HyperFake: Hyperspectral Reconstruction and Attention-Guided Analysis for Advanced Deepfake Detection

Pavan C Shekar; Pawan Soni; Vivek kanhangad

REVIEW 4 major objections 6 minor 11 references

HyperFake: Hyperspectral Reconstruction and Attention-Guided Analysis for Advanced Deepfake Detection

T0 review · 4 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 HyperFake claims that reconstructing 31-channel hyperspectral data from RGB video, then weighting the spectral bands by learned attention, exposes deepfake artifacts that RGB-only classifiers miss, reaching 92% validation accuracy.

desk verdict The spectral-reconstruction claim is unsupported as stated, but the pipeline is a concrete first attempt that deserves a proper ablation study. read the letter →

arxiv 2505.18587 v1 pith:C64EG6CP submitted 2025-05-24 cs.CV cs.AI

Pavan C Shekar , Pawan Soni , Vivek Kanhangad This is my paper

classification cs.CVcs.AI

keywords deepfakedetectionhyperspectralreconstructionspectralattentionMST++transformerEfficientNetRGB-to-hyperspectralmappingmediaforensicsfacemanipulation

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 tries to establish that a deepfake detector can gain a decisive advantage by first reconstructing 31-channel hyperspectral images from ordinary RGB video frames and then classifying with a spectral-attention-guided EfficientNet. It claims that artifacts invisible in RGB become detectable in the reconstructed spectral representation, so the method reaches 92% validation accuracy on a standard face-manipulation benchmark, well above the reported RGB baselines of 63.75% for ResNet-50 and 71.75% for EfficientNet-B7. If this is right, hyperspectral-style forensic analysis becomes available without hyperspectral cameras. The paper itself presents the results as preliminary and limits the evaluation to a single benchmark.

What carries the argument

The load-bearing object is the reconstructed hyperspectral cube $H = R(I_{\text{RGB}})$ produced by an improved MST++ model, together with the spectral attention reduction $F_{\text{reduced}} = \sum_{i=1}^{31} \alpha_i H_i$. MST++ is a multi-stage spectral-wise transformer for estimating 31-band hyperspectral images from RGB; the paper's FlexiAttention variant downsamples the spectral features, applies attention, and upsamples them, raising reported reconstruction PSNR from 34.32 dB to 34.8 dB. The learned weights $\alpha_i$ are intended to emphasize bands that carry deepfake artifacts while compressing the cube into the 3-channel input EfficientNet-B0 expects. Training uses BCEWithLogitsLoss with Adam and cosine learning-rate decay.

What would settle it

Train a classifier of comparable capacity directly on the same RGB frames with the same training protocol and evaluation split; if it reaches or exceeds 92% validation accuracy, the hyperspectral reconstruction is adding no information. The claimed advantage should also disappear if the reconstruction branch is replaced by a fixed random nonlinear projection to 31 channels.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a deterministic RGB-to-hyperspectral reconstruction, followed by learned spectral attention, exposes manipulation traces that conventional RGB classifiers miss. Concretely: an improved MST++ transformer converts each RGB frame into a hyperspectral cube $H = R(I_{\text{RGB}}) \in \mathbb{R}^{B \times 31 \times H \times W}$; a spectral attention module forms a learned weighted sum of the 31 bands, $F_{\text{reduced}} = \sum_{i=1}^{31} \alpha_i H_i$, collapsing them to 3 channels; and an EfficientNet-B0 classifier labels the result real or fake. On a standard face-manipulation benchmark, the HyperFake pipeline reports 98.94% training accuracy and 92% validation accuracy, against 63.75% and 71.75% for the two RGB baselines. The paper presents this as the first hyperspectral-reconstruction-based deepfake detection pipeline and as evidence that estimated spectral information, not just RGB, helps detection and generalization.

Load-bearing premise

The load-bearing premise is that a deterministic mapping trained on natural RGB images and their spectral counterparts exposes manipulation traces that a similarly capable classifier could not already recover from the original RGB pixels.

Editorial extensions

If this is right

If the central claim is correct, ordinary RGB video can be screened for spectral-style manipulation traces without any hyperspectral camera hardware.
The reported gap between 92% validation accuracy and the 63.75% and 71.75% baselines implies that the learned 31-to-3-channel spectral projection is a useful inductive bias for face-forensics classification.
The claimed reconstruction PSNR improvement from 34.32 dB to 34.8 dB is presented as making deepfake artifacts more distinguishable while FlexiAttention also speeds inference.
Because the evaluation uses a single benchmark, the paper's broader generalization claims depend on the planned cross-dataset tests that it says are future work.

Reading between the lines

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

Because Eq. (1) defines the hyperspectral cube as a learned function of the same RGB input, the 'hidden traces' are manufactured by the mapping; a fair test is to compare against an RGB classifier of comparable capacity trained with the same protocol.
The reconstruction network is trained on natural-image spectral pairs, while the detector sees faces; if the useful signal is face-specific, the mapping may need face-domain fine-tuning to transfer across datasets and manipulation styles.
A direct way to isolate the attention mechanism's role is to feed all 31 reconstructed bands into a classifier without the 3-channel projection and compare accuracy; that would show whether band selection or dimensionality reduction is what helps.

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

Colleague, here's my read on arXiv:2505.18587. The headline: the paper's central claim is not supported as stated, but there's a real idea underneath that deserves a sharper test.

What's new: this is the first paper I've seen that applies RGB-to-hyperspectral reconstruction to deepfake detection. The pipeline is concrete: freeze a MST++ trained on ARAD 1K, run RGB through it to get 31 channels, take a learned weighted sum of those channels to get back to 3, then feed that into EfficientNet-B0. They report 92% validation accuracy on FaceForensics++, which beats their two baselines by a wide margin. That's a new empirical result.

What the paper does well: the writing is clear, the method section is easy to follow, and the authors are honest that this is preliminary. Freezing the reconstruction network is a sensible choice that avoids overfitting the reconstruction to the deepfake task. They also report a small PSNR gain from their FlexiAttention change.

The soft spots are serious. The conceptual claim that hyperspectral reconstruction 'reveals manipulation artifacts invisible in conventional RGB space' is wrong as stated. Eq. (1) defines H = R(IRGB), so every spectral channel is a deterministic function of the RGB input. The spectral attention output in Eq. (3) is a learned weighted sum of those channels. The whole preprocessing chain is therefore a learned 3-channel transformation of the original RGB image. No deterministic transform can add information that wasn't already in the input; the data processing inequality rules that out. So the claim of 'invisible' traces is not grounded.

That said, there's a weaker version that could still hold: the frozen ARAD-pretrained transform might recode RGB into a representation that is easier for the classifier to exploit. That is a legitimate and testable hypothesis. The problem is the experiments don't test it. The baselines are only ResNet-50 and EfficientNet-B7, trained with no apparent protocol match. There is no control that feeds RGB directly to the same EfficientNet-B0, no ablation where MST++ is replaced with a random or identity transform, and no ablation removing the spectral attention. Without those, the 92% accuracy is uninterpretable as evidence for the spectral approach; it could just be the classifier architecture, initialization, or training details.

The metrics are also thin: they list EER and AUC in Section V but only report train/val accuracy in Table I. No code or data is released. A couple of references, [5] and [6], have 'Authors' in place of author names, which looks unfinished.

Who is this for? Researchers working on deepfake detection who are curious about whether fixed spectral transforms can help as a preprocessing step. The idea is worth a follow-up with proper controls, but this paper doesn't yet give us the evidence. I would send it to peer review – the question is legitimate and the authors need a referee to push them to run the right ablations. A serious editor should not desk-reject it. My own verdict would be reject or major revision until the ablations are done.

Referee Report

4 major / 6 minor

Summary. This manuscript introduces HyperFake, a deepfake detection pipeline that reconstructs 31-channel hyperspectral data from RGB frames using an 'improved MST++' model, then applies a spectral attention module that reduces the 31 channels to 3 via a learned weighted sum, and finally classifies the result with EfficientNet-B0. The reconstruction network is trained on ARAD 1K and frozen, while the classifier is trained on FaceForensics++, where the authors report 98.94% training accuracy and 92% validation accuracy, compared with validation accuracies of 63.75% for ResNet-50 and 71.75% for EfficientNet-B7. The paper claims to be the first hyperspectral-reconstruction-based deepfake detector and states that the reconstructed spectral data exposes manipulation artifacts invisible in conventional RGB space.

Significance. If the central claims were supported, the paper would open a practically attractive direction: obtaining hyperspectral-style analysis without the cost of hyperspectral cameras. The authors are transparent that the results are preliminary, they define standard metrics such as EER and AUC-ROC, and they identify future cross-dataset evaluation as necessary. However, the central conceptual claim is not derivable from the method as described. Because Eq. (1) defines the hyperspectral representation H as a deterministic function R(IRGB) of the RGB input, the attention-weighted combination in Eq. (3) is itself a deterministic transformation of RGB; no information absent from the RGB input can be introduced by the reconstruction. The reported accuracy advantage is therefore uninterpretable as evidence of spectral information gain without controlled ablations. The novelty of applying learned spectral reconstruction to deepfake detection is worth acknowledging, but the current manuscript does not establish that the reconstruction is the source of the reported improvement.

major comments (4)

[Secs. I, II-A, IV-B, Eq. (1)] The claim that reconstructed 31-channel hyperspectral data unveils manipulation artifacts invisible in conventional RGB space is not supported by the architecture. Equation (1) sets H = R(IRGB), so every spectral band is a deterministic function of the RGB input, and Eq. (3) forms the classifier input as a learned weighted sum of these bands. The entire preprocessing chain is thus a deterministic three-channel transformation of IRGB followed by EfficientNet, and by the data processing inequality it cannot literally contain information absent from the RGB input. The paper should either replace the information-gain claim with a testable reparameterization claim or provide controlled experiments that substantiate the weaker claim.
[Table I and Sec. VI] No ablation isolates the contribution of the reconstruction stage. The comparison in Table I is between HyperFake and two unrelated baselines (ResNet-50 and EfficientNet-B7) without matching the training protocol, optimizer, epochs, augmentations, input resolution, or data split. A control that feeds the original RGB frames to the same EfficientNet-B0 head with the same training recipe, and a control that replaces the trained MST++ with a random or identity transformation, are necessary to attribute the 92% validation accuracy to the spectral reconstruction. Without such controls, the improvement could arise from architecture capacity, initialization, or training details.
[Secs. V-B, VII, Table I] The evaluation is too sparse to support the paper's generalization claims. Although EER and AUC-ROC are defined in Sec. V-B, Table I reports only training and validation accuracy. The FaceForensics++ protocol is underspecified: the paper does not state which compression level (raw, c23, or c40), which manipulation types, whether classification is frame-level or video-level, or the train/validation split. The authors themselves state in Sec. VII that 'this work presents preliminary results' and relegate DFDC, Celeb-DF, and cross-dataset evaluation to future work. The reported 92% validation accuracy is therefore not interpretable as evidence for robust and generalizable deepfake detection.
[Sec. IV-B.1] The claimed PSNR improvement from 34.32 dB to 34.8 dB and the inference-speed improvement from FlexiAttention are not accompanied by any experimental setup, dataset, or runtime measurement. Since ARAD 1K is a natural-image spectral reconstruction benchmark, the paper also does not show that the frozen reconstruction produces meaningful spectral estimates on manipulated face images, which is essential for the pipeline's rationale.

minor comments (6)

[Secs. I and IV] There are formatting artifacts in the section headings: 'M OTIVATION' and 'M ETHODOLOGY' should read 'MOTIVATION' and 'METHODOLOGY'.
[References [5] and [6]] References [5] and [6] list 'Authors' as the author name; complete bibliographic entries are needed.
[Sec. IV-C, Eqs. (2)-(3)] The spectral attention mechanism is introduced through the generic self-attention formula in Eq. (2), but the text does not explain how the attention output produces the scalar weights alpha_i in Eq. (3); this connection should be made explicit.
[Secs. IV-A and IV-E] The abstract and Sec. IV-A mention 'spectral-aware recalibration layers' in the classification module, but no such layers are defined or described in Sec. IV-E.
[Fig. 1 caption and Sec. IV-E] The figure caption uses 'EfficientNetB0' while the text uses 'EfficientNet-B0'; please standardize the notation.
[Sec. V-A] The description of FaceForensics++ is imprecise: the dataset contains 1,000 original video sequences with multiple manipulation methods, not simply '1,000 real videos and their manipulated counterparts'; please specify the exact subset, compression, and split used.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild conceptual circularity in the 'RGB-invisible traces' claim; experimental evaluation is not circular.

self definitional [Abstract; Sec. I (Motivation); Sec. IV-B, Eq. (1)]
"By reconstructing 31-channel hyperspectral data using an improved MST++ model, HyperFake unveils manipulation artifacts invisible in conventional RGB space. ... H = R(IRGB) ∈ R^{B×31×H×W} (1)"

Equation (1) defines the reconstructed hyperspectral tensor H as a deterministic function of the RGB input IRGB. Therefore every 'spectral' channel, and any artifact the paper claims to unveil, is by construction a function of the same RGB values. The claim that these artifacts are 'invisible in conventional RGB space' is only true in the weak sense that standard classifiers do not use this particular transformation; it is not true that H supplies information absent from RGB. Attributing the classification gain to 'hyperspectral information' is thus a self-definitional explanation: the new representation is generated from the input it is said to reveal.

full rationale

The reported 92% validation accuracy is a holdout measurement, so the experimental claim is not a fitted artifact. MST++ is frozen after training on the external ARAD 1K dataset, and the attention weights are learned on the FF++ training split, which is standard supervised learning rather than circular prediction. The circularity is confined to the interpretive claim: since Eq. (1) defines H as a deterministic function of IRGB, the 'hidden spectral traces' are constructed from the RGB input, and the paper's assertion that they are 'invisible in conventional RGB space' is an over-statement. A properly controlled comparison (same head, same training protocol, with and without MST++/spectral attention) would be needed to support the spectral explanation; its absence is a correctness and ablation concern, not circularity. No load-bearing self-citation chain or uniqueness import was found.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the assumption that a learned RGB-to-hyperspectral mapping provides new, useful information for deepfake detection. The attention weights are a small set of fitted parameters, while the large network parameters are treated as part of the method. No new physical entities are introduced.

free parameters (3)

Learned attention weights alpha_i (Eq. 3) = Not reported
The attention weights in the spectral attention module are learned on the FaceForensics++ training data, and their values are not provided. The final 3-channel representation is a weighted sum of the 31 reconstructed channels, so these weights directly control the input to the classifier.
MST++ reconstruction network parameters = Not reported
The improved MST++ model is trained on the ARAD 1K dataset and then frozen. These millions of learned parameters define the deterministic mapping from RGB to 31 spectral channels, and the central claim depends on this mapping being meaningful for face images.
FlexiAttention downsampling ratio = Not reported
The FlexiAttention mechanism is said to downsample spectral features, apply attention, and upsample. No exact downsampling factor or other configuration is given, so this is an unspecified design choice.

assumptions (3)

domain assumption A model trained on ARAD 1K natural images can produce meaningful hyperspectral reconstructions for face images in deepfake datasets.
The MST++ model is trained on ARAD 1K, a dataset of natural scenes, and then directly applied to face crops from FaceForensics++. This transferability is assumed without any validation.
domain assumption The reconstructed 31 spectral channels contain discriminative information for deepfake detection that is not trivially available in the RGB input.
This is the core premise of the paper and is stated in Sections I and II. Since the reconstruction is deterministic from RGB, the assumption is that the nonlinear mapping produces features that make manipulative traces more separable, which is never tested.
domain assumption The FaceForensics++ train/validation split used in the experiments is unbiased and standard.
The paper does not describe the split; it only says it uses the FaceForensics++ dataset. The unusually low baseline accuracies suggest the split or preprocessing may differ from common practice, potentially affecting the fairness of the comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HyperFake: Hyperspectral Reconstruction and Attention-Guided Analysis for Advanced Deepfake Detection." pith.science (2026). https://pith.science/paper/C64EG6CP

@misc{pith2026250518587,
  author       = {Pith},
  title        = {Pith review of: HyperFake: Hyperspectral Reconstruction and Attention-Guided Analysis for Advanced Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C64EG6CP}},
  note         = {Machine review of arXiv:2505.18587}
}

read the original abstract

Deepfakes pose a significant threat to digital media security, with current detection methods struggling to generalize across different manipulation techniques and datasets. While recent approaches combine CNN-based architectures with Vision Transformers or leverage multi-modal learning, they remain limited by the inherent constraints of RGB data. We introduce HyperFake, a novel deepfake detection pipeline that reconstructs 31-channel hyperspectral data from standard RGB videos, revealing hidden manipulation traces invisible to conventional methods. Using an improved MST++ architecture, HyperFake enhances hyperspectral reconstruction, while a spectral attention mechanism selects the most critical spectral features for deepfake detection. The refined spectral data is then processed by an EfficientNet-based classifier optimized for spectral analysis, enabling more accurate and generalizable detection across different deepfake styles and datasets, all without the need for expensive hyperspectral cameras. To the best of our knowledge, this is the first approach to leverage hyperspectral imaging reconstruction for deepfake detection, opening new possibilities for detecting increasingly sophisticated manipulations.

Figures

Figures reproduced from arXiv: 2505.18587 by the authors.

**Figure 1.** HyperFake pipeline architecture for Deepfake detection. RGB input is processed by MST++ to generate a 31-channel hyperspectral image. The spectral [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Spectral Attention Module: Transforming 31-Channel Hyperspectral [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Comparison of real and fake images with hyperspectral reconstructions.The first column shows real images, the second shows fake images, and the last [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 6 canonical work pages

[5]

Blockwise spectral analysis for deepfake detection in high-fidelity videos,

Authors, “Blockwise spectral analysis for deepfake detection in high-fidelity videos,” in Proceedings of the IEEE Conference , 2023. [Online]. Available: https://ieeexplore.ieee.org/document/10032370

work page arXiv 2023
[6]

A review of deepfake techniques: Architecture, detection, and datasets,

——, “A review of deepfake techniques: Architecture, detection, and datasets,” IEEE Access , 2022. [Online]. Available: https://ieeexplore.ieee.org/document/10711187

work page arXiv 2022
[1]

Deepfake detection based on discrepancies between faces and their context,

Y . Nirkin, L. Wolf, Y . Keller, and T. Hassner, “Deepfake detection based on discrepancies between faces and their context,” arXiv preprint arXiv:2008.12262 , 2020. [Online]. Available: https://arxiv.org/abs/2008.12262

arXiv 2008
[2]

Combining efficientnet and vision transformers for video deepfake detection,

D. Coccomini, N. Messina, C. Gennaro, and F. Falchi, “Combining efficientnet and vision transformers for video deepfake detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022. [Online]. Avail- able: https://github.com/davide-coccomini/Combining-EfficientNet-and- Vision-Transformers-for-Video-Deepfake-Detection

work page 2022
[3]

LAA-Net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection,

D. Nguyen, N. Mejri, I. P. Singh, P. Kuleshova, M. Astrid, A. Kacem, E. Ghorbel, and D. Aouada, “LAA-Net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection,” arXiv preprint arXiv:2401.13856, 2024. [Online]. Available: https://arxiv.org/abs/2401.13856

arXiv 2024
[4]

Fakeout: Leveraging out-of-domain self-supervision for multi-modal video deepfake detection,

G. Knafo and O. Fried, “Fakeout: Leveraging out-of-domain self-supervision for multi-modal video deepfake detection,” arXiv preprint arXiv:2212.00773 , 2022. [Online]. Available: https://arxiv.org/abs/2212.00773

arXiv 2022
[8]

MST++: Multi-stage spectral-wise transformer for efficient spectral reconstruction,

Y . Cai, J. Lin, Z. Lin, H. Wang, Y . Zhang, H. Pfister, R. Timofte, and L. V . Gool, “MST++: Multi-stage spectral-wise transformer for efficient spectral reconstruction,” in Proceedings of the NTIRE Spectral Reconstruction Challenge, 2022

work page 2022
[9]

Ntire 2022 spectral recovery challenge and data set,

B. Arad, R. Timofte, R. Yahel, N. Morag, A. Bernat, Y . Cai, J. Lin, Z. Lin, H. Wang, Y . Zhang, H. Pfister, L. Van Gool, S. Liu, Y . Li, C. Feng, L. Lei, J. Li, S. Du, C. Wu, Y . Leng, R. Song, M. Zhang, C. Song, S. Zhao, Z. Lang, W. Wei, L. Zhang, R. Dian, T. Shan, A. Guo, C. Feng, J. Liu, M. Agarla, S. Bianco, M. Buzzelli, L. Celona, R. Schettini, J. H...

work page 2022

Show all 11 references

[10]

Faceforensics++: Learning to detect manipulated facial images,

A. R ¨ossler, 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
[11]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2016, pp. 770–778. [Online]. Available: https://arxiv.org/abs/1512.03385

2016 arXiv
[12]

Efficientnet: Rethinking model scaling for convolutional neural networks,

M. Tan and Q. V . Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” in Proceedings of the 36th International Conference on Machine Learning , 2019, pp. 6105–6114. [Online]. Available: https://arxiv.org/abs/1905.11946

2019 arXiv

Pith tools

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

[1] [5]

Blockwise spectral analysis for deepfake detection in high-fidelity videos,

Authors, “Blockwise spectral analysis for deepfake detection in high-fidelity videos,” in Proceedings of the IEEE Conference , 2023. [Online]. Available: https://ieeexplore.ieee.org/document/10032370

work page arXiv 2023

[2] [6]

A review of deepfake techniques: Architecture, detection, and datasets,

——, “A review of deepfake techniques: Architecture, detection, and datasets,” IEEE Access , 2022. [Online]. Available: https://ieeexplore.ieee.org/document/10711187

work page arXiv 2022

[3] [1]

Deepfake detection based on discrepancies between faces and their context,

Y . Nirkin, L. Wolf, Y . Keller, and T. Hassner, “Deepfake detection based on discrepancies between faces and their context,” arXiv preprint arXiv:2008.12262 , 2020. [Online]. Available: https://arxiv.org/abs/2008.12262

arXiv 2008

[4] [2]

Combining efficientnet and vision transformers for video deepfake detection,

D. Coccomini, N. Messina, C. Gennaro, and F. Falchi, “Combining efficientnet and vision transformers for video deepfake detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022. [Online]. Avail- able: https://github.com/davide-coccomini/Combining-EfficientNet-and- Vision-Transformers-for-Video-Deepfake-Detection

work page 2022

[5] [3]

LAA-Net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection,

D. Nguyen, N. Mejri, I. P. Singh, P. Kuleshova, M. Astrid, A. Kacem, E. Ghorbel, and D. Aouada, “LAA-Net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection,” arXiv preprint arXiv:2401.13856, 2024. [Online]. Available: https://arxiv.org/abs/2401.13856

arXiv 2024

[6] [4]

Fakeout: Leveraging out-of-domain self-supervision for multi-modal video deepfake detection,

G. Knafo and O. Fried, “Fakeout: Leveraging out-of-domain self-supervision for multi-modal video deepfake detection,” arXiv preprint arXiv:2212.00773 , 2022. [Online]. Available: https://arxiv.org/abs/2212.00773

arXiv 2022

[7] [8]

MST++: Multi-stage spectral-wise transformer for efficient spectral reconstruction,

Y . Cai, J. Lin, Z. Lin, H. Wang, Y . Zhang, H. Pfister, R. Timofte, and L. V . Gool, “MST++: Multi-stage spectral-wise transformer for efficient spectral reconstruction,” in Proceedings of the NTIRE Spectral Reconstruction Challenge, 2022

work page 2022

[8] [9]

Ntire 2022 spectral recovery challenge and data set,

B. Arad, R. Timofte, R. Yahel, N. Morag, A. Bernat, Y . Cai, J. Lin, Z. Lin, H. Wang, Y . Zhang, H. Pfister, L. Van Gool, S. Liu, Y . Li, C. Feng, L. Lei, J. Li, S. Du, C. Wu, Y . Leng, R. Song, M. Zhang, C. Song, S. Zhao, Z. Lang, W. Wei, L. Zhang, R. Dian, T. Shan, A. Guo, C. Feng, J. Liu, M. Agarla, S. Bianco, M. Buzzelli, L. Celona, R. Schettini, J. H...

work page 2022