Pith. sign in

REVIEW 4 major objections 6 minor 48 references

HexMIL: Hierarchical Attention MIL for Ante-Hoc Explainable Detection of AI-Manipulated CT Volumes

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Two-level attention trained on binary volume labels detects and localizes AI-manipulated CT scans, beating all baselines under cross-generator evaluation.

desk verdict Strong hierarchical MIL for CT forensics, but the 'volume-level only' supervision claim is contradicted by Stage 1's slice labels. read the letter →

arxiv 2608.05101 v1 pith:F7G42ITT submitted 2026-08-05 cs.CV cs.MM

classification cs.CVcs.MM
keywords medicaldeepfakesCTmanipulationdetectionmultipleinstancelearninggatedattentionexplainableAIcross-generatorgeneralization3Dlocalizationante-hocattribution
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

HexMIL aims to prove that a medical deepfake detector can be built without pixel-level masks: it needs only binary volume-level labels to decide whether a CT volume is tampered, and the same attention weights that make that decision also produce a full-resolution 3D heatmap localizing the manipulated sub-volume. The paper evaluates under a strict protocol where training uses volumes from one generative architecture and testing uses unseen ones. It reports gains of +9.1 AUC and +9.4 F1 over the strongest baselines in out-of-domain classification, and the best average IoU and Pointing Game scores in localization. The central point is that spatial attribution can be ante-hoc and structurally faithful, since the attention weights are the exact forward computation, not a post-hoc gradient approximation.

What carries the argument

The central object is a two-level hierarchy of Gated Attention modules, one over image patches within each axial slice and one over slices within the volume. Both follow the attention-based MIL formulation where each weight is a softmax over a gated combination of tanh and sigmoid projections, so each weight has an unambiguous spatial identity at its level. The two weight sets are combined into a 3D attention volume through multiplication, upsampling, global smoothing, and thresholding, which is then used to extract a tight axis-aligned bounding box around the manipulated sub-volume.

What would settle it

On a held-out set of tampered CT volumes with known manipulation masks, run HexMIL after corrupting the Stage 1 slice labels by shifting their axial coordinates by several slices; if the 3D attention volume still localizes the true manipulated region with unchanged IoU, the claim that localization is driven by genuinely learned forensic features would be undermined.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hierarchical gated-attention multiple instance learning architecture can simultaneously detect AI manipulation in CT volumes and localize the tampered region, using only binary volume-level supervision. The patch-level attention weights and the slice-level attention weights, learned independently in a two-stage procedure, are multiplied together at inference to form a full-resolution 3D attention volume. Because these weights are the pooling mechanism that produces both the slice representation and the volume representation, they constitute the exact forward computation driving the classification decision, which the paper contrasts with post-hoc methods like Grad-CAM and Grad-CAM++. In the reported cross-generator experiments on M3DSynth and CT-GAN, the method reaches 92.0 AUC and 90.2 F1 for out-of-domain classification and achieves the best average IoU of 42.4 and Pointing Game score of 70.6 for localization.

Load-bearing premise

The method assumes the axial coordinate metadata used to label tampered slices in Stage 1 is accurate and that those slice labels, although not pixel masks, provide enough spatial signal for patch attention to learn the true manipulated region.

Editorial extensions

If this is right

  • Medical deepfake detectors can be trained without expensive pixel-level annotations, since localization emerges from binary volume labels alone.
  • Explanations are structurally tied to the decision: removing or randomizing the attention weights collapses the model toward uniform averaging, so the heatmap and the classification share the same forward mechanism.
  • A detector trained on one generative family can flag manipulations from unseen generative families, with the reported margins of +9.1 AUC and +9.4 F1 over the strongest baseline.
  • Localization out-of-domain is attainable without spatial supervision, with an average IoU of 42.4 and Pointing Game score of 70.6, surpassing methods that use pixel-level supervision.
  • The same hierarchical design could extend to other volumetric modalities and anatomical sites, pending validation beyond the thorax as the authors note.

Reading between the lines

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

  • The 'binary volume-level supervision' claim is literal only at the volume level: Stage 1 consumes slice labels derived from the nodule axial coordinate in the dataset metadata, which is a spatially informed signal even if it is not a pixel mask; whether this counts as 'no spatial supervision' is a distinction the paper leaves open.
  • Because the 3D attention volume is the product of two normalized weight sets, its spatial resolution is capped by the patch grid, so a second finer-resolution pass over high-attention regions is a natural testable extension that could improve IoU on small nodules.
  • The ante-hoc faithfulness argument suggests a direct stress test beyond the paper: randomize the learned attention weights at inference on out-of-domain volumes and compare the resulting AUC drop with the effect of randomizing weights in a post-hoc saliency method; HexMIL's explanation should degrade exactly when its decision degrades.
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

4 major / 6 minor

Summary. HexMIL proposes a two-stage hierarchical attention MIL framework for detecting AI-manipulated CT volumes and localizing the manipulated sub-volume. Stage 1 trains a slice encoder (SliceMIL) on a slice-level binary classification task; Stage 2 freezes SliceMIL and trains a volume-level gated attention aggregator on volume-level binary labels. At inference, patch-level and slice-level attention maps are combined into a 3D attention volume, thresholded, and converted to a bounding box. The paper evaluates on M3DSynth and CT-GAN under a cross-generator protocol, reporting state-of-the-art classification (92.0 AUC, 90.2 F1) and localization (42.4 IoU, 70.6 Pointing Game) with no pixel-level masks. The abstract and conclusion claim that this is achieved 'using only binary volume-level supervision' and 'without any spatial supervision.'

Significance. If the reported results are reproducible, HexMIL is a useful contribution to medical deepfake detection: the cross-generator evaluation is rigorous, the baseline set is broad, and the ablations (Tables 3-5) help isolate the contribution of each architectural component. The ante-hoc framing, where attention weights are part of the forward computation rather than a post-hoc surrogate, is conceptually appealing and well motivated by the critique of Grad-CAM-style methods. However, the central supervision claim is overstated: Stage 1 uses slice-level labels derived from manipulation metadata, which is a form of spatial supervision along the slice axis. This does not invalidate the method, but it fundamentally changes the contribution from 'localization without spatial supervision' to 'localization with weak slice-level supervision.' The absence of error bars, multiple seeds, or significance tests further weakens the quantitative claims. These issues are fixable with additional experiments and a corrected narrative, but they are load-bearing for the paper's headline claims.

major comments (4)
  1. [Section 3.1, Eq. (4); Abstract; Conclusion] The training of SliceMIL in Stage 1 uses slice-level labels D^S_train constructed as 'tampered slices are those intersecting the manipulated sub-volume, identified via the nodule axial coordinate' (Section 3.1). The BCE loss in Eq. (4) is a per-slice classification loss on these labels. This is slice-axis spatial supervision: the model is explicitly told which slices contain the manipulation. Therefore the Abstract's 'using only binary volume-level supervision' and the Conclusion's 'without any spatial supervision' are inaccurate. The slice-axis component of the localization heatmap in Eq. (11) is not emergent from volume-level labels alone, because the volume attention beta_k operates on SliceMIL representations that were trained to recognize tampered slices. I recommend either retraining HexMIL from volume-level labels only (which would be a much stronger claim), or accurately reframing the contribution as 'no pixel-level masks' and adding an ablation that quantifies how much slice-level pretraining contributes to the reported localization accuracy.
  2. [Section 4.1, Tables 1 and 2] All classification and localization results are reported as single point estimates without error bars, multiple seeds, or significance tests. Given the high variance typical of cross-domain medical deepfake detection and the small number of transfer cells (12 directions), the +9.1 AUC and +9.4 F1 improvements over the strongest baselines are not statistically established. For example, the gap between HexMIL (92.0 AUC) and ManTraNet (82.9 AUC) may be genuine, but without variance estimates the reader cannot judge whether the gap is robust. The authors should report mean and standard deviation over at least five seeds, and preferably paired significance tests for the key comparisons.
  3. [Section 3.4, Implementation details, Table 2] The localization metrics depend on the thresholds tau_beta = 0.1, tau_3D = 0.65, and the Gaussian smoothing sigma = 0.03 * max(K,H,W), but no sensitivity analysis or principled selection procedure is provided. Since the 3D bounding box in Eq. (14) is extracted by thresholding the globally normalized heatmap, the reported IoU and Pointing Game scores could vary substantially with these hyperparameters. Please add a threshold sweep or a validation-based selection protocol to demonstrate that the localization results are not artifacts of a single favorable threshold setting.
  4. [Sections 3.3-3.4 and 4.2] The paper implicitly assumes that high attention weights correspond to manipulated voxels rather than to other discriminative cues such as anatomical structures or scanner artifacts. This assumption is load-bearing for the localization claim but is not tested. The in-domain evaluation (Figure 3) only compares HexMIL's heatmaps to Grad-CAM and Grad-CAM++ on a single backbone; it does not show that the attention maps track the manipulated sub-volume under label randomization, nor that removing the most-attended voxels changes the classification decision. I recommend adding control experiments, e.g., training with shuffled slice labels and showing that localization collapses, or comparing attention maps against an occlusion-based ground-truth attribution, to substantiate the 'structurally faithful' claim.
minor comments (6)
  1. [Section 3.4, Eq. (13)] The text says a normalized value v 'ranks in the top (1-v) fraction' of attended voxels, but min-max normalization does not produce ranks; please rephrase this to describe the relative threshold interpretation accurately.
  2. [Section 3.2, Eq. (5)] The sinusoidal positional encoding is added to slice representations, but the dimension of PE(z_k) is not specified. Please state that PE is D-dimensional and clarify how the slice index z_k is scaled before encoding.
  3. [Table 4] The row labels 'Standard' and 'Self-Attention' are not defined in the main text; please specify the exact attention formulations used in these ablations (e.g., Bahdanau-style additive attention vs. transformer-style key-query attention) so the comparison is reproducible.
  4. [Figure 3] Pixel AUC is used in the caption but not defined in the main text; please define it in the caption or in Section 4.2.
  5. [Table 1] The entry for DFX-SN appears as '73.1/76.070.9/70.6' with a missing delimiter; please fix the formatting.
  6. [Section 3.1] The phrase 'identified via the nodule axial coordinate available as standard dataset metadata' is vague; please specify the exact metadata field from LIDC-IDRI and confirm that this metadata is available for all training and test volumes, since it is the basis for the slice-level labels.

Circularity Check

1 steps flagged · score 6.0 of 10

Stage-1 slice-level labels derived from the nodule axial coordinate constitute spatial supervision, so the 'no spatial supervision' localization claim partially reduces to fitted input.

  1. fitted input called prediction [Section 3.1 (slice-level dataset construction), Section 3.3 (training claim), Section 3.4 (Eq. 11 heatmap)]
    "tampered slices T S train are those intersecting the manipulated sub-volume, identified via the nodule axial coordinate available as standard dataset metadata. ... Neither Lslc nor LV employs any spatial supervision; the 3D attention structure that delineates the manipulated region emerges entirely as a by-product of the binary classification objective."

    The slice-level dataset D^S_train is built by labeling every slice intersecting the nodule's axial coordinate as tampered. SliceMIL is trained on these slice-level labels via Eq. (4), so it learns to recognize tampered slices along the z-axis. This is spatial supervision: the model is explicitly told which axial slices contain the manipulation. The 3D attention volume in Eq. (11) multiplies the intra-slice attention A_k by the inter-slice weight beta_k, and beta_k aggregates SliceMIL representations that were trained to separate tampered from pristine slices. The slice-axis component of the localization heatmap is therefore a direct readout of Stage-1's slice-level fitted classifier, not an emergent by-product of volume-level binary labels.

full rationale

The paper's central classification result (out-of-domain AUC/F1) is evaluated against external baselines and does not reduce to its inputs, so the method retains independent content. However, the localization claim is partially circular. Stage 1 (Section 3.2) trains a slice classifier on slice-level labels that are constructed from the manipulation's axial coordinate metadata (Section 3.1). These labels are spatial supervision along the slice axis, directly contradicting the abstract's 'using only binary volume-level supervision' and Section 3.3's 'neither Lslc nor LV employs any spatial supervision.' The 3D heatmap in Eq. (11) combines the patch attention with the slice attention beta, and the slice attention operates on features trained to discriminate tampered slices. Thus the z-axis component of the localization is a fitted input renamed as an emergent prediction. The within-slice patch attention is not directly supervised, which preserves some independent grounding for the 2D localization, but the overclaim about supervision is load-bearing for the paper's novelty. No other circular steps, self-citation chains, or ansatz-smuggling were found; the citations to the authors' prior work are not load-bearing for the method's derivation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central method is built on the cited gated attention framework and introduces no new physical entities. The main extra assumptions are that slice-level metadata is reliable and that attention weights are spatially faithful. The primary hidden cost is the slice-level supervision that the abstract omits from its 'binary volume-level supervision' claim.

free parameters (6)
  • Patch size P = 64
    Selected as the operating point balancing OOD classification and localization in Table 5. Large P reduces IoU, small P reduces AUC. This is a test-set-dependent choice.
  • Window size K = 32
    Number of axial slices per sub-volume. If K is larger than the manipulation extent it may dilute the signal; if smaller it may bisect regions, a limitation acknowledged by the authors.
  • Attention internal dimensions L, L' = 128, 256
    Hidden dimensions of the gated attention projections in Eqs. (2) and (6). Chosen without reported sensitivity analysis.
  • Slice threshold tau_beta = 0.1
    Filters slices before assembling the 3D attention volume; hand-set in Section 3.4, no cross-validation reported.
  • Volume threshold tau_3D = 0.65
    Selects the top voxels of the normalized attention volume and defines the extracted bounding box. Used for all reported IoU and Pointing Game numbers, and hand-set in Section 3.4.
  • Gaussian smoothing sigma = 0.03 * max(K,H,W)
    Smoothing kernel width for the attention heatmap in Eq. (13). Hand-set.
assumptions (4)
  • standard math Gated attention aggregation from Ilse et al. (2018) is a valid permutation-invariant pooling that produces interpretable instance weights.
    The method builds on the attention-based MIL equations (Eqs. 2, 3, 6), which are cited from [14] and not re-derived.
  • domain assumption Tampered regions in CT volumes leave local detectable artifacts that survive across generative architectures and are spatially localized.
    The entire detection and localization approach assumes that patch-level and slice-level attention signals exist and transfer across unseen generators. This is the premise of the cross-generator protocol.
  • domain assumption The nodule axial coordinate in the dataset metadata accurately identifies all slices that contain manipulation.
    Stage 1 constructs slice-level labels from this metadata (Section 3.1). If the coordinate is noisy or the manipulation extends beyond the nodule, SliceMIL is trained on incorrect labels.
  • ad hoc to paper Attention weights alpha and beta are spatially faithful, i.e., high weight corresponds to manipulated voxels rather than to other discriminative cues such as anatomy or scanner artifacts.
    The XAI module (Section 3.4) treats attention as a localization map. The paper provides empirical localization results as support, but no proof that attention cannot focus on benign structures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HexMIL: Hierarchical Attention MIL for Ante-Hoc Explainable Detection of AI-Manipulated CT Volumes." pith.science (2026). https://pith.science/paper/F7G42ITT

@misc{pith2026260805101,
  author       = {Pith},
  title        = {Pith review of: HexMIL: Hierarchical Attention MIL for Ante-Hoc Explainable Detection of AI-Manipulated CT Volumes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7G42ITT}},
  note         = {Machine review of arXiv:2608.05101}
}
abstract

The emergence of medical deepfakes, i.e., medical images manipulated by deep generative models, poses a significant threat to clinical workflows. However, existing detectors suffer from two critical limitations: poor generalization to unseen generative architectures for manipulation detection and lack of interpretability. In this context, we present HexMIL (Hierarchical EXplainable Multiple Instance Learning), a mask-free medical deepfake detector that simultaneously addresses both limitations using only binary volume-level supervision. HexMIL decomposes each CT volume into a two-level hierarchy of patches and slices, aggregated via independent Gated Attention modules whose weights are directly combined into a full-resolution 3D attention volume that localizes the manipulated sub-region without any pixel-level annotation. Unlike post-hoc methods such as Grad-CAM, HexMIL's attention weights constitute the exact forward computation driving the classification decision, providing ante-hoc and structurally faithful spatial attribution. We evaluate HexMIL on M3DSynth and CT-GAN datasets under a rigorous cross-generator generalization protocol, training on a single generative architecture and testing on unseen ones. HexMIL outperforms all baselines by $+9.1$ AUC and $+9.4$ F1 in out-of-domain classification, and achieves the best average IoU and Pointing Game score in localization. Project page: opontorno.github.io/hexmil.

Figures

Figures reproduced from arXiv: 2608.05101 by the authors.

Figure 1
Figure 1. HexMIL graphical overview. (a) SliceMIL training: We train SliceMIL encoder module to distinguish between pristine and tampered slices based on their features (Section 3.2). (b) HexMIL Training: We train HexMIL model to identify and classify the presence of AI-manipulation within volume (Section 3.3). (c) Inference: A generic K’-slices volume is divided into K-slices sub-volumes during testing, and HexMIL processes … view at source ↗
Figure 2
Figure 2. XAI Module output. (a) Projection of attention maps and extracted Attention Volume along three axes. (b) 3D visualization of the attention heatmap (top) and the resulting 3D bounding box for spatial localization (bottom). Hierarchical heatmap construction. At inference time, both attention maps are jointly available for each sub-volume: the inter-slice weights βk from the Volume Gated Attention and the intra-slice p… view at source ↗
Figure 3
Figure 3. (a) Qualitative localization on manipulated CT slices (green boxes: ground truth), comparing our ante-hoc attention against post-hoc methods (Grad-CAM, Grad-CAM++) applied to the same HexMIL model. (b) Quantitative evaluation using IoU. (c), Pixel AUC, and Pointing Game. HexMIL is consistently strong across all axes. A plausible explanation is architectural: instead of a 2D slice-level evalu￾ation followed by a late… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of Gated (GA, top) and Self-Attention (SA, bottom) maps. Green boxes denote ground truth manipulations [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 42 canonical work pages

  1. [1]

    Toward medical deepfake detection: A comprehensive dataset and novel method

    Shuaibo Li, Zhaohu Xing, Hongqiu Wang, Pengfei Hao, Xingyu Li, Zekai Liu, and Lei Zhu. Toward medical deepfake detection: A comprehensive dataset and novel method. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 626–637. Springer, 2025

  2. [2]

    Generative adversar- ial nets.Advances in Neural Information Processing Systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversar- ial nets.Advances in Neural Information Processing Systems, 27, 2014

  3. [3]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InIn- ternational Conference on Machine Learning, pages 2256–2265. pmlr, 2015

  4. [4]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022

  5. [5]

    Advance- ments and challenges in deepfake medical imaging: gen- eration and detection techniques.Computers and Elec- trical Engineering, 132:111038, 2026

    Sayeli Dey, Smita Das, and Penumala Nani. Advance- ments and challenges in deepfake medical imaging: gen- eration and detection techniques.Computers and Elec- trical Engineering, 132:111038, 2026

  6. [6]

    Machine learn- ing based medical image deepfake detection: A com- parative study.Machine Learning with Applications, 8:100298, 2022

    Siddharth Solaiyappan and Yuxin Wen. Machine learn- ing based medical image deepfake detection: A com- parative study.Machine Learning with Applications, 8:100298, 2022

  7. [7]

    {CT-GAN}: Malicious tampering of 3d medi- cal imagery using deep learning

    Yisroel Mirsky, Tom Mahler, Ilan Shelef, and Yuval Elovici. {CT-GAN}: Malicious tampering of 3d medi- cal imagery using deep learning. In28th USENIX Secu- rity Symposium (USENIX Security 19), pages 461–478, 2019

  8. [8]

    M3dsynth: A dataset of medical 3d images with ai-generated local ma- nipulations

    Giada Zingarini, Davide Cozzolino, Riccardo Corvi, Giovanni Poggi, and Luisa Verdoliva. M3dsynth: A dataset of medical 3d images with ai-generated local ma- nipulations. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13176–13180. IEEE, 2024

Show all 48 references
  1. [9]

    Hybrid multimodal deepfake detection in medical images using convnextv2- tiny with triplet attention.Journal of Computational Science, page 102815, 2026

    R Dhanyalakshmi, Alexander Zakharov, Natalia Ro- manchuk, Cristina Popirlan, Claudiu-Ionut Popirlan, J Anitha, and D Jude Hemanth. Hybrid multimodal deepfake detection in medical images using convnextv2- tiny with triplet attention.Journal of Computational Science, page 102815, 2026

  2. [10]

    Forensic detection of generated mri imagery using au- toregressive modeling and frequency analysis

    Arpan Mahara, Naphtali Rishe, and Malek Adjouadi. Forensic detection of generated mri imagery using au- toregressive modeling and frequency analysis. InPro- ceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, pages 222–231, 2026

  3. [11]

    Back-in-time diffusion: Unsupervised detection of medical deepfakes.ACM Transactions on Intelligent Systems and Technology, 16(6):1–26, 2025

    Fred M Grabovski, Lior Yasur, Guy Amit, and Yisroel Mirsky. Back-in-time diffusion: Unsupervised detection of medical deepfakes.ACM Transactions on Intelligent Systems and Technology, 16(6):1–26, 2025

  4. [12]

    A comprehensive review of deepfakes in medical imaging: ethical concerns, detection tech- niques and future directions.Applied Computer Science, 21(2):139–153, 2025

    P Pradepan et al. A comprehensive review of deepfakes in medical imaging: ethical concerns, detection tech- niques and future directions.Applied Computer Science, 21(2):139–153, 2025

  5. [13]

    Jointdiffusion: Joint representation 9 learning for generative, predictive, and self-explainable ai in healthcare.Computerized Medical Imaging and Graphics, page 102619, 2025

    Joanna Kaleta, Paweł Skier´s, Jan Dubi´nski, Przemysław Korzeniowski, Tomasz Trzci´nski, Jakub M Tomczak, and Kamil Deja. Jointdiffusion: Joint representation 9 learning for generative, predictive, and self-explainable ai in healthcare.Computerized Medical Imaging and Graphics...

  6. [14]

    Attention-based deep multiple instance learning

    Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International Conference on Machine Learning, pages 2127–2136. PMLR, 2018

  7. [15]

    Grad-cam: Visual explanations from deep net- works via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep net- works via gradient-based localization. InProceedings of the IEEE International Conference on Computer Vision, pages 618–626, 2017

  8. [16]

    Grad- cam++: Generalized gradient-based visual explanations for deep convolutional networks

    Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad- cam++: Generalized gradient-based visual explanations for deep convolutional networks. In2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839–847. IEEE, 2018

  9. [17]

    Faceforensics++: Learning to detect manipulated facial images

    Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1–11, 2019

  10. [18]

    Exploiting visual artifacts to expose deepfakes and face manipulations

    Falko Matern, Christian Riess, and Marc Stamminger. Exploiting visual artifacts to expose deepfakes and face manipulations. In2019 IEEE Winter Applications of Computer Vision Workshops (WACVW), pages 83–92. IEEE, 2019

  11. [19]

    Mesonet: a compact facial video forgery detection network

    Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In2018 IEEE International Work- shop on Information Forensics and Security (WIFS), pages 1–7. IEEE, 2018

  12. [20]

    Cnn-generated images are surprisingly easy to spot

    Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8695–8704, 2020

  13. [21]

    Lever- aging frequency analysis for deep fake image recogni- tion

    Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Lever- aging frequency analysis for deep fake image recogni- tion. InInternational Conference on Machine Learning, pages 3247–3258. PMLR, 2020

  14. [22]

    Rethinking the up- sampling operations in cnn-based generative network for generalizable deepfake detection

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up- sampling operations in cnn-based generative network for generalizable deepfake detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

  15. [23]

    Dˆ 3: Scaling up deepfake detection by learning from discrepancy

    Yongqi Yang, Zhihao Qian, Ye Zhu, Olga Russakovsky, and Yu Wu. Dˆ 3: Scaling up deepfake detection by learning from discrepancy. InProceedings of the Com- puter Vision and Pattern Recognition Conference, pages 23850–23859, 2025

  16. [24]

    Deepfeaturex-sn: Generalization of deepfake de- tection via contrastive learning.Multimedia Tools and Applications, pages 1–20, 2025

    Orazio Pontorno, Luca Guarnera, and Sebastiano Bat- tiato. Deepfeaturex-sn: Generalization of deepfake de- tection via contrastive learning.Multimedia Tools and Applications, pages 1–20, 2025

  17. [25]

    On the detection of synthetic images generated by diffu- sion models

    Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffu- sion models. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), page...

  18. [26]

    On the exploitation of dct-traces in the generative- ai domain

    Orazio Pontorno, Luca Guarnera, and Sebastiano Bat- tiato. On the exploitation of dct-traces in the generative- ai domain. In2024 IEEE international Conference on Image Processing (ICIP), pages 3806–3812. IEEE, 2024

  19. [27]

    Mednet: Medi- cal deepfakes detection using an improved deep learn- ing approach.Multimedia Tools and Applications, 83(16):48357–48375, 2024

    Saleh Albahli and Marriam Nawaz. Mednet: Medi- cal deepfakes detection using an improved deep learn- ing approach.Multimedia Tools and Applications, 83(16):48357–48375, 2024

  20. [28]

    Image manipulation detection by multi- view multi-scale supervision

    Xinru Chen, Chengbo Dong, Jiaqi Ji, Juan Cao, and Xirong Li. Image manipulation detection by multi- view multi-scale supervision. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 14185–14193, 2021

  21. [29]

    Mantra-net: Manipulation tracing network for detec- tion and localization of image forgeries with anomalous features

    Yue Wu, Wael AbdAlmageed, and Premkumar Natara- jan. Mantra-net: Manipulation tracing network for detec- tion and localization of image forgeries with anomalous features. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9543–9552, 2019

  22. [30]

    Trufor: Leverag- ing all-round clues for trustworthy image forgery detec- tion and localization

    Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leverag- ing all-round clues for trustworthy image forgery detec- tion and localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages...

  23. [31]

    Dietterich, Richard H

    Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles.Artificial Intelligence, 89(1–2):31–71, 1997

  24. [32]

    Data-efficient and weakly supervised computational pathology on whole-slide images.Nature Biomedical Engineering, 5(6):555–570, 2021

    Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images.Nature Biomedical Engineering, 5(6):555–570, 2021. 10

  25. [33]

    Dual-stream multi- ple instance learning network for whole slide image clas- sification with self-supervised contrastive learning

    Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multi- ple instance learning network for whole slide image clas- sification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318–14328, 2021

  26. [34]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in Neural Informa- tion Processing Systems, 34:2136–2147, 2021

    Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in Neural Informa- tion Processing Systems, 34:2136–2147, 2021

  27. [35]

    Dtfd-mil: Double-tier feature distillation mul- tiple instance learning for histopathology whole slide image classification

    Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E Coupland, and Yalin Zheng. Dtfd-mil: Double-tier feature distillation mul- tiple instance learning for histopathology whole slide image classification. InProceedings of the IEEE/CVF Conference on Comput...

  28. [36]

    Learning deep features for discriminative localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 2921–2929, 2016

  29. [37]

    Sanity checks for saliency maps.Advances in Neural Informa- tion Processing Systems, 31, 2018

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps.Advances in Neural Informa- tion Processing Systems, 31, 2018

  30. [38]

    Attention is all you need.Advances in Neural Information Processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing systems, 30, 2017

  31. [39]

    Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The lung image database consortium (lidc) and image database resource initiative (idri): a c...

  32. [40]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition, pages 770–778, 2016

  33. [41]

    Frequency-aware deep- fake detection: Improving generalizability through fre- quency space domain learning

    Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deep- fake detection: Improving generalizability through fre- quency space domain learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5052–5060, 2024

  34. [42]

    Xception: Deep learning with depth- wise separable convolutions

    François Chollet. Xception: Deep learning with depth- wise separable convolutions. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 1251–1258, 2017

  35. [43]

    Localization of deep in- painting using high-pass fully convolutional network

    Haodong Li and Jiwu Huang. Localization of deep in- painting using high-pass fully convolutional network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8301–8310, 2019

  36. [44]

    A closer look at spatiotemporal convolutions for action recognition

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition, pages 6450–6459, 2018

  37. [45]

    Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? InProceedings of the IEEE con- ference on Computer Vision and Pattern Recognition, pages 6546–6555, 2018

    Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? InProceedings of the IEEE con- ference on Computer Vision and Pattern Recognition, pages 6546–6555, 2018

  38. [46]

    Vivit: A video vision transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vision transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 6836–6846, 2021

  39. [47]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

  40. [48]

    Neural machine translation by jointly learning to align and translate

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- gio. Neural machine translation by jointly learning to align and translate. In3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 11

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.