability of natural-image features to forensic traces is assumed (Section III-A Implementation Details).
domain assumption The CAT-Net training set of over 800k forged images is sufficient to train a model that generalizes across the 10 test domains.
The method is only trained on CAT-Net; the paper assumes this training distribution covers splicing, copy-move, removal, and other manipulations needed for the test sets (Section III-A Datasets).
domain assumption SRM, Noiseprint++, and Bayar convolution provide complementary and reliable manipulation traces.
These noise extractors are used as fixed (or pretrained) feature extractors; the paper assumes their outputs are informative for the decoder (Section II-B).
domain assumption The SS2D selective scan preserves 2D spatial context necessary for pixel-wise localization.
The method relies on cross-scan/cross-merge to serialize image patches for S6; the paper assumes no loss of spatial structure (Section II-A).

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Lightweight and Effective Image Tampering Localization Network with Vision Mamba." pith.science (2026). https://pith.science/paper/KOP2LEHX

@misc{pith2026250209941,
  author       = {Pith},
  title        = {Pith review of: A Lightweight and Effective Image Tampering Localization Network with Vision Mamba},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KOP2LEHX}},
  note         = {Machine review of arXiv:2502.09941}
}

read the original abstract

Current image tampering localization methods primarily rely on Convolutional Neural Networks (CNNs) and Transformers. While CNNs suffer from limited local receptive fields, Transformers offer global context modeling at the expense of quadratic computational complexity. Recently, the state space model Mamba has emerged as a competitive alternative, enabling linear-complexity global dependency modeling. Inspired by it, we propose a lightweight and effective FORensic network based on vision MAmba (ForMa) for blind image tampering localization. Firstly, ForMa captures multi-scale global features that achieves efficient global dependency modeling through linear complexity. Then the pixel-wise localization map is generated by a lightweight decoder, which employs a parameter-free pixel shuffle layer for upsampling. Additionally, a noise-assisted decoding strategy is proposed to integrate complementary manipulation traces from tampered images, boosting decoder sensitivity to forgery cues. Experimental results on 10 standard datasets demonstrate that ForMa achieves state-of-the-art generalization ability and robustness, while maintaining the lowest computational complexity. Code is available at https://github.com/multimediaFor/ForMa.

Figures

Figures reproduced from arXiv: 2502.09941 by the authors.

**Figure 2.** (a) Architecture of proposed ForMa. Li={2, 2, 9, 2}. Linear, Conv, and Shuffle refers to the linear, convolution and pixel shuffle layers, respectively. ⊕ represents element-wise addition. (b) Structure of the VSS Block. It includes a depthwise convolutional layar (DWConv), SiLU activation function, SS2D module, and linear normalization (LN). The VSS Block, Shuffle-based Decoder, and Noise Extractor are learnable. …… view at source ↗

**Figure 3.** Illustration of the SS2D module. The S6 used is from Mamba [11]. [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Results on example images from NIST, Korus and CASIAv1 datasets. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: Robustness evaluation against different image post-processing tech [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Unmasking Synthetic Realities in Generative AI: A Comprehensive Review of Adversarially Robust Deepfake Detection Systems
cs.CR 2025-07 conditional novelty 3.0 of 10

A systematic review of deepfake detection finds a pervasive lack of adversarial robustness evaluation across all modalities and calls for resilient, modality-agnostic detectors.

Reference graph

Works this paper leans on

28 extracted references · 22 canonical work pages · cited by 1 Pith paper

[1]

Lightweight and high-precision net- work for image copy-move forgery detection,

Y . Shi, S. Weng, L. Yu, and L. Li, “Lightweight and high-precision net- work for image copy-move forgery detection,” IEEE Signal Processing Letters, vol. 31, pp. 1409–1413, 2024

work page 2024
[2]

Effective image tampering localization via enhanced transformer and co-attention fusion,

K. Guo, H. Zhu, and G. Cao, “Effective image tampering localization via enhanced transformer and co-attention fusion,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, 2024, pp. 4895–4899

work page 2024
[3]

Effective image tampering localization with multi-scale convnext feature fusion,

H. Zhu, G. Cao, M. Zhao, H. Tian, and W. Lin, “Effective image tampering localization with multi-scale convnext feature fusion,”Journal of Visual Communication and Image Representation, vol. 98, p. 103981, 2024

work page 2024
[4]

MVSS-Net: Multi- view multi-scale supervised networks for image manipulation detec- tion,

C. Dong, X. Chen, R. Hu, J. Cao, and X. Li, “MVSS-Net: Multi- view multi-scale supervised networks for image manipulation detec- tion,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 3, pp. 3539–3553, 2022

work page 2022
[5]

Learning jpeg compression artifacts for image manipulation detection and localization,

M.-J. Kwon, S.-H. Nam, I.-J. Yu, H.-K. Lee, and C. Kim, “Learning jpeg compression artifacts for image manipulation detection and localization,” International Journal of Computer Vision , vol. 130, no. 8, pp. 1875– 1895, 2022

work page 2022
[6]

PSCC-Net: Progressive spatio- channel correlation network for image manipulation detection and localization,

X. Liu, Y . Liu, J. Chen, and X. Liu, “PSCC-Net: Progressive spatio- channel correlation network for image manipulation detection and localization,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 11, pp. 7505–7517, 2022

2022
[7]

Hierarchical fine-grained image forgery detection and localization,

X. Guo, X. Liu, Z. Ren, S. Grosz, I. Masi, and X. Liu, “Hierarchical fine-grained image forgery detection and localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

work page 2023
[8]

TruFor: Leveraging all-round clues for trustworthy image forgery detection and localization,

F. Guillaro, D. Cozzolino, A. Sud, N. Dufour, and L. Verdoliva, “TruFor: Leveraging all-round clues for trustworthy image forgery detection and localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 20 606–20 615

work page 2023

Show all 28 references

[9]

IML-ViT: Image manipulation localization by vision transformer,

X. Ma, B. Du, X. Liu, A. Y . A. Hammadi, and J. Zhou, “IML-ViT: Image manipulation localization by vision transformer,” Proceedings of the AAAI Conference on Artificial Intelligence , 2024

2024
[10]

A new approach to linear filtering and prediction problems,

R. E. Kalman, “A new approach to linear filtering and prediction problems,” Journal of Basic Engineering , pp. 35–45, 1960

1960
[11]

Mamba: Linear-time sequence modeling with selective state spaces,

A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

2023 arXiv
[12]

VMamba: Visual state space model,

Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, and Y . Liu, “VMamba: Visual state space model,” arXiv preprint arXiv:2401.10166, 2024

2024 arXiv
[13]

ManTra-Net: Manipulation tracing network for detection and localization of image forgeries with anomalous features,

Y . Wu, W. AbdAlmageed, and P. Natarajan, “ManTra-Net: Manipulation tracing network for detection and localization of image forgeries with anomalous features,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 9543–9552

2019
[14]

Efficiently modeling long sequences with structured state spaces,

A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,” arXiv preprint arXiv:2111.00396 , 2021

2021 arXiv
[15]

Rich models for steganalysis of digital images,

J. Fridrich and J. Kodovsky, “Rich models for steganalysis of digital images,” IEEE Transactions on Information Forensics and Security , vol. 7, no. 3, pp. 868–882, 2012

2012
[16]

Can we get rid of handcrafted feature extractors? SparseViT: Nonsemantics- centered, parameter-efficient image manipulation localization through spare-coding transformer,

L. Su, X. Ma, X. Zhu, C. Niu, Z. Lei, and J.-Z. Zhou, “Can we get rid of handcrafted feature extractors? SparseViT: Nonsemantics- centered, parameter-efficient image manipulation localization through spare-coding transformer,” in Proceedings of the AAAI Conference on Artificia...

2025
[17]

Casia image tampering detection evaluation database,

J. Dong, W. Wang, and T. Tan, “Casia image tampering detection evaluation database,” in Proceedings of the IEEE China Summit and International Conference on Signal and Information Processing , 2013, pp. 422–426

2013
[18]

Columbia uncompressed image splicing detection evaluation dataset,

J. Hsu and S. Chang, “Columbia uncompressed image splicing detection evaluation dataset,” Columbia DVMM Research Lab , 2006

2006
[19]

MFC Datasets: Large- scale benchmark datasets for media forensic challenge evaluation,

H. Guan, M. Kozak, E. Robertson, Y . Lee, A. N. Yates, A. Delgado, D. Zhou, T. Kheyrkhah, J. Smith, and J. Fiscus, “MFC Datasets: Large- scale benchmark datasets for media forensic challenge evaluation,” in Proceedings of the IEEE Winter Applications of Computer Vision Worksho...

2019
[20]

Exposing digital image forgeries by illumination color classification,

T. J. De Carvalho, C. Riess, E. Angelopoulou, H. Pedrini, and A. de Rezende Rocha, “Exposing digital image forgeries by illumination color classification,” IEEE Transactions on Information Forensics and Security, 2013

2013
[21]

COVERAGE—A novel database for copy-move forgery detection,

B. Wen, Y . Zhu, R. Subramanian, T.-T. Ng, X. Shen, and S. Winkler, “COVERAGE—A novel database for copy-move forgery detection,” in Proceedings of the IEEE International Conference on Image Processing, 2016, pp. 161–165

2016
[22]

Evaluation of random field models in multi- modal unsupervised tampering localization,

P. Korus and J. Huang, “Evaluation of random field models in multi- modal unsupervised tampering localization,” in Proceedings of the IEEE International Workshop on Information Forensics and Security, 2016, pp. 1–6

2016
[23]

Fighting Fake News: Image splice detection via learned self-consistency,

M. Huh, A. Liu, A. Owens, and A. A. Efros, “Fighting Fake News: Image splice detection via learned self-consistency,” in Proceedings of the European Conference on Computer Vision , 2018, pp. 101–117

2018
[24]

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

2021
[25]

FaceForensics++: Learning to detect manipulated facial images,

A. Rossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner, “FaceForensics++: Learning to detect manipulated facial images,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 1–11

2019
[26]

Exploring multi-view pixel contrast for general and robust image forgery localization,

Z. Lou, G. Cao, K. Guo, H. Zhu, and L. Yu, “Exploring multi-view pixel contrast for general and robust image forgery localization,” IEEE Transactions on Information Forensics and Security , 2025

2025
[27]

Focal loss for dense object detection,

T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in 2017 IEEE International Conference on Computer Vision (ICCV) , 2017

2017
[28]

Robust image forgery detection against transmission over online social networks,

H. Wu, J. Zhou, J. Tian, J. Liu, and Y . Qiao, “Robust image forgery detection against transmission over online social networks,” IEEE Trans- actions on Information Forensics and Security , vol. 17, pp. 443–456, 2022

2022

Pith tools

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

[1] [1]

Lightweight and high-precision net- work for image copy-move forgery detection,

Y . Shi, S. Weng, L. Yu, and L. Li, “Lightweight and high-precision net- work for image copy-move forgery detection,” IEEE Signal Processing Letters, vol. 31, pp. 1409–1413, 2024

work page 2024

[2] [2]

Effective image tampering localization via enhanced transformer and co-attention fusion,

K. Guo, H. Zhu, and G. Cao, “Effective image tampering localization via enhanced transformer and co-attention fusion,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, 2024, pp. 4895–4899

work page 2024

[3] [3]

Effective image tampering localization with multi-scale convnext feature fusion,

H. Zhu, G. Cao, M. Zhao, H. Tian, and W. Lin, “Effective image tampering localization with multi-scale convnext feature fusion,”Journal of Visual Communication and Image Representation, vol. 98, p. 103981, 2024

work page 2024

[4] [4]

MVSS-Net: Multi- view multi-scale supervised networks for image manipulation detec- tion,

C. Dong, X. Chen, R. Hu, J. Cao, and X. Li, “MVSS-Net: Multi- view multi-scale supervised networks for image manipulation detec- tion,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 3, pp. 3539–3553, 2022

work page 2022

[5] [5]

Learning jpeg compression artifacts for image manipulation detection and localization,

M.-J. Kwon, S.-H. Nam, I.-J. Yu, H.-K. Lee, and C. Kim, “Learning jpeg compression artifacts for image manipulation detection and localization,” International Journal of Computer Vision , vol. 130, no. 8, pp. 1875– 1895, 2022

work page 2022

[6] [6]

PSCC-Net: Progressive spatio- channel correlation network for image manipulation detection and localization,

X. Liu, Y . Liu, J. Chen, and X. Liu, “PSCC-Net: Progressive spatio- channel correlation network for image manipulation detection and localization,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 11, pp. 7505–7517, 2022

2022

[7] [7]

Hierarchical fine-grained image forgery detection and localization,

X. Guo, X. Liu, Z. Ren, S. Grosz, I. Masi, and X. Liu, “Hierarchical fine-grained image forgery detection and localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

work page 2023

[8] [8]

TruFor: Leveraging all-round clues for trustworthy image forgery detection and localization,

F. Guillaro, D. Cozzolino, A. Sud, N. Dufour, and L. Verdoliva, “TruFor: Leveraging all-round clues for trustworthy image forgery detection and localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 20 606–20 615

work page 2023