Pith. sign in

REVIEW 3 major objections 5 minor 54 references

EVAS: Efficient Multimodal Temporal Forgery Localization via Audio-Visual Synergy and Steered Boundary Calibration

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

Pith's one-line read EVAS localizes sparse deepfake segments in long videos by deep audio-visual synergy and training-only boundary sharpening, reporting much higher strict-threshold accuracy than prior methods.

desk verdict Solid multimodal TFL system paper with real SOTA margins and a clean efficiency story; the BAR transfer story is under-isolated but not a reason to dismiss the work. read the letter →

arxiv 2607.04472 v1 pith:5ASV4ROU submitted 2026-07-05 cs.CV

classification cs.CV
keywords MultimodalTemporalForgeryLocalizationMulti-StageAudio-VisualSynergyBoundary-AwareRefinementdeepfakedetectiondesynchronizationHourglassFFNanchor-free
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

Deepfakes that hide brief forged intervals inside long authentic footage evade whole-video detectors and also defeat ordinary action-localization pipelines that smooth away the tiny audio-visual desynchronizations that betray neural rendering. EVAS attacks that gap with two design moves. First, a multi-stage synergy module repeatedly lets each modality query the other so that high-frequency sync artifacts stay visible instead of being washed out by late fusion. Second, a boundary-aware refinement procedure trains several detection heads under ground-truth invalid-frame masks that force attention onto true transitions, then discards every auxiliary head at test time and runs only the primary detector. Together with a compressed feed-forward block that cuts latency, the resulting model records state-of-the-art average precision and recall on three standard temporal-forgery benchmarks, with the largest gains at the strictest localization thresholds. A sympathetic reader cares because reliable interval-level forensics is what content platforms and investigators actually need once generation quality makes binary “fake or real” labels insufficient.

What carries the argument

Multi-Stage Audio-Visual Synergy (MAVS) plus Boundary-Aware Refinement (BAR): MAVS iteratively updates each modality by treating the other as basis and message while accumulating residual history; BAR trains N detection heads under ground-truth-guided masks that zero out pristine frames, then at inference keeps only the first detector so that recursive error cannot propagate.

What would settle it

Retrain EVAS with the same architecture and loss but without any invalid-frame masking or auxiliary heads; if the AP@0.95 gap versus the full model collapses on the official LAV-DF full set, the claimed transfer of boundary calibration fails.

Watch

Extended reading notes

Core claim

The paper claims that progressive multi-stage audio-visual querying combined with a decoupled boundary-refinement strategy—teacher-forced invalid-frame masking during training only—produces deep multimodal forensic features that recover sparse, high-frequency forgery boundaries far more accurately than shallow-fusion or cascaded-inference baselines, yielding state-of-the-art average precision and recall on LAV-DF, AV-Deepfake1M and TVIL.

Load-bearing premise

The boundary-sharpening that ground-truth masking teaches the network during training must fully transfer to a single primary detector run alone at inference, without the distribution shift the paper itself says would collapse accuracy if the auxiliary heads stayed active.

Editorial extensions

If this is right

  • Strict-threshold localization (AP@0.95) becomes practical for content moderation pipelines that previously had to accept coarse or whole-video decisions.
  • Real-time or near-real-time forensic scanning of long-form video is feasible once the HourglassFFN latency reduction is retained.
  • Unimodal visual-only variants can still outperform prior multimodal baselines when audio is unavailable or corrupted.
  • Cascaded multi-stage detectors that keep every head at inference are shown to be unnecessary and actively harmful for sparse forgery localization.

Reading between the lines

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

  • The same teacher-forced masking plus single-head inference pattern could be ported to other sparse temporal tasks (e.g., rare-event detection or micro-expression localization) where recursive refinement is known to amplify early errors.
  • If the residual multi-stage querying truly isolates phase jitter, adversarial generators that deliberately re-synchronize audio and video at the sample level may become the next natural attack surface.
  • Because the method freezes large pretrained encoders and only tunes adapters plus heads, the same recipe may scale to higher-resolution or multi-speaker video without re-training the entire backbone.
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

3 major / 5 minor

Summary. The manuscript proposes EVAS, an end-to-end multimodal framework for temporal forgery localization (TFL) of sparsely distributed deepfake segments. It introduces Multi-Stage Audio-Visual Synergy (MAVS / Hierarchical State Synchronization) that iteratively updates audio and visual states via dense mutual queries (Eqs. 2–5, L=2), Boundary-Aware Refinement (BAR) that trains N cascaded detectors with ground-truth-guided invalid-frame masking (γ=0.2) then discards auxiliary heads at inference (Eq. 11), and a lightweight HourglassFFN with SSLReLU. On LAV-DF, AV-Deepfake1M and TVIL the method reports state-of-the-art mAP and average recall, with a particularly large gain at the strict threshold AP@0.95 = 88.63 on LAV-DF versus prior best ≈37.61 (UMMAFormer). Ablations (Tables 5–7, Figs. 4–6) examine component contributions, stage count, mask ratio and robustness to compression/noise.

Significance. If the reported strict-threshold gains hold under fair comparison, the work is a useful advance for fine-grained multimodal forensics: sparse, short forgeries are practically important and most prior TFL systems degrade sharply at high IoU. The explicit train/inference decoupling of BAR, the efficiency numbers (Table 3, ~50 ms vs tens of seconds), the robustness curves (Fig. 6) and the qualitative localization examples (Fig. 7) are concrete engineering contributions. The paper also supplies systematic ablations of L, N and γ, which is more than many architecture papers provide. The result is therefore of interest to the multimedia-forensics and temporal-localization communities provided the attribution of the headline AP@0.95 leap can be tightened.

major comments (3)
  1. [§3.3, Table 5, Table 7, Fig. 4] §3.3 (Eqs. 7–11) and the central claim that BAR’s multi-stage GT-masked training “steers” the primary detector used alone at inference. Table 5 shows A+V → A+V+BAR raises AP@0.95 from 83.96 to 88.25, and Table 7 / Fig. 4 show that N=3 is best and that using later heads at test time collapses. What is missing is a controlled single-head baseline: the identical architecture and loss trained with N=1 (no auxiliary heads, no GT invalid-frame masking) versus the full BAR regime that still evaluates only Detector1. Without that comparison it remains possible that the strict-threshold jump is driven by end-to-end VideoMAE/BYOL-a fine-tuning, feature-pyramid averaging or the multi-head loss rather than the claimed transfer of boundary sharpening. This ablation is load-bearing for the mechanism that is supposed to explain the SOTA numbers.
  2. [Table 1, §4.1] Table 1 (and similarly Tables 2, 4): almost all published baselines are marked E2E = × and use frozen or offline features, while EVAS is fully end-to-end with a strong VideoMAE-S temporal adapter + BYOL-a backbone. The AP@0.95 gap (88.63 vs 37.61) is therefore confounded by training regime and feature strength. A feature-matched or at least end-to-end re-implementation of the strongest prior (UMMAFormer / ActionFormer) under the same backbone and training schedule is needed before the architectural claims (MAVS + BAR) can be credited with the full margin. At minimum the paper should quantify how much of the gain remains when EVAS is trained with frozen encoders.
  3. [§4, Tables 1–4] No multi-seed statistics or error bars accompany the extreme AP@0.95 claim. Given that the hyper-parameters L=2, N=3, γ=0.2 were selected by validation search and that early stopping is used (max 12 epochs), a modest variance estimate (3–5 independent runs) is required to establish that the reported leap is stable rather than a lucky seed. This is especially important for a result that more than doubles the previous best strict-threshold score.
minor comments (5)
  1. [Fig. 1, Fig. 2, Fig. 7] Figure captions and body text inconsistently render the method name as “EV AS” (with a space) versus “EVAS”. Standardize throughout.
  2. [§3.4] The SSLReLU moment derivation (Eqs. 13–15) assumes the pre-activation is exactly standard normal. After residual blocks and LayerNorm this is only approximate; a short empirical check of the actual mean/variance entering HourglassFFN would strengthen the claim that the closed-form s, b are optimal.
  3. [Table 5] Table 5: adding MAVS after BAR slightly lowers mAP (94.63 → 94.55) while raising AP@0.95. A one-sentence discussion of this trade-off would help readers interpret the synergy claim.
  4. [§3.5, §4 Training Settings] Hyper-parameter λ (regression weight) and the Hourglass bottleneck width h / SSLReLU slope α are listed as free parameters but never given numerical values or sensitivity plots. Please report the final settings used for all tables.
  5. [§2] Related-work coverage of recent audio-visual deepfake localization (e.g., AVFF, Delocate, Face Forgery Video Detection via Temporal Forgery Cue Unraveling) is thin; a short paragraph situating MAVS relative to these concurrent cross-attention designs would improve positioning.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical end-to-end architecture paper whose SOTA claims rest on held-out benchmark metrics, not quantities defined from fitted constants or self-referential equations.

full rationale

EVAS is a standard multimodal detection architecture (MAVS cross-modal fusion + BAR multi-head training with GT-guided invalid-frame masking + HourglassFFN). All reported AP/AR numbers (Tables 1-4) are computed against independent ground-truth labels on public test splits of LAV-DF, AV-Deepfake1M and TVIL; none of the equations (1-16) define a metric in terms of a free parameter that is later re-presented as a prediction. Hyper-parameters L=2, N=3, γ=0.2 are selected by ordinary validation ablations (Tables 5-7, Figs. 4-5) and do not force the test-set scores by construction. The decoupled train/infer strategy (Eq. 11) is a design choice whose benefit is measured empirically, not assumed. Citations are to external prior art (ActionFormer, UMMAFormer, VideoMAE, etc.); no load-bearing uniqueness theorem or ansatz is imported from overlapping authors. The paper therefore contains no self-definitional loop, fitted-input-as-prediction, or self-citation chain that reduces its central claims to their inputs.

Assumptions & free parameters 5 free parameters · 5 assumptions · 3 invented entities

The central SOTA claim rests on standard deep-learning practice plus a handful of hand-chosen architectural hyperparameters and the modeling assumption that GT-guided multi-stage masking during training improves a single-head inference detector. No new physical entities are postulated; the “invented” items are architectural modules whose only evidence is the paper’s own ablations.

free parameters (5)
  • MAVS depth L = 2
    Number of cascaded cross-modal stages; set to 2 after ablation (Table 6).
  • BAR stages N = 3
    Number of cascaded detection heads used only in training; set to 3 after ablation (Table 7).
  • mask relaxation ratio γ = 0.2
    Temporal padding fraction around predicted segments when building invalid-frame masks; set to 0.2 after sensitivity plot (Fig. 5).
  • regression loss weight λ
    Balances Focal classification vs Distance-IoU regression in the total loss (Eq. 16); value not numerically fixed in text beyond being a hyperparameter.
  • Hourglass bottleneck width h and SSLReLU slope α
    Compression ratio and negative slope of the custom activation; chosen for efficiency/accuracy trade-off, not derived.
assumptions (5)
  • domain assumption Audio and visual streams are temporally aligned at input and can be encoded into fixed-length sequences of equal temporal resolution for cross-modal attention.
    Stated in §3.1–3.2 feature encoding; required for the Probe/Basis/Message construction.
  • domain assumption Ground-truth forgery intervals are available and accurate during training for teacher-forced invalid-frame masks.
    BAR §3.3 builds masks from G; without reliable GT the multi-stage training signal collapses.
  • ad hoc to paper Discarding auxiliary heads at inference still retains the benefit of multi-stage GT-masked training (no recursive error propagation).
    Core design choice of BAR (Eq. 11); justified by Fig. 4 but not independently proven outside this architecture.
  • domain assumption Standard Transformer-style scaled-dot-product attention and residual accumulation preserve high-frequency forensic cues better than late fusion.
    Motivates MAVS equations (2)–(5) against shallow fusion baselines.
  • standard math Focal Loss and Distance-IoU Loss are appropriate surrogates for sparse temporal localization quality.
    Cited losses [44,45] used in §3.5 without re-derivation.
invented entities (3)
  • Multi-Stage Audio-Visual Synergy (MAVS / Hierarchical State Synchronization)
    purpose: Progressive bidirectional cross-modal state updates that keep each modality’s topology while exposing sync artifacts.
    Defined in §3.2; evidence is only the paper’s ablations and SOTA tables, not external independent measurements.
  • Boundary-Aware Refinement (BAR) with invalid-frame masking and decoupled train/infer heads
    purpose: Sharpen forgery boundaries during training without recursive error at test time.
    Defined in §3.3; independent_evidence false because the mechanism is validated only on the same benchmarks.
  • HourglassFFN with SSLReLU
    purpose: Reduce FFN compute while preserving gradient flow through a narrow bottleneck.
    §3.4 introduces the activation and normalization constants; no external validation beyond latency tables in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EVAS: Efficient Multimodal Temporal Forgery Localization via Audio-Visual Synergy and Steered Boundary Calibration." pith.science (2026). https://pith.science/paper/5ASV4ROU

@misc{pith2026260704472,
  author       = {Pith},
  title        = {Pith review of: EVAS: Efficient Multimodal Temporal Forgery Localization via Audio-Visual Synergy and Steered Boundary Calibration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ASV4ROU}},
  note         = {Machine review of arXiv:2607.04472}
}
read the original abstract

The rapid proliferation of artificial intelligence-generated content necessitates reliable multimodal forensics. Beyond video-level binary classification, precisely localizing sparsely distributed forged segments in long-form videos remains a critical challenge. This task is particularly difficult when manipulations are subtly embedded and cross-modal signals are weak and temporally diffuse. To address these challenges, we propose EVAS, an end-to-end multimodal framework for temporal forgery localization. At its core, a Multi-Stage Audio-Visual Synergy mechanism facilitates progressive cross-modal interaction to learn deep multimodal forensic representations and capture high-order semantic traces of sparse manipulations. Furthermore, we introduce a Boundary-Aware Refinement strategy to achieve steered boundary calibration. By incorporating invalid-frame masking, this strategy suppresses ambiguous regions and sharpens transition predictions. We adopt a decoupled training paradigm with auxiliary heads to disentangle representation learning from inference objectives, enhancing model generalization and stability. Additionally, a lightweight HourglassFFN is incorporated to reduce computational overhead. Extensive experiments demonstrate that EVAS achieves state-of-the-art average localization accuracy and average recall across three benchmark datasets, validating its effectiveness for fine-grained temporal forgery localization.

Figures

Figures reproduced from arXiv: 2607.04472 by the authors.

Figure 1
Figure 1. (a) The qualitative visualization demonstrates that EVAS aligns more precisely with ground truth boundaries compared to the audio [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed EVAS framework and schematic illustration of the MAVS mechanism. Dashed cubes within the framework [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the HourglassFFN. These regions are then converted into a binary mask vector 𝑀𝑘 ∈ {0, 1} 𝑇 using an indicator function I: 𝑀𝑘 [𝑡] = I(𝑡 ∈ R𝑘 ) (9) The mask is applied to the fused features from the previous stage 𝐹 𝑘−1 𝑚 via the Hadamard product ⊙ to generate the input for the next stage: 𝐹 𝑘 𝑚 = 𝐹 𝑘−1 𝑚 ⊙ 𝑀𝑘 (10) By enforcing this operation, the mechanism completely zeros out pristine background informat… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Analysis of model inference performance using detection [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Analysis of model robustness under varying video compression and audio noise intensities on the LAVDF dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visualization of localization results on the Lav-DF benchmark. We compare the ground truth shown in purple with predictions from [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 12 canonical work pages

  1. [1]

    2025.Drifting A way from Truth: GenAI-Driven News Diversity Challenges LVLM- Based Misinformation Detection

    Fanxiao Li, Jiaying Wu, Tingchao Fu, Yunyun Dong, Bingbing Song, Wei Zhou. 2025.Drifting A way from Truth: GenAI-Driven News Diversity Challenges LVLM- Based Misinformation Detection. https://doi.org/10.48550/arXiv.2508.12711

  2. [2]

    2025.Zooming In on Fakes: A Novel Dataset for Localized AI- Generated Image Detection with Forgery Amplification Approach

    Lvpan Cai, Haowei Wang, Jiayi Ji, Yanshu Zhoumen, Shen Chen, Taiping Yao, Xiaoshuai Sun. 2025.Zooming In on Fakes: A Novel Dataset for Localized AI- Generated Image Detection with Forgery Amplification Approach. https://doi.org/ 10.48550/arXiv.2504.11922

  3. [3]

    Tong Qiao, Shichuang Xie, Yanli Chen, Florent Retraint, Xiangyang Luo. 2024. Fully Unsupervised Deepfake Video Detection Via Enhanced Contrastive Learning. https://doi.org/10.1109/TPAMI.2024.3356814

  4. [4]

    2015.Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich. 2015.Going deeper with convolutions. https://doi.org/10.1109/CVPR.2015.7298594

  5. [5]

    2018.Cascade R-CNN: Delving into High Quality Object Detection

    Zhaowei Cai, Nuno Vasconcelos. 2018.Cascade R-CNN: Delving into High Quality Object Detection. https://doi.org/10.1109/CVPR.2018.00644

  6. [6]

    2022.Do You Really Mean That? Content Driven Audio-Visual Deepfake Dataset and Multimodal Method for Temporal Forgery Localization

    Zhixi Cai, Kalin Stefanov, Abhinav Dhall, Munawar Hayat. 2022.Do You Really Mean That? Content Driven Audio-Visual Deepfake Dataset and Multimodal Method for Temporal Forgery Localization. In 2022 International Conference on Digital Image Computing: Techniques and Applications (DICTA), 1–10. https://doi.org/ 10.1109/DICTA56598.2022.10034605

  7. [7]

    2016.Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks

    Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun. 2016.Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. https://doi.org/10.1109/ TPAMI.2016.2577031

  8. [8]

    Rojas, Ali Thabet, Bernard Ghanem

    Mengmeng Xu, Chen Zhao, David S. Rojas, Ali Thabet, Bernard Ghanem. 2020. G-TAD: Sub-Graph Localization for Temporal Action Detection. https://doi.org/10. 1109/CVPR42600.2020.01017

Show all 54 references
  1. [9]

    2019.Exposing DeepFake Videos By Detecting Face Warping Artifacts

    Yuezun Li, Siwei Lyu. 2019.Exposing DeepFake Videos By Detecting Face Warping Artifacts. https://doi.org/10.48550/arXiv.1811.00656

  2. [10]

    Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, Alexei A. Efros. 2019.Detecting Photoshopped Faces by Scripting Photoshop. https://doi.org/10. 48550/arXiv.1906.05856

  3. [11]

    2020.DeepFake Detection via Facial Landmark Analysis

    Tackhyun Jung, Sangwon Kim, Keecheon Kim. 2020.DeepFake Detection via Facial Landmark Analysis. https://doi.org/10.1109/ACCESS.2020.2988660

  4. [12]

    2020.FakeCatcher: Detection of Syn- thetic Portrait Videos using Biological Signals

    Umur Aybars Ciftci, Ilke Demir, Lijun Yin. 2020.FakeCatcher: Detection of Syn- thetic Portrait Videos using Biological Signals. https://doi.org/10.1109/TPAMI.2020. 3009287

  5. [13]

    2019.FaceForensics++: Learning to Detect Manipulated Facial Images

    Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, Matthias Niessner. 2019.FaceForensics++: Learning to Detect Manipulated Facial Images. https://doi.org/10.1109/ICCV.2019.00009

  6. [14]

    2018.MesoNet: a Compact Facial Video Forgery Detection Network

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, Isao Echizen. 2018.MesoNet: a Compact Facial Video Forgery Detection Network. https://doi.org/10.1109/WIFS. 2018.8630761

  7. [16]

    2023.Dis- criminative Feature Mining Based on Frequency Information and Metric Learning for Face Forgery Detection

    Jiaming Li, Hongtao Xie, Lingyun Yu, Xingyu Gao, Yongdong Zhang. 2023.Dis- criminative Feature Mining Based on Frequency Information and Metric Learning for Face Forgery Detection. https://doi.org/10.1109/TKDE.2021.3117003

  8. [17]

    2022.Adaptive Face Forgery Detection in Cross Domain

    Luchuan Song, Zheng Fang, Xiaodan Li, Xiaoyi Dong, Zhenchao Jin, Yuefeng Chen, Siwei Lyu. 2022.Adaptive Face Forgery Detection in Cross Domain. https: //doi.org/10.1007/978-3-031-19830-4_27

  9. [18]

    2021.Learning Self-Consistency for Deepfake Detection

    Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, Wei Xia. 2021.Learning Self-Consistency for Deepfake Detection. https://doi.org/10.1109/ ICCV48922.2021.01475

  10. [19]

    2023.Deep Learning-Based Action Detection in Untrimmed Videos: A Survey

    Elahe Vahdani, Yingli Tian. 2023.Deep Learning-Based Action Detection in Untrimmed Videos: A Survey. https://doi.org/10.1109/TPAMI.2022.3193611

  11. [21]

    2023.PivoTAL: Prior-Driven Supervision for Weakly- Supervised Temporal Action Localization

    Mamshad Nayeem Rizve, Gaurav Mittal, Ye Yu, Matthew Hall, Sandra Sajeev, Mubarak Shah, Mei Chen. 2023.PivoTAL: Prior-Driven Supervision for Weakly- Supervised Temporal Action Localization. https://doi.org/10.1109/CVPR52729.2023. 02202

  12. [22]

    2024.Blind and Low Vision Individuals’ Detec- tion of Audio Deepfakes

    Filipo Sharevski, Rawan Zeidieh. 2024.Blind and Low Vision Individuals’ Detec- tion of Audio Deepfakes. In Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS ’24). https://doi.org/10.1145/ 3658644.3670353

  13. [23]

    Zhan Tong, Yibing Song, Jue Wang, Limin Wang. 2022.VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training.https://proceedings.neurips.cc/paper_files/paper/2022/file/ 416f9cb3276121c42eebb86352a4354a-Paper-Conference.pdf

  14. [24]

    2021.An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy et al. 2021.An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale. https://openreview.net/pdf?id=YicbFdNTTy

  15. [25]

    2024.End-to- End Temporal Action Detection with 1B Parameters Across 1000 Frames

    Shuming Liu, Chen-Lin Zhang, Chen Zhao, Bernard Ghanem. 2024.End-to- End Temporal Action Detection with 1B Parameters Across 1000 Frames. https: //doi.org/10.1109/CVPR52733.2024.01759

  16. [26]

    2021.BYOL for Audio: Self-Supervised Learning for General-Purpose Audio Representation

    Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, Kunio Kashino. 2021.BYOL for Audio: Self-Supervised Learning for General-Purpose Audio Representation. https://doi.org/10.1109/IJCNN52387.2021.9534474

  17. [27]

    Weinberger

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, Kilian Q. Weinberger. 2017. Densely Connected Convolutional Networks. https://doi.org/10.1109/CVPR.2017. 243

  18. [28]

    2021.Cross- Attentional Audio-Visual Fusion for Weakly-Supervised Action Localization

    Jun-Tae Lee, Mihir Jain, Hyoungwoo Park, Sungrack Yun. 2021.Cross- Attentional Audio-Visual Fusion for Weakly-Supervised Action Localization. https: //openreview.net/pdf?id=hWr3e3r-oH5

  19. [29]

    2024.MLCA-A VSR: Multi-Layer Cross Attention Fusion Based Audio-Visual Speech Recognition

    He Wang, Pengcheng Guo, Pan Zhou, Lei Xie. 2024.MLCA-A VSR: Multi-Layer Cross Attention Fusion Based Audio-Visual Speech Recognition. https://doi.org/10. 1109/ICASSP48485.2024.10446769

  20. [30]

    Komal Chugh, Parul Gupta, Abhinav Dhall, Ramanathan Subramanian. 2020. Not Made for Each Other: Audio-Visual Dissonance-based Deepfake Detection and Localization. In Proceedings of the 28th ACM International Conference on Multimedia, 439–447. https://doi.org/10.1145/3394171.3413700

  21. [31]

    2021.Activity Graph Transformer for Temporal Action Localization

    Megha Nawhal, Greg Mori. 2021.Activity Graph Transformer for Temporal Action Localization. https://doi.org/10.48550/arXiv.2101.08540

  22. [32]

    2019.BMN: Boundary- Matching Network for Temporal Action Proposal Generation

    Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, Shilei Wen. 2019.BMN: Boundary- Matching Network for Temporal Action Proposal Generation. https://doi.org/10. 1109/ICCV.2019.00399

  23. [33]

    2022.Hear Me Out: Fusional Approaches for Audio Augmented Temporal Action Localization

    Anurag Bagchi, Jazib Mahmood, Dolton Fernandes, Ravi Kiran Sarvadevabhatla. 2022.Hear Me Out: Fusional Approaches for Audio Augmented Temporal Action Localization. https://doi.org/10.48550/arXiv.2106.14118

  24. [34]

    2021.SOFT: Softmax-free Transformer with Linear Complexity.https://proceedings.neurips.cc/paper_files/paper/2021/file/ b1d10e7bafa4421218a51b1e1f1b0ba2-Paper.pdf

    Jiachen Lu et al. 2021.SOFT: Softmax-free Transformer with Linear Complexity.https://proceedings.neurips.cc/paper_files/paper/2021/file/ b1d10e7bafa4421218a51b1e1f1b0ba2-Paper.pdf

  25. [35]

    2022.ActionFormer: Localizing Moments of Actions with Transformers

    Chen-Lin Zhang, Jianxin Wu, Yin Li. 2022.ActionFormer: Localizing Moments of Actions with Transformers. In European Conference on Computer Vision (ECCV), 492–510. https://doi.org/10.1007/978-3-031-19772-7_29

  26. [36]

    2023.Ummaformer: A Universal Multimodal-Adaptive Transformer Framework for Temporal Forgery Localization

    Rui Zhang et al. 2023.Ummaformer: A Universal Multimodal-Adaptive Transformer Framework for Temporal Forgery Localization. https://doi.org/10.1145/3581783. 3613767

  27. [37]

    2024.DiModiF: Discourse Modality- Information Differentiation for Audio-Visual Deepfake Detection and Localization

    Christos Koutlis, Symeon Papadopoulos. 2024.DiModiF: Discourse Modality- Information Differentiation for Audio-Visual Deepfake Detection and Localization. https://doi.org/10.48550/arXiv.2411.10193

  28. [38]

    2018.Mesonet: a compact facial video forgery detection network

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, Isao Echizen. 2018.Mesonet: a compact facial video forgery detection network. In 2018 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1–7. IEEE. https: //doi.org/10.1109/WIFS.2018.8630761

  29. [39]

    2023.Glitch in the matrix: A large scale benchmark for content driven audio-visual forgery detection and localization

    Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, Munawar Hayat. 2023.Glitch in the matrix: A large scale benchmark for content driven audio-visual forgery detection and localization. Computer Vision and Image Un- derstanding, 236: 103818. https://doi.org/10....

  30. [40]

    2023.Tridet: Temporal Action Detection with Relative Boundary Modeling

    Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, Dacheng Tao. 2023.Tridet: Temporal Action Detection with Relative Boundary Modeling. https://doi.org/10. 1109/CVPR52729.2023.01808

  31. [41]

    Zhang et al

    Y. Zhang et al. 2024.MFMS: Learning Modality-Fused and Modality-Specific Features for Deepfake Detection and Localization Tasks. https://doi.org/10.1145/ 3664647.3688984

  32. [42]

    2017.Attention Is All You Need

    Ashish Vaswani et al. 2017.Attention Is All You Need. https://proceedings.neurips. cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf

  33. [43]

    2024.MetaFormer Baselines for Vision

    Weihao Yu et al. 2024.MetaFormer Baselines for Vision. IEEE Transactions on Pattern Analysis and Machine Intelligence. https://doi.org/10.1109/TPAMI.2023. 3329173

  34. [44]

    2020.Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression

    Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, Dongwei Ren. 2020.Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression. https://doi.org/10.1609/aaai.v34i07.6999

  35. [45]

    2017.Focal Loss for Dense Object Detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, Piotr Dollár. 2017.Focal Loss for Dense Object Detection. https://doi.org/10.1109/ICCV.2017.324

  36. [46]

    2025.Face Forgery Video Detection via Temporal Forgery Cue Unraveling

    Zonghui Guo, Yingjie Liu, Jie Zhang, Haiyong Zheng, Shiguang Shan. 2025.Face Forgery Video Detection via Temporal Forgery Cue Unraveling. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). https://doi. org/10.1109/CVPR52734.2025.00693

  37. [47]

    2024.A VFF: Audio-Visual Feature Fusion for Video Deepfake Detection

    Trevine Oorloff, Surya Koppisetti, Nicolò Bonettini, Divyaraj Solanki, Ben Col- man, Yaser Yacoob, Ali Shahriyari, Gaurav Bharaj. 2024.A VFF: Audio-Visual Feature Fusion for Video Deepfake Detection. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)...

  38. [48]

    2024.Delocate: Detection and Localization for Deepfake Videos with Randomly-Located Tampered Traces

    Juan Hu, Xin Liao, Difei Gao, Satoshi Tsutsui, Qian Wang, Zheng Qin, Mike Zheng Shou. 2024.Delocate: Detection and Localization for Deepfake Videos with Randomly-Located Tampered Traces. In Proceedings of the Thirty-Third Inter- national Joint Conference on Artificial Intellig...

  39. [49]

    2025.Trusted Video Inpainting Localization via Deep Attentive Noise Learning

    Zijie Lou, Gang Cao, Man Lin, Lifang Yu, Shaowei Weng. 2025.Trusted Video Inpainting Localization via Deep Attentive Noise Learning. In IEEE Transactions on Dependable and Secure Computing. https://doi.org/10.1109/TDSC.2025.3595960

  40. [50]

    2025.Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer

    Ziyi Liu, Yangcen Liu. 2025.Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). https://doi.org/10.1109/ CVPR52734.2025.00814

  41. [51]

    Jianyang Xie, Yitian Zhao, Yanda Meng, He Zhao, Anh Nguyen, Yalin Zheng. 2025. Are Spatial-Temporal Graph Convolution Networks for Human Action Recognition Over-Parameterized?. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). https://doi.org/10.11...

  42. [52]

    2024.SafeEar: Content Privacy-Preserving Audio Deepfake Detection

    Xinfeng Li, Kai Li, Yifan Zheng, Chen Yan, Xiaoyu Ji, Wenyuan Xu. 2024.SafeEar: Content Privacy-Preserving Audio Deepfake Detection. In Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security (CCS ’24). https://doi.org/10.1145/3658644.3690292

  43. [53]

    2022.Localizing Fake Segments in Speech

    Bowen Zhang, Terence Sim. 2022.Localizing Fake Segments in Speech. https: //doi.org/10.1109/ICPR56361.2022.9956134

  44. [54]

    2022.Proposal-Free Temporal Action Detection via Global Segmentation Mask Learning

    Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, Tao Xiang. 2022.Proposal-Free Temporal Action Detection via Global Segmentation Mask Learning. In European Conference on Computer Vision (ECCV), 645–662. https://doi.org/10.1007/978-3-031-20062- 5_37

  45. [55]

    2022.DCAN: Improving Temporal Action Detection via Dual Context Aggregation

    Guo Chen, Yin-Dong Zheng, Limin Wang, Tong Lu. 2022.DCAN: Improving Temporal Action Detection via Dual Context Aggregation. In Proceedings of the AAAI Conference on Artificial Intelligence, 248–257. https://doi.org/10.1609/aaai. v36i1.19900

  46. [56]

    2024.A V-Deepfake1M: A Large-Scale LLM-Driven Audio-Visual Deepfake Dataset

    Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, Kalin Stefanov. 2024.A V-Deepfake1M: A Large-Scale LLM-Driven Audio-Visual Deepfake Dataset. In Proceedings of the 32nd ACM International Conference on Multimedia, 7414–7423. https://doi.org...

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.