Pith. sign in

REVIEW 4 major objections 6 minor 48 references

A deepfake detector whose training is refreshed by an open adversarial competition keeps pace with the generative frontier, and dated snapshots outperform static baselines on recent fakes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 05:47 UTC pith:7MO7KI4L

load-bearing objection A serious, honestly-reported evaluation of a continuously refreshed detector; the refresh mechanism is plausible but not yet causally isolated from architecture churn. the 4 major comments →

arxiv 2607.13234 v1 pith:7MO7KI4L submitted 2026-07-14 cs.CV cs.AI

Continuously Evolving Deepfake Detection: An Architecture and Public-Benchmark Evaluation of a Dynamic Detection System

Ken Jon Miyachi , Dylan Uys This is my paper
classification cs.CV cs.AI MSC 68T4568T05
keywords deepfake detectioncontinuous adaptationadversarial competitiontemporal generalizationin-the-wild benchmarkgenerative mediaensemble modelcontamination audit
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the real-world failure of deepfake detectors is not architectural but temporal: a model trained once on a fixed corpus decays as new generators appear. It presents a detection system whose training distribution is continually refreshed by an open competition in which generative miners serve new models and discriminative miners compete to catch them. The authors evaluate one dated snapshot across nineteen public benchmarks and report strong performance, including improvements over the best published results on several in-the-wild suites. The load-bearing evidence is a temporal study: successive dated snapshots improve AUC on media from generators the static baseline never saw, on both image and video tracks.

Core claim

The central claim is that the system adapts even though every individual snapshot is static: each frozen export inherits a training distribution that is days, not years, behind the generative frontier. The evidence is a dated April 2026 export that, without per-benchmark tuning, reaches 0.936 AUC on Sumsub original images, 0.915 on the Deepfake-Eval-2024 image track, and 0.822 on its video track; and a temporal back-test in which successive snapshots improve from 0.842 to 0.902 AUC (image) and 0.864 to 0.936 (video) on test sets drawn from generators absent from a November 2025 static baseline's training.

What carries the argument

The mechanism is an open adversarial competition (the paper's SN34 loop): generative miners serve state-of-the-art generators under validator checks, discriminative miners submit detector checkpoints scored by a combined discrimination-and-calibration reward, and the winner seeds the production export. This loop, together with the GAS-Station dataset of verified adversarial media, is what continually refreshes the training distribution; the detector itself is a heterogeneous ensemble of four vision backbones with separate image, general-video, and face-video checkpoints.

Load-bearing premise

The entire adaptation argument rests on the forward GAS-Station weeks being genuinely post-export and on the perceptual-hash contamination audit catching all training near-duplicates; if either fails, the temporal and cross-dataset improvements are unsupported.

What would settle it

Take the production API snapshot dated April 15, 2026, run it on the timestamped GAS-Station weeks 26–28, and verify the published recall; then check a random sample of those weeks against the training corpora with a stronger near-duplicate search than perceptual hashing. If recall is much lower, or if duplicates surface, the mechanism's claimed freshness is refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, static benchmark retraining cycles are too slow; detection systems need continuous data refresh to remain usable.
  • Public benchmarks should adopt temporal protocols (train on early generators, test on later ones) as a standard evaluation axis.
  • The incentive mechanism turns the adversarial process into a scalable data engine: new generators enter the challenge distribution within hours rather than months.
  • The contamination-aware evaluation subsets (frame-level near-duplicate audit) should become standard practice for reporting cross-dataset face-video results.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same incentive-driven refresh idea could transfer to other non-stationary detection problems, e.g., fraud, malware, and spam, where the adversary's distribution also drifts weekly.
  • The paper's own failure analysis suggests pixel-space (non-latent) generators like Hourglass are a current blind spot; a testable prediction is that once such generators enter the competition stream, subsequent snapshots will close that gap.
  • The forward GAS-Station evaluation is only as strong as the guarantee that those weeks are truly post-export; an independent third party could re-run the exact API snapshot on those timestamped weeks to verify the recall figures.
  • The reported parity with the best commercial detector on Deepfake-Eval-2024 image track depends on a 95% CI that includes the published value; a larger in-the-wild sample would sharpen whether the system genuinely matches or exceeds commercial tools.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents BitMind Forensics (BMF), a deepfake-detection system built on the Bittensor SN34 incentive loop. The authors argue that static detectors are structurally doomed by generative drift, and that BMF's continuous, incentive-driven refresh of the training distribution is what keeps it current. They evaluate a single April 15, 2026 export (image, general-video, and human-video checkpoints) on nineteen public datasets, reporting strong results on in-the-wild benchmarks (Sumsub, Deepfake-Eval-2024), AI-image panels (Community Forensics, GenImage), and AI-video suites (GenVidBench, GenVideo). A temporal study shows successive dated exports improving on fixed test sets of recent-generator media, and a forward test on the GAS-Station stream shows the frozen snapshot flagging a stable fraction of post-export synthetic media. The paper is transparent about unfavorable numbers (e.g., 24% TPR@1%FPR on Deepfake-Eval video, ECE 0.479 on Sumsub, FF++ accuracy 31%) and releases its evaluation harness.

Significance. If the central claim were established, the paper would make a meaningful contribution: reframing deepfake detection as a process-level, non-stationary problem rather than an architecture-level one, and providing a large public benchmark evaluation of a deployed system. Genuine strengths include: one fixed snapshot with no per-benchmark tuning; DeLong confidence intervals on primary benchmarks; honest reporting of adverse operating points; a forward evaluation on post-export GAS-Station weeks; and a public harness (gasbench). The GAS-Station release is also a useful community resource. However, the paper's headline causal claim — that successive snapshot improvements are caused by training-data freshness rather than by architecture evolution or other confounds — is not yet supported by the evidence presented. The paper itself acknowledges this gap in §5.4 and §6.3, which makes the issue load-bearing rather than incidental.

major comments (4)
  1. [§5.4, §6.3, §3.4] The central claim that continuous data refresh, not architecture, keeps BMF current is not isolated by the temporal study. As the paper states in §5.4, successive snapshots 'may differ in architecture as well as training data'; §3.4 explains that each round is winner-take-all and the winning basis seeds the next export, so the architecture can change between snapshots. The observed monotone AUC gains (image 0.842→0.902; video 0.864→0.936) could therefore reflect better architectures surfaced by the competition, more training data in general, different model selection, or a combination. The paper's own §6.3 leaves 'an ablation isolating training-data freshness (the same model retrained without the most recent months of data)' to future work. That ablation is not optional for the paper's thesis; it is the decisive experiment. Without it, or without a comparison that holds the architecture
  2. [§5.4, Table 20, §4.2] The forward GAS-Station evaluation does not resolve the confound identified above. Showing that a frozen April snapshot maintains roughly stable recall on post-export weeks (image 0.929→0.917; video 0.789→0.814) is consistent with the mechanism working, but it is equally consistent with a static detector that generalizes reasonably to new generators; no comparison against a detector trained once on older data is provided. In addition, the claim that the GAS-Station weeks are genuinely post-export and free of leakage rests on the validators' C2PA/prompt checks and timestamped partitions, but §4.2 and §3.4 do not describe how timestamps are verified or how an adversarial generative miner could be prevented from backdating or mislabeling submissions. Since the forward test is a key piece of evidence for the deployment model, this verification needs to be described concretely.
  3. [§5.3.3, Table 19, §7] The face-video cross-dataset claims against Effort (DFDC 0.947 vs 0.843; Celeb-DF v2 0.9985 vs 0.956) rely on the contamination audit, which the paper itself limits to 'media-level duplication' and explicitly does not exclude 'person-level identity overlap' for celebrity-sourced benchmarks. If identity leakage inflates scores, those headline numbers are unsupported. Moreover, the comparison is not split-matched: BMF is scored on audit-verified subsets of roughly 1K clips, while Effort's published numbers come from the canonical protocol. The paper labels the comparison 'indicative' in the Table 19 caption, but the abstract and conclusion state the cross-dataset victory more flatly. Please either provide an identity-controlled analysis (e.g., train/test disjoint by person where feasible) or temper the abstract/conclusion claims to match the actual strength of the evidence.
  4. [§4.3, Table 4] All baseline comparisons use published point values rather than re-run baselines, and several rows are explicitly 'indicative' due to metric or aggregation mismatches (Community Forensics, DF40). This is disclosed, but it means the paper's comparative claims are weaker than a head-to-head evaluation would be. In particular, the Deepfake-Eval video comparison (BMF 0.822 vs best commercial 0.79) is reported as a numerical win even though the published 0.79 falls just below BMF's CI [0.791, 0.848] and the baseline is reported to two decimals. A matched re-run of at least the primary in-the-wild baselines would substantially strengthen the comparative claims.
minor comments (6)
  1. [§3.4, Eq. (3)] The quantities ^MCC and eB are described as 'normalized' but the normalization is not defined. Please specify the exact transformations and state whether α=1.2 and β=1.8 are fixed across all rounds.
  2. [Table 1] The backbone name 'EV A-L/14' appears to be a typo for 'EVA-L/14'; please correct.
  3. [§5.2.1] The attribution of Sumsub robustness 'primarily to full-frame scoring without a face-crop stage' is a plausible post-hoc hypothesis, but the ablations in §6.3 show that under degradation the ensemble can be diluted by weaker branches. Consider marking this as a hypothesis and testing it explicitly, e.g., by adding a face-crop condition to the ablation.
  4. [§5.4, Table 21] The dates are inconsistent: the text says 'November 7, 2025' while the table header says 'Nov 7 (static)'. Use a single format throughout.
  5. [Table 22] The Sumsub ECE of 0.479 is reported for the pooled all-conditions set, but Table 8 does not break ECE down by manipulation condition. Reporting per-condition ECE would help readers see where calibration degrades.
  6. [Reproducibility Statement] Because model weights are not public, independent verification must go through the production API. It would be helpful to state whether the API exposes per-sample softmax scores (needed to recompute DeLong CIs) and whether the released gasbench harness includes the exact preprocessing and routing used for the paper's reported numbers.

Circularity Check

0 steps flagged

No significant circularity: external benchmarks carry the central results; the one self-citation is not load-bearing, and the temporal claim is limited by an acknowledged confound rather than circular.

full rationale

The paper's headline benchmark claims are not defined in terms of its own outputs: one dated export is scored on nineteen public datasets with no per-benchmark tuning, and the metrics are compared to published external baselines. No parameter is fitted to the headline test sets, and no reported AUC is constructed from the system's own training signal. The only self-citation ([29], the BitMind SN34 framework) is background: the present paper specifies the mechanism in detail in §3.4 and tests it independently, so the citation is not load-bearing for the central claims. The closest thing to a circular risk is the temporal study in §5.4, where later snapshots are evaluated on generators they have already seen in training. The paper states this explicitly ('The test generators are absent from the training data of the static baseline only; the later snapshots have seen media from these generator families in training (though not the test items themselves), by design'), and frames the study as an end-to-end mechanism back-test rather than a zero-shot freshness test. The paper's own §6.3 also concedes the causal attribution is incomplete: 'An ablation isolating training-data freshness (the same model retrained without the most recent months of data) is left to future work; the temporal study (§5.4) provides the system-level version of that evidence.' §5.4 additionally concedes that successive snapshots 'may differ in architecture as well as training data.' That is a real internal-validity gap, but it is not circularity: the observed monotone improvement is not forced by definition, no equation equates the temporal gain to a fitted parameter, and the test items are not the training items. Score 2 reflects the minor non-load-bearing self-citation and the deferred freshness ablation, not substantive circular derivation.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The paper's core empirical claims depend mostly on the benchmark data and the audit, not on the fitted parameters, which mostly affect operating points. The mechanism premises (adversarial challenge distribution, valid timestamping) are assumed, not independently demonstrated here.

free parameters (5)
  • Ensemble fusion weights = A1 1/6, A2 1/6, B 1/3, C 1/3
    Hand-set in Eqs. (1)-(2); not learned. Modest reweighting would change operating points; the ablation shows uniform fusion can hurt under degradation.
  • SN34 scoring weights (α, β) = α=1.2, β=1.8
    Eq. (3): chosen to weight Brier over MCC in the competition objective; shapes which detector designs win rounds and therefore influences the evaluated snapshot.
  • Cosine-head scale and fake-logit bias = s=30, bias +0.5
    §3.1: hand-set bounds on per-branch confidence; affects calibration, which is reported separately.
  • Decision threshold = 0.5 fixed
    §4.1: fixed across all benchmarks; affects accuracy and recall operating points but not AUC.
  • Video inference budget = 16 frames, 1.3× face margin, confidence-weighted pooling
    §3.3: hand-set inference choices; video results would vary with budget or full-tiling mode, which is not evaluated.
axioms (6)
  • standard math DeLong method gives valid AUC confidence intervals for correlated samples
    Used for all confidence intervals; standard nonparametric method [41].
  • domain assumption The public benchmarks are representative of real circulated deepfakes
    §2.2 argues this from prior audits; the paper's in-the-wild value claim inherits benchmark validity.
  • domain assumption Validator C2PA and prompt-alignment checks keep GAS-Station media genuinely synthetic and correctly timestamped
    §3.4, §4.2: the forward temporal test and the no-leakage claim rely on this.
  • ad hoc to paper Delayed fool-rate reward makes the challenge distribution track the generative frontier
    §3.4: central mechanism premise; no direct measurement of how closely the challenge distribution approximates real-world circulation.
  • domain assumption Frame-level perceptual-hash audit detects training near-duplicates
    §7: self-conducted; identity overlap explicitly not excluded.
  • domain assumption MTCNN face crops are an adequate uniform preprocessing for face-video benchmarks
    §3.2: if MTCNN fails on a class of faces, those results would be affected.
invented entities (3)
  • BMF (dated production export) no independent evidence
    purpose: Deployed deepfake detector evaluated in the paper; the artifact whose numbers populate Tables 3–22.
    Weights and per-sample outputs not public; 'production API serves the exact evaluated snapshot' at publication (§4.1), but this is unverifiable from the paper and the API evolves afterwards.
  • Bittensor Subnet 34 incentive loop no independent evidence
    purpose: Continually refreshes training distribution and surfaces detector architectures (§3.4); the adaptation mechanism the paper argues is central.
    No external audit of the competition or its logs in this paper; the temporal study is the only evidence and does not isolate the mechanism from architecture changes.
  • GAS-Station independent evidence
    purpose: Adversarial media dataset produced by SN34; used for forward temporal evaluation and released for third-party training/testing (§4.2).
    Public Hugging Face release with weekly partitions and timestamps enables third-party leakage audits (https://huggingface.co/datasets/gasstation/gs-images-v4).

pith-pipeline@v1.3.0-alltime-deepseek · 22576 in / 15559 out tokens · 146187 ms · 2026-08-02T05:47:05.085227+00:00 · methodology

0 comments
read the original abstract

Deepfake detectors that achieve near-perfect scores on academic benchmarks collapse on real-world content: recent in-the-wild evaluations report AUC drops of 45-50% for state-of-the-art open-source models. We argue this gap is structural: static detectors are trained once against a moving generative frontier. We present BitMind Forensics (BMF), trained through Bittensor SN34, an open adversarial competition that continually refreshes the training distribution. We evaluate one dated export comprising image, general-video, and human-video checkpoints across nineteen public datasets: the canonical face-swap suites (FaceForensics++, Celeb-DF v1/v2/++, DFDC, DFD, UADFV, DF40) and recent in-the-wild and AI-generated-media benchmarks (Sumsub, Deepfake-Eval-2024, WildRF, Community Forensics, AIGCDetectBench, GenImage, AI-GenBench, AIGIBench, RAID, GenVidBench, GenVideo-100K). BMF reaches 0.936 AUC on Sumsub's original images and 0.872 pooled AUC over its full four-condition manipulation battery (1.4M images), staying robust under perturbation (0.855 JPEG, 0.799 downscaled), while GPEN enhancement improves detection (0.996). On Deepfake-Eval-2024, it matches the best commercial detector on images (0.915 vs 0.90) and exceeds it on video (0.822 vs 0.79), far above the best open-source detectors (0.56 and 0.63). It reaches 0.991 AUC on a 21-generator AI-image panel and 0.918 on GenVidBench, and exceeds the FF++-trained frontier on DFDC (0.947 vs 0.843) and Celeb-DF v2 (0.9985 vs 0.956), both contamination-audited, with statistical parity on Celeb-DF++. In a temporal study, successive dated exports improve on held-out media from generators absent from the static baseline's training (image 0.842 to 0.902; video 0.864 to 0.936). Our evaluation harness is public, and at publication the production API serves the exact evaluated snapshot for independent verification.

Figures

Figures reproduced from arXiv: 2607.13234 by Dylan Uys, Ken Jon Miyachi.

Figure 1
Figure 1. Figure 1: The SN34 loop. Generative miners, validators, and discriminative miners interact through Subnet 34: verified miner￾generated media are versioned in the GAS-Station data plane, and detector submissions are scored by the s34 objective in gasbench runs over GAS-Station data, public benchmarks, and a private hold￾out. Each round’s winning basis seeds the production export (BMF), which we fine-tune and extend b… view at source ↗
Figure 2
Figure 2. Figure 2: Adaptation over time, plotted from [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 31 linked inside Pith

  1. [1]

    Orthogonal subspace decomposition for gen- eralizable AI-generated image detection (Effort),

    Z. Yanet al., “Orthogonal subspace decomposition for gen- eralizable AI-generated image detection (Effort),” inICML,

  2. [2]

    GenImage: A million-scale benchmark for de- tecting AI-generated image,

    M. Zhuet al., “GenImage: A million-scale benchmark for de- tecting AI-generated image,” inNeurIPS Datasets & Bench- marks, 2023. arXiv:2306.08571

    Pith/arXiv arXiv 2023

  3. [3]

    Deepfake-Eval-2024: A multi-modal in-the-wild benchmark of Deepfakes circulated in 2024,

    N. A. Chandraet al., “Deepfake-Eval-2024: A multi-modal in-the-wild benchmark of Deepfakes circulated in 2024,” in CVPR Workshops, 2026. arXiv:2503.02857

    Pith/arXiv arXiv 2024

  4. [4]

    A sanity check for AI-generated image detec- tion,

    S. Yanet al., “A sanity check for AI-generated image detec- tion,” inICLR, 2025. arXiv:2406.19435

    Pith/arXiv arXiv 2025

  5. [5]

    Evaluating Deepfake detec- tors in the wild,

    V . Pirogov and M. Artemev, “Evaluating Deepfake detec- tors in the wild,” inICML DataWorld Workshop, 2025. arXiv:2507.21905

    Pith/arXiv arXiv 2025

  6. [6]

    Fit for purpose? Deepfake detection in the real world,

    G. Lin, L. Lin, C. P. Walker, D. S. Schiff, and S. Hu, “Fit for purpose? Deepfake detection in the real world,” 2025. arXiv:2510.16556

    arXiv 2025

  7. [7]

    How effective are publicly accessible Deepfake detection tools? a comparative evaluation of open-source and free-to-use platforms,

    M. Rettinger, B. Beaumont, N.-A. Le-Khac, and H.-H. Nguyen-Le, “How effective are publicly accessible Deepfake detection tools? a comparative evaluation of open-source and free-to-use platforms,” 2026. arXiv:2603.04456

    arXiv 2026

  8. [8]

    Perfor- mance decay in Deepfake detection: The limitations of train- ing on outdated data,

    J. Richings, M. Leblanc, I. Groves, and V . Nockles, “Perfor- mance decay in Deepfake detection: The limitations of train- ing on outdated data,” 2025. arXiv:2511.07009

    arXiv 2025

  9. [9]

    Community forensics: Using thou- sands of generators to train fake image detectors,

    J. Park and A. Owens, “Community forensics: Using thou- sands of generators to train fake image detectors,” inCVPR,

  10. [10]

    FaceForensics++: Learning to detect manip- ulated facial images,

    A. Rössleret al., “FaceForensics++: Learning to detect manip- ulated facial images,” inICCV, 2019. arXiv:1901.08971

    Pith/arXiv arXiv 2019

  11. [11]

    Detecting Deepfakes with self- blended images,

    K. Shiohara and T. Yamasaki, “Detecting Deepfakes with self- blended images,” inCVPR, 2022. arXiv:2204.08376

    Pith/arXiv arXiv 2022

  12. [12]

    Implicit identity leakage: The stumbling block to improving Deepfake detection generalization,

    S. Donget al., “Implicit identity leakage: The stumbling block to improving Deepfake detection generalization,” inCVPR,

  13. [13]

    Transcending forgery specificity with latent space augmentation for generalizable Deepfake detection,

    Z. Yanet al., “Transcending forgery specificity with latent space augmentation for generalizable Deepfake detection,” in CVPR, 2024. arXiv:2311.11278

    Pith/arXiv arXiv 2024

  14. [14]

    Towards universal fake image detectors that generalize across generative models,

    U. Ojha, Y . Li, and Y . J. Lee, “Towards universal fake image detectors that generalize across generative models,” inCVPR,

  15. [15]

    Forgery-aware adaptive transformer for generalizable synthetic image detection,

    H. Liuet al., “Forgery-aware adaptive transformer for generalizable synthetic image detection,” inCVPR, 2024. arXiv:2312.16649

    Pith/arXiv arXiv 2024

  16. [16]

    C2P-CLIP: Injecting category common prompt in CLIP to enhance generalization in Deepfake detection,

    C. Tanet al., “C2P-CLIP: Injecting category common prompt in CLIP to enhance generalization in Deepfake detection,” in AAAI, 2025. arXiv:2408.09647

    Pith/arXiv arXiv 2025

  17. [17]

    DIRE for diffusion-generated image detec- tion,

    Z. Wanget al., “DIRE for diffusion-generated image detec- tion,” inICCV, 2023. arXiv:2303.09295

    Pith/arXiv arXiv 2023

  18. [18]

    DRCT: Diffusion reconstruction contrastive training towards universal detection of diffusion generated im- ages,

    B. Chenet al., “DRCT: Diffusion reconstruction contrastive training towards universal detection of diffusion generated im- ages,” inICML, 2024. PMLR 235:7621–7639

    2024

  19. [19]

    Rethinking the up-sampling operations in CNN- based generative network for generalizable Deepfake detec- tion,

    C. Tanet al., “Rethinking the up-sampling operations in CNN- based generative network for generalizable Deepfake detec- tion,” inCVPR, 2024. arXiv:2312.10461

    Pith/arXiv arXiv 2024

  20. [20]

    A bias-free training paradigm for more general AI-generated image detection,

    F. Guillaro, G. Zingarini, B. Usman, A. Sud, D. Coz- zolino, and L. Verdoliva, “A bias-free training paradigm for more general AI-generated image detection,” inCVPR, 2025. arXiv:2412.17671

    Pith/arXiv arXiv 2025

  21. [21]

    Can ChatGPT detect deepfakes? a study of us- ing multimodal large language models for media forensics,

    S. Jiaet al., “Can ChatGPT detect deepfakes? a study of us- ing multimodal large language models for media forensics,” in CVPR Workshops, 2024. arXiv:2403.14077

    Pith/arXiv arXiv 2024

  22. [22]

    FakeShield: Explainable image forgery detection and localization via multi-modal large language models,

    Z. Xuet al., “FakeShield: Explainable image forgery detection and localization via multi-modal large language models,” in ICLR, 2025. arXiv:2410.02761

    Pith/arXiv arXiv 2025

  23. [23]

    DeepfakeBench: A comprehensive benchmark of Deepfake detection,

    Z. Yanet al., “DeepfakeBench: A comprehensive benchmark of Deepfake detection,” inNeurIPS Datasets & Benchmarks,

  24. [24]

    DF40: Toward next-generation Deep- fake detection,

    Z. Yanet al., “DF40: Toward next-generation Deep- fake detection,” inNeurIPS Datasets & Benchmarks, 2024. arXiv:2406.13495

    Pith/arXiv arXiv 2024

  25. [25]

    Fake or JPEG? revealing com- mon biases in generated image detection datasets,

    P. Grommeltet al., “Fake or JPEG? revealing com- mon biases in generated image detection datasets,” 2024. arXiv:2403.17608

    Pith/arXiv arXiv 2024

  26. [26]

    DeMamba: AI-generated video de- tection on million-scale GenVideo benchmark,

    H. Chenet al., “DeMamba: AI-generated video de- tection on million-scale GenVideo benchmark,” 2024. arXiv:2405.19707

    Pith/arXiv arXiv 2024

  27. [27]

    GenVidBench: A challenging bench- mark for detecting AI-generated video,

    Z. Niet al., “GenVidBench: A challenging bench- mark for detecting AI-generated video,” inAAAI, 2026. arXiv:2501.11340

    arXiv 2026

  28. [28]

    AI-GenBench: A new ongoing benchmark for AI-generated image detection,

    L. Pellegriniet al., “AI-GenBench: A new ongoing benchmark for AI-generated image detection,” inV erimedia Workshop @ IJCNN, 2025. arXiv:2504.20865

    arXiv 2025

  29. [29]

    Survival of the fittest detec- tors: A decentralized framework for evolving Deep- fake detection,

    BitMind Research Team, “Survival of the fittest detec- tors: A decentralized framework for evolving Deep- fake detection,” tech. rep., BitMind Research, 2024. https://assets.bitmindlabs.ai/pdfs/ survival-of-the-fittest-detectors.pdf

    2024

  30. [30]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inCVPR, 2022. arXiv:2201.03545

    Pith/arXiv arXiv 2022

  31. [31]

    EV A-02: A visual representation for neon genesis,

    Y . Fang, Q. Sun, X. Wang, T. Huang, X. Wang, and Y . Cao, “EV A-02: A visual representation for neon genesis,”Image and Vision Computing, vol. 149, 2024. arXiv:2303.11331

    Pith/arXiv arXiv 2024

  32. [32]

    Learning trans- ferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy,et al., “Learning trans- ferable visual models from natural language supervision,” in ICML, 2021. arXiv:2103.00020

    Pith/arXiv arXiv 2021

  33. [33]

    DI- NOv3,

    O. Siméoni, H. V . V o, M. Seitzer, F. Baldassarre,et al., “DI- NOv3,” 2025. arXiv:2508.10104

    Pith/arXiv arXiv 2025

  34. [34]

    V-JEPA 2: Self-supervised video mod- els enable understanding, prediction and planning,

    M. Assranet al., “V-JEPA 2: Self-supervised video mod- els enable understanding, prediction and planning,” 2025. arXiv:2506.09985

    Pith/arXiv arXiv 2025

  35. [35]

    Joint face detec- tion and alignment using multi-task cascaded convolutional networks,

    K. Zhang, Z. Zhang, Z. Li, and Y . Qiao, “Joint face detec- tion and alignment using multi-task cascaded convolutional networks,”IEEE Signal Processing Letters, vol. 23, no. 10, pp. 1499–1503, 2016. arXiv:1604.02878

    Pith/arXiv arXiv 2016

  36. [36]

    Is artificial intelligence generated image detection a solved problem? (AIGIBench),

    Z. Liet al., “Is artificial intelligence generated image detection a solved problem? (AIGIBench),” 2025. arXiv:2505.12335

    arXiv 2025

  37. [37]

    Real- time Deepfake detection in the real-world,

    B. Cavia, E. Horwitz, T. Reiss, and Y . Hoshen, “Real- time Deepfake detection in the real-world,” 2024. arXiv:2406.09398. 15

    Pith/arXiv arXiv 2024

  38. [38]

    RAID: A dataset for testing the adversar- ial robustness of AI-generated image detectors,

    H. Eddoubiet al., “RAID: A dataset for testing the adversar- ial robustness of AI-generated image detectors,” inNeurIPS Datasets & Benchmarks, 2025. arXiv:2506.03988

    arXiv 2025

  39. [39]

    CNN-generated images are surprisingly easy to spot. . . for now,

    S.-Y . Wang, O. Wang, R. Zhang, A. Owens, and A. A. Efros, “CNN-generated images are surprisingly easy to spot. . . for now,” inCVPR, 2020. arXiv:1912.11035

    Pith/arXiv arXiv 2020

  40. [40]

    Tracing hyperparame- ter dependencies for model parsing via learnable graph pooling network,

    X. Guo, V . Asnani, S. Liu, and X. Liu, “Tracing hyperparame- ter dependencies for model parsing via learnable graph pooling network,” inNeurIPS, 2024. arXiv:2312.02224

    Pith/arXiv arXiv 2024

  41. [41]

    Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach,

    E. R. DeLong, D. M. DeLong, and D. L. Clarke-Pearson, “Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach,” Biometrics, vol. 44, no. 3, pp. 837–845, 1988

    1988

  42. [42]

    Celeb-DF: A large- scale challenging dataset for DeepFake forensics,

    Y . Li, X. Yang, P. Sun, H. Qi, and S. Lyu, “Celeb-DF: A large- scale challenging dataset for DeepFake forensics,” inCVPR,

  43. [43]

    The DeepFake detection challenge (DFDC) dataset,

    B. Dolhanskyet al., “The DeepFake detection challenge (DFDC) dataset,” 2020. arXiv:2006.07397

    Pith/arXiv arXiv 2020

  44. [44]

    SimSwap: An efficient framework for high fidelity face swapping,

    R. Chen, X. Chen, B. Ni, and Y . Ge, “SimSwap: An efficient framework for high fidelity face swapping,” inACM Multime- dia, 2020. arXiv:2106.06340

    Pith/arXiv arXiv 2020

  45. [45]

    GAN prior embedded network for blind face restoration in the wild,

    T. Yang, P. Ren, X. Xie, and L. Zhang, “GAN prior embedded network for blind face restoration in the wild,” inCVPR, 2021. arXiv:2105.06070

    Pith/arXiv arXiv 2021

  46. [46]

    Exposing deep fakes using incon- sistent head poses,

    X. Yang, Y . Li, and S. Lyu, “Exposing deep fakes using incon- sistent head poses,” inICASSP, 2019. arXiv:1811.00656

    Pith/arXiv arXiv 2019

  47. [47]

    Celeb-DF++: A large-scale challenging video DeepFake benchmark for generalizable forensics,

    Y . Liet al., “Celeb-DF++: A large-scale challenging video DeepFake benchmark for generalizable forensics,” 2025. arXiv:2507.18015

    Pith/arXiv arXiv 2025

  48. [48]

    Scalable high-resolution pixel- space image synthesis with Hourglass diffusion transformers,

    K. Crowson, S. A. Baumann, A. Birch, T. M. Abraham, D. Z. Kaplan, and E. Shippole, “Scalable high-resolution pixel- space image synthesis with Hourglass diffusion transformers,” inICML, 2024. arXiv:2401.11605. 16

    arXiv 2024