Pith. sign in

REVIEW 5 major objections 6 minor 41 references

Enhancing Abnormality Identification: Robust Out-of-Distribution Strategies for Deepfake Detection

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Deepfake detection reframed as out-of-distribution recognition: a two-module pipeline flags unfamiliar generators by combining reconstruction residual, latent encoding, and softmax confidence.

desk verdict Useful engineering study with a new benchmark split, but the synthetic-only Abnormality module is near or below the softmax baseline on the scenarios that actually test unseen generators, so the central open-set claim is unsupported. read the letter →

arxiv 2506.02857 v1 pith:NOUJRRAP submitted 2025-06-03 cs.CV

classification cs.CV
keywords deepfakedetectionout-of-distributionanomalyreconstructionresidualattentionmapU-NetDeiTtransformeropen-setrecognition
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

Deepfake detectors trained on known generators often fail on new generators, so the paper recasts detection as an out-of-distribution problem: the model should not only say real or fake but also flag that an input is unlike anything it trained on. The proposed pipeline has two cooperating modules: an in-distribution classifier that also reconstructs its input, and an Abnormality module that turns the reconstruction residual, the latent encoding, and the softmax confidence vector into an OOD risk score. Two variants are tested, one using a U-Net that reconstructs pixels and one using a DeiT Transformer whose attention heatmaps are reconstructed by an autoencoder. On the adapted CDDB benchmark, the authors report that the V2 and V3 Abnormality modules beat the MSP and ODIN baselines in the Content scenario, and that training with real outliers yields near-perfect AUROC, while synthetic-only training is more practical but weaker in the hardest Group scenario. If the claim holds, the practical payoff is a built-in retraining signal for deployed detectors as new generative models appear.

What carries the argument

The load-bearing mechanism is the Abnormality module, a compact MLP that consumes a stack of three signals per image: the softmax probability vector from the deepfake classifier, the encoder's latent embedding, and a squared reconstruction residual. The residual is image-space for the U-Net variant and attention-map-space for the DeiT variant, where a variational autoencoder reconstructs the heatmap. Three versions of the module (V1, V2, V3) rebalance the dimensionality of these signals; V3 compresses both the encoding and the residual branches to ten elements before a final single-layer risk head, so the softmax vector is no longer swamped. The module's job is to map how badly the model fails to reconstruct or represent an input into an OOD risk score, with training done on synthetic alterations, real outliers, or a mix of both.

What would settle it

Measure per-generator mean squared residuals from the trained U-Net and attention-map autoencoder on held-out generators: if any unseen generator's residual distribution overlaps or sits below the in-distribution residual distribution, the OOD signal fails. The paper's own Group scenario is near that boundary, with synthetic-only V2S at AUROC 63.35 versus MSP at 62.06.

Watch

Extended reading notes

Core claim

The core claim is that a network trained with a weighted loss that combines binary cross-entropy for real/fake classification and a reconstruction loss can serve both as a deepfake classifier and as an OOD detector. At inference, three auxiliary signals are pushed into the Abnormality module: the squared reconstruction residual, the encoder's latent representation, and the softmax probability vector. The CNN variant computes the residual in image space from a U-Net Scorer; the Transformer variant computes it in attention-map space by reconstructing DeiT heatmaps with an autoencoder. The authors report that on their CDDB partitioning the V2 and V3 modules surpass MSP and ODIN in the Content scenario (V2S AUROC 91.6 versus 77.85 for MSP), that real-outlier training approaches 100 AUROC there, that the Group scenario remains hard for synthetic-only training (V2S 63.35 AUROC, about tied with MSP at 62.06), and that on the CIFAR10 OOD benchmark the mixed-training V2M module reaches far-OOD AUROC values of 99.24 on MNIST, 98.06 on SVHN, and 99.43 on Textures with 97.5% in-distribution accuracy.

Load-bearing premise

The load-bearing premise is that images produced by generators or content types the model never saw produce larger reconstruction (or attention-map) residuals than in-distribution images, so residual size can serve as an OOD signal.

Editorial extensions

If this is right

  • A deployed deepfake detector built this way can emit an OOD alarm alongside its real/fake decision, giving an explicit retraining cue when a new generator appears.
  • Combining softmax probability, latent encoding, and reconstruction residual is the load-bearing design; the V3 variant that compresses all branches to ten dimensions shows that rebalancing these signals is enough to beat the MSP baseline in the reported settings.
  • Attention-map reconstruction with a tiny DeiT reaches Content-scenario OOD results comparable to the CNN approach while using roughly 57% less memory (1,118 MB versus 2,760 MB).
  • When real outlier images are available, the method saturates the Content scenario (AUROC above 99.9), indicating that most of the remaining gap in practical settings is the quality and relevance of the OOD training data.
  • On the CIFAR10 OOD benchmark, the mixed-training V2M module reaches far-OOD AUROC values of 99.24 (MNIST), 98.06 (SVHN), and 99.43 (Textures) at 97.5% ID accuracy, placing it among the top trained-with-outlier techniques in the reported tables.

Reading between the lines

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

  • My inference, not the paper's: the same Abnormality stack could transfer to other media-forensics tasks such as splicing detection or GAN attribution, because the residual-plus-encoding-plus-confidence recipe does not depend on the deepfake label structure.
  • My reading: the large gap between Content and Group scenarios suggests the method is most sensitive to semantic content shifts (faces versus non-faces) and less sensitive to generator novelty; on the Group scenario, synthetic-only training performs at the MSP baseline, so the open-set benefit in that setting is not yet established.
  • A testable extension the paper leaves implicit is to replace the hand-crafted synthetic corruptions (blur, JPEG, noise, rotation) with perturbations that mimic generator artifacts, or with GAN-generated outliers; that could close the Group-scenario gap while staying in the unsupervised regime.
  • The reported failure of the attention-map variant on Group/Mix scenarios, attributed to uniform attention maps on varied content, implies that the Transformer route needs a way to keep attention maps discriminative when in-distribution content is diverse; this is an inference, not a paper conclusion.
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.

Referee Report

5 major / 6 minor

Summary. The paper proposes two OOD detection approaches for deepfake detection: a CNN-based U-Net that jointly classifies and reconstructs images, and a Transformer-based DeiT that reconstructs attention heatmaps. Three Abnormality module variants (V1, V2, V3) fuse the reconstruction residual, the bottleneck/CLS encoding, and the softmax probability vector via MLPs to produce an OOD risk score. The modules are trained with synthetic-only, real-only, or mixed OOD data. Evaluation is carried out on a custom split of the CDDB dataset into Content, Group, and Mix scenarios, and on the CIFAR10 OpenOOD benchmark. The paper claims that the approach achieves promising results and ranks among the top-performing configurations.

Significance. The paper introduces a useful task formulation and a new benchmark split of CDDB into Content, Group, and Mix OOD scenarios, which could serve as a resource for future open-set deepfake detection research. The systematic comparison of synthetic versus real outlier exposure is also informative. However, the central claim that the Abnormality modules provide robust OOD detection for unseen deepfake generators is not supported by the most direct experiments. The Group and Mix scenarios, which are the only ones that test OOD samples from unseen generators, show synthetic-only performance essentially at the MSP baseline (or below), while the high Content-scenario scores are confounded by semantic face/non-face separation. The paper is transparent about the Transformer's failure on Group/Mix, but the abstract and conclusion overstate the method's effectiveness. The CIFAR10 results further show that V2S is not competitive on near-OOD detection, undermining the 'state-of-the-art' phrasing.

major comments (5)
  1. [§5.3, Table 11] The Group and Mix scenarios are the only experiments that directly test OOD samples from unseen deepfake generators, yet V2S achieves AUROC 63.35 on Group and 60.14 on Mix, compared with MSP baselines of 62.06 and 69.74, respectively. The 1.3-point gap on Group is within any reasonable noise given that no error bars or multiple runs are reported, and on Mix the proposed method is worse than the simple softmax baseline. These results do not support the abstract's claim that the method 'achieves promising results' or the conclusion's claim of 'remarkable efficacy' for open-set detection. At minimum, the claims must be softened and the discrepancy with the Content scenario explicitly discussed.
  2. [§5.1, §5.3, Table 10] The high Content-scenario score of V2S (AUROC 91.6) is not evidence for the paper's core premise that reconstruction residuals separate unseen generators. The Content split separates faces from non-faces after removing content-overlapping samples, so the OOD set is semantically distinct from the ID set. The model likely learns this low-level content cue rather than any generator-agnostic artifact. This does not test the statement in Section 3 that 'we expect the reconstruction error in the OOD module to increase when inputs come from unknown distributions.' The authors should either run a controlled experiment where ID and OOD contain the same semantic classes (as in Group/Mix) or explicitly state that their method targets content shift rather than generator novelty.
  3. [§4.4, §5.3] The V2R and V2M variants train on real OOD samples drawn from the same benchmark distribution that is later used for testing. For example, in the Group scenario, non-GAN images serve as both training outliers and test OOD. This is standard outlier exposure with test-distribution leakage, so the near-perfect AUROC values (e.g., 92.72 on Group, 100 on Content) are expected and do not validate an independent OOD detector. The paper should explicitly acknowledge this leakage and avoid presenting V2M/V2R as evidence that the Abnormality module generalizes to unknown distributions; these results only show that the module can be a strong supervised classifier when the outlier distribution is known.
  4. [§4.4] The synthetic-only variant is repeatedly called 'unsupervised' (e.g., Section 5.3: 'providing an interesting unsupervised approach'), but the Abnormality module is trained with labels indicating which inputs are altered ID images treated as OOD. This is supervised outlier exposure with automatically generated labels, not an unsupervised method. The terminology is misleading and should be corrected throughout, for example by calling it 'synthetic outlier exposure' or 'self-supervised labeling.'
  5. [§5.4, Tables 13-14] The CIFAR10 benchmark comparison is not properly controlled. The proposed ID classifier has 97.5% accuracy, while most compared baselines report around 95% ID accuracy; OOD detection metrics are sensitive to the quality of the ID classifier. Additionally, V2S achieves 69.07 AUROC on near-OOD CIFAR100, which is far below trained methods such as RotPred (91.19) and LogitNorm (90.95), contradicting the impression given by the abstract of top-ranked performance. The authors should either use a common ID classifier across methods or clearly label that the comparison is not ID-matched, and they should temper the 'state-of-the-art' claims accordingly.
minor comments (6)
  1. [§1] In the Introduction, 'and-hoc pre-processing' should be 'ad-hoc pre-processing.'
  2. [§4, §5.3] Several typos appear: 'Hendrycks and Gimpe' should be 'Hendrycks and Gimpel'; in Table 10, the FPR95 value '08e-4' should be '8e-4'; in the header of Table 12, the text says 'V1 and V2' but the table lists V2 and V3.
  3. [§4.3, Table 4] The text states that the V3 MLP blocks have 'a consistent design featuring five fully connected layers,' but Table 4 shows six FC layers (FC R1-R6 and FC E1-E6) before the concatenation. Please reconcile this inconsistency.
  4. [§4.3] The sentence 'These 3 abnormality module versions are all valid for the CNN approach...' is followed later by the statement that for the Transformer approach only V2 and V3 are appropriate. This is clear, but the earlier sentence could be read as implying all versions work for both. Consider rewording.
  5. [§4.4, Figure 5] The caption of Figure 5 lists 'normal noise distorted, normal noise, and random rotation' but the sequence appears to include overlapping categories. Please clarify which distortions are actually shown.
  6. [§4.4, §5.2] The paper does not report hyperparameters for training the Abnormality modules (learning rate, optimizer, number of epochs, batch size, or loss weights beyond the BCE loss). These details are necessary for reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

Real-outlier variants are fit on the same OOD distributions they are later scored on, so their top AUROC is a fit; the synthetic-only runs that actually test unseen deepfake generators are near baseline.

  1. fitted input called prediction [Section 5.4 (CIFAR10 benchmark, Tables 13-14); training setup in Section 4.4]
    "Real OOD are data sampled from near OOD datasets as outliers. Thus, V2S belongs to the second block, while V2R is in the third one."

    The paper trains V2M with real outliers sampled from the near-OOD datasets and then reports AUROC on the same near-OOD datasets (CIFAR-100 and TinyImageNet) in Table 14. Because the training outliers and the test OOD are drawn from the same two datasets, the Abnormality module is not detecting unseen OOD; it is fitting those distributions. The 93.36/99.66 AUROC values are therefore in-distribution accuracy for the second-stage classifier, not a prediction. The same pattern appears in the CDDB Content scenario, where 'available OOD samples' from the benchmark being tested are used for supervised training and produce 99.9-100 AUROC.

full rationale

The paper's derivation chain is mostly self-contained: the Abnormality module is a supervised MLP over the ID module's residual, encoding, and softmax output, and the reconstruction-based premise is directly probed by the synthetic-only (V2S) runs and by the external CIFAR far-OOD numbers. The self-references [1,2] are contextual and are not load-bearing. The main circular step is the real-outlier (V2R/V2M) evaluation: training outliers are drawn from the same near-OOD datasets that Table 14 scores as test OOD, and in the CDDB Content scenario 'available OOD samples' from the same benchmark split are used for training, yielding near-perfect AUROC. Those numbers reduce by construction because the training labels and test labels come from the same distribution, making them outlier-exposure fits rather than predictions of unseen OOD. Separately, the Group/Mix V2S results (63.35/60.14 vs MSP 62.06/69.74) undercut the central OOD premise, but that is a support or correctness problem rather than circularity; similarly, the synthetic-only training is supervised on self-generated labels despite being called 'unsupervised,' which is a terminology issue. Overall circularity is partial: it affects the real-outlier validation, while the synthetic-only and external-benchmark results retain independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The pipeline rests on hand-set loss weights, arbitrary synthetic OOD perturbations, and two empirical assumptions about reconstruction error and attention maps. No code or data splits are released, and the real-outlier variants depend on curated OOD examples.

free parameters (6)
  • Loss weight alpha = 0.9
    Eq. 1; hand-set to prioritize classification over reconstruction, affecting all CNN models.
  • Loss weight beta = 0.1
    Eq. 1; coupled with alpha=0.9, chosen by hand and not optimized.
  • JPEG quality for synthesized OOD = 10%
    Section 4.4; hand-chosen JPEG compression level used to create synthetic OOD training data.
  • Gaussian blur kernel = not specified
    Section 4.4 lists blurring as an OOD synthesis technique but gives no kernel size or sigma, leaving a free choice.
  • Noise standard deviation = random, unspecified
    Section 4.4; noise with random standard deviation is used to synthesize OOD, but the distribution is not quantified.
  • Rotation angle = pi/2
    Section 4.4; random image rotation by pi/2 is used as an OOD augmentation.
assumptions (5)
  • domain assumption The reconstruction residual of an autoencoder trained on ID data is higher for OOD than for ID inputs.
    Invoked throughout Section 4 as the core signal for the Abnormality module.
  • domain assumption Attention maps from a DeiT classifier trained on ID data are reconstructible for ID samples and poorly reconstructible for OOD samples.
    Section 3.2; adopted from Cultrera et al. [28] without independent verification on this dataset.
  • ad hoc to paper Removing content-overlapping samples from CDDB leaves a valid OOD benchmark.
    Section 5.1; this data exclusion directly inflates the separability of the Content scenario.
  • ad hoc to paper Synthetic distortions such as blur, JPEG, noise, and rotation simulate OOD data for deepfakes.
    Section 4.4; the Group scenario result, AUROC 63.35 for V2S, suggests this proxy is weak.
  • domain assumption Real OOD samples used in V2R and V2M training are representative of the OOD test distribution.
    Section 4.4 and Tables 10 to 12; this is outlier exposure, not a parameter-free guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Abnormality Identification: Robust Out-of-Distribution Strategies for Deepfake Detection." pith.science (2026). https://pith.science/paper/NOUJRRAP

@misc{pith2026250602857,
  author       = {Pith},
  title        = {Pith review of: Enhancing Abnormality Identification: Robust Out-of-Distribution Strategies for Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOUJRRAP}},
  note         = {Machine review of arXiv:2506.02857}
}
read the original abstract

Detecting deepfakes has become a critical challenge in Computer Vision and Artificial Intelligence. Despite significant progress in detection techniques, generalizing them to open-set scenarios continues to be a persistent difficulty. Neural networks are often trained on the closed-world assumption, but with new generative models constantly evolving, it is inevitable to encounter data generated by models that are not part of the training distribution. To address these challenges, in this paper, we propose two novel Out-Of-Distribution (OOD) detection approaches. The first approach is trained to reconstruct the input image, while the second incorporates an attention mechanism for detecting OODs. Our experiments validate the effectiveness of the proposed approaches compared to existing state-of-the-art techniques. Our method achieves promising results in deepfake detection and ranks among the top-performing configurations on the benchmark, demonstrating their potential for robust, adaptable solutions in dynamic, real-world applications.

Figures

Figures reproduced from arXiv: 2506.02857 by the authors.

Figure 1
Figure 1. Pipeline architecture composed by ID Module and Ab [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Attention heatmap from DeiT model. Unlike the U-Net-based solution, this approach trains the classifier and the reconstruction model separately. For the reconstruction process, two network architectures are proposed: the classic AutoEncoder (AE) and the Variational AutoEncoder (VAE). The AE is trained using the standard Mean Absolute Error (MAE) loss function. In contrast, the VAE employs a loss function that combin… view at source ↗
Figure 3
Figure 3. Abnormality module Encoder V1 and V2 scheme. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Abnormality module Encoder V3 scheme. Layer/Module Output Shape Branch Input 1 (B, 2) Softmax Input 3 (B, 3, 112, 112) Residual Encoder block 1 (B, 16, 56, 56) Encoder block 2 (B, 32, 28, 28) Encoder block 3 (B, 64, 14, 14) Flatten (B, 12544) FC R1 + BN + GELU (B, 4096…
Figure 5
Figure 5. Figure 5: Alteration techniques for synthesized OOD data. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages

  1. [1]

    Cont inuous fake media detection: Adapting deepfake detectors to new generative techniques

    Francesco Tassone, Luca Maiano, and Irene Amerini. Cont inuous fake media detection: Adapting deepfake detectors to new generative techniques. Computer Vision and Image Understanding , 249:104143, 2024

  2. [2]

    Deep learning for multimedia forensics

    Irene Amerini, Aris Anagnostopoulos, Luca Maiano, Lore nzo Ricciardi Celsi, et al. Deep learning for multimedia forensics. F oundations and Trends® in Computer Graphics and Vision, 12(4):309–457, 2021

  3. [3]

    Synthetic Image Verification in the Era of Generative AI: What Works and What Isn't There Yet

    Diangarti Tariang, Riccardo Corvi, Davide Cozzolino, G iovanni Poggi, Koki Nagano, and Luisa V erdoliva. Synthetic image verification in the era of generative ai: Wha t works and what isn’t there yet. arXiv preprint arXiv:2405.00196, 2024

  4. [4]

    Detecting gan-ge nerated imagery using color cues

    Scott McCloskey and Michael Albright. Detecting gan-ge nerated imagery using color cues. arXiv preprint arXiv:1812.08247, 2018

  5. [5]

    Attributing fake images to gans: Learning and analyzing gan finger- prints

    Ning Y u, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan finger- prints. In Proceedings of the IEEE/CVF international conference on co mputer vision, pages 7556–7566, 2019

  6. [6]

    Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR) , pages 506–511

    Francesco Marra, Diego Gragnaniello, Luisa V erdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR) , pages 506–511. IEEE, 2019

  7. [7]

    Deep learning based one-class detection system for fake faces generated by gan network

    Shengyin Li, Vibekananda Dutta, Xin He, and Takafumi Mat sumaru. Deep learning based one-class detection system for fake faces generated by gan network. Sensors, 22(20):7767, 2022

  8. [8]

    Fakespotter: A simple yet robust baseline for spotting ai-synthesized fake faces

    Run Wang, Felix Juefei-Xu, Lei Ma, Xiaofei Xie, Yihao Hua ng, Jian Wang, and Y ang Liu. Fakespotter: A simple yet robust baseline for spotting ai-synthesized fake faces . arXiv preprint arXiv:1909.06122 , 2019

Show all 41 references
  1. [9]

    Multimodal forgery detection using ensemble learning

    Ammarah Hashmi, Sahibzada Adil Shahzad, Wasim Ahmad, Ch ia Wen Lin, Y u Tsao, and Hsin-Min Wang. Multimodal forgery detection using ensemble learning. In 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) , pages 1524–1532....

  2. [10]

    Not made for each other-audio-visual dissonance-based deepfake detection and localization

    Komal Chugh, Parul Gupta, Abhinav Dhall, and Ramanatha n Subramanian. Not made for each other-audio-visual dissonance-based deepfake detection and localization. In Proceedings of the 28th ACM international conference on multimedia, pages 439–447, 2020

  3. [11]

    Emotions don’t lie: An audio-visual deepfake detection method using affective cues

    Trisha Mittal, Uttaran Bhattacharya, Rohan Chandra, A niket Bera, and Dinesh Manocha. Emotions don’t lie: An audio-visual deepfake detection method using affective cues. In Proceedings of the 28th ACM international conference on multimedia, pages 2823–2832, 2020

  4. [12]

    G eneralized out-of-distribution detection: A survey

    Jingkang Y ang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. G eneralized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334 , 2021

  5. [13]

    Multimodal forgery detection using ensemble learning

    Ammarah Hashmi, Sahibzada Adil Shahzad, Wasim Ahmad, C hia Wen Lin, Y u Tsao, and Hsin-Min Wang. Multimodal forgery detection using ensemble learning. In 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) , pages 1524–1532....

  6. [14]

    Fakelocator: Robust localization of gan-based face manipulations

    Yihao Huang, Felix Juefei-Xu, Qing Guo, Y ang Liu, and Ge guang Pu. Fakelocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information F orensics and Security, 17:2657–2672, 2022

  7. [15]

    Deepfake detection by analyzing convolutional traces

    Luca Guarnera, Oliver Giudice, and Sebastiano Battiat o. Deepfake detection by analyzing convolutional traces. In Proceedings of the IEEE/CVF conference on computer vision a nd pattern recognition workshops , pages 666– 667, 2020

  8. [16]

    Deepfake detection based on the discrepancy between the face and its context

    Y uval Nirkin, Lior Wolf, Y osi Keller, and Tal Hassner. Deepfake detection based on the discrepancy between the face and its context. arXiv preprint arXiv:2008.12262 , 2020

  9. [17]

    Deepfake video dete ction using convolutional vision transformer

    Deressa Wodajo and Solomon Atnafu. Deepfake video dete ction using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021

  10. [18]

    A siamese-based verification system for open-set architecture attribution of synthetic images

    Lydia Abady, Jun Wang, Benedetta Tondi, and Mauro Barni . A siamese-based verification system for open-set architecture attribution of synthetic images. Pattern Recognition Letters, 180:75–81, 2024. 16

  11. [19]

    Syn- thetic image verification in the era of generative artificial intelligence: What works and what isn’t there yet

    Diangarti Tariang, Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa V erdoliva. Syn- thetic image verification in the era of generative artificial intelligence: What works and what isn’t there yet. IEEE Security & Privacy , 2024

  12. [20]

    Reliable out-of-dis tribution recognition of synthetic images

    Anatol Maier and Christian Riess. Reliable out-of-dis tribution recognition of synthetic images. Journal of Imaging, 10(5):110, 2024

  13. [21]

    Improving gener alization of deepfake detection with data farming and few-shot learning

    Pavel Korshunov and Sébastien Marcel. Improving gener alization of deepfake detection with data farming and few-shot learning. IEEE Transactions on Biometrics, Behavior , and Identity Science, 4(3):386–397, 2022

  14. [22]

    Generalized zero and few-shot transfer for facial forgery detection

    Shivangi Aneja and Matthias Nießner. Generalized zero and few-shot transfer for facial forgery detection. arXiv preprint arXiv:2006.11863, 2020

  15. [23]

    A baseline for detectin g misclassified and out-of-distribution examples in neural networks

    Dan Hendrycks and Kevin Gimpel. A baseline for detectin g misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136 , 2016

  16. [24]

    Enhanc ing the reliability of out-of-distribution image detec- tion in neural networks

    Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhanc ing the reliability of out-of-distribution image detec- tion in neural networks. arXiv preprint arXiv:1706.02690 , 2017

  17. [25]

    Learning confiden ce for out-of-distribution detection in neural net- works

    Terrance DeVries and Graham W Taylor. Learning confiden ce for out-of-distribution detection in neural net- works. arXiv preprint arXiv:1802.04865 , 2018

  18. [26]

    Can multi-label classification networks know what they don’t know? Advances in Neural Information Processing Systems , 34:29074–29087, 2021

    Haoran Wang, Weitang Liu, Alex Bocchieri, and Yixuan Li . Can multi-label classification networks know what they don’t know? Advances in Neural Information Processing Systems , 34:29074–29087, 2021

  19. [27]

    En ergy-based out-of-distribution detection

    Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. En ergy-based out-of-distribution detection. Advances in neural information processing systems , 33:21464–21475, 2020

  20. [28]

    Leveraging visual attention for out-of-distribution detection

    Luca Cultrera, Lorenzo Seidenari, and Alberto Del Bimb o. Leveraging visual attention for out-of-distribution detection. In Proceedings of the IEEE/CVF International Conference on Co mputer Vision, pages 4447–4456, 2023

  21. [29]

    Cutmix: Regularization strategy to train strong classifiers with lo calizable features

    Sangdoo Y un, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun , Junsuk Choe, and Y oungjoon Y oo. Cutmix: Regularization strategy to train strong classifiers with lo calizable features. In Proceedings of the IEEE/CVF international conference on computer vision , pages 6023–6032, 2019

  22. [30]

    Detecting o ut-of-distribution examples with in-distribution ex- amples and gram matrices

    Chandramouli Shama Sastry and Sageev Oore. Detecting o ut-of-distribution examples with in-distribution ex- amples and gram matrices. arXiv e-prints, pages arXiv–1912, 2019

  23. [31]

    Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data

    Y en-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10951–10960, 2020

  24. [32]

    A sim ple unified framework for detecting out-of- distribution samples and adversarial attacks

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A sim ple unified framework for detecting out-of- distribution samples and adversarial attacks. Advances in neural information processing systems , 31, 2018

  25. [33]

    Mixture densities, maximum likelihood and the em algorithm

    Richard A Redner and Homer F Walker. Mixture densities, maximum likelihood and the em algorithm. SIAM review, 26(2):195–239, 1984

  26. [34]

    Bayesian learning for neural networks , volume 118

    Radford M Neal. Bayesian learning for neural networks , volume 118. Springer Science & Business Media, 2012

  27. [35]

    Markov chain Monte Carlo: stochastic simulation for Bayesian inference

    Dani Gamerman and Hedibert F Lopes. Markov chain Monte Carlo: stochastic simulation for Bayesian inference. Chapman and Hall/CRC, 2006

  28. [36]

    Mixture outlier exposure: Towards out-of-distribution detection in fine-grained env ironments

    Jingyang Zhang, Nathan Inkawhich, Randolph Linderman , Yiran Chen, and Hai Li. Mixture outlier exposure: Towards out-of-distribution detection in fine-grained env ironments. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 5531–5540, 2023

  29. [37]

    U- net: Convolutional networks for biomedical image seg- mentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image seg- mentation. In Medical image computing and computer-assisted interventi on–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedin gs, ...

  30. [38]

    Randaugment: Practical automated data augmen- tation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le . Randaugment: Practical automated data augmen- tation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision a nd pattern recognition workshops, pages 702–703, 2020

  31. [39]

    Learning representations by back-propagating errors

    David E Rumelhart, Geoffrey E Hinton, and Ronald J Willi ams. Learning representations by back-propagating errors. nature, 323(6088):533–536, 1986

  32. [40]

    A continual deepfake detection benchmark: Datase t, methods, and essentials

    Chuqiao Li, Zhiwu Huang, Danda Pani Paudel, Y abin Wang, Mohamad Shahbazi, Xiaopeng Hong, and Luc V an Gool. A continual deepfake detection benchmark: Datase t, methods, and essentials. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vis ion, pages...

  33. [41]

    Openood v1.5: Enhanced benchmark for out-of-distribution detection, 2023

    Jingyang Zhang, Jingkang Y ang, Pengyun Wang, Haoqi Wang, Y ueqian Lin, Haoran Zhang, Yiyou Sun, Xuefeng Du, Kaiyang Zhou, Wayne Zhang, Yixuan Li, Ziwei Liu, Yiran Ch en, and Hai Li. Openood v1.5: Enhanced benchmark for out-of-distribution detection, 2023. 18

Pith tools

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