{"id":"fd2003d8-8c2c-4ebb-a545-99768b522435","arxiv_id":"2412.20156","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A distilled transformer with mixture-of-experts, local enhancement, and attention scaling is reported to improve face forgery detection on five deepfake benchmarks.","lead":"This paper presents a transformer-based face forgery detector that combines local and global cues, uses self-generated soft labels during training, and adds a lightweight attention-scaling module to keep attention maps diverse. The authors report gains over several baselines on five deepfake datasets, but the headline state-of-the-art claim is weakened by an incomplete baseline comparison and the absence of error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 itself contradicts the SOTA claim: on Celeb-DF, DTN AUC is 75.32 versus 95.14 for Guan et al., 86.40 for ResNet34, and 76.13 for FoCus.","rationale":"The reader's filed weakest_assumption concerns whether DSD soft labels generalize, but the load-bearing problem is more direct: the paper's own Table 1 contradicts the headline SOTA claim on Celeb-DF. The reader did note this contradiction in the rationale, so my stress-test agrees with the overall conditional verdict rather than moving it. I keep the verdict at CONDITIONAL rather than REJECT because the architecture and internal ablations are plausible, and the central claim could be corrected by removing or substantially qualifying the SOTA statement and by adding a properly controlled comparison against Guan et al. and other strong baselines. However, the current version should not be accepted without that correction, and the concrete protocol check should be performed before any positive claims about cross-dataset generality are retained.","tokens_in":19119,"tokens_out":8118,"duration_ms":86047,"concrete_test":"Run the official code or checkpoint of Guan et al. [36] and the proposed DTN under the exact protocol of Sections 5.1-5.2: train on FF++ HQ with the official split, use dlib face cropping at 256x256 with the same frame sampling and augmentation, and evaluate both models on the same Celeb-DF test frames. If Guan et al. remains near 95.1 AUC while DTN remains near 75.3, the SOTA claim on Celeb-DF is false; if the 95.14 value came from a different protocol, report both methods under one unified protocol and revise the claim accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Abstract and Section 5.3 claim that the proposed method 'surpasses the state of the arts on five deepfake datasets' and is 'more general to various deepfake datasets.' The paper's own Table 1 directly refutes this on the Celeb-DF cross-dataset benchmark: after training on FF++, DTN obtains 75.32 AUC, while Guan et al. [36] obtains 95.14, ResNet34 [35] obtains 86.40, and FoCus [34] obtains 76.13. A 20-point gap against a published method on a named dataset means the central generality claim, as written, is false. The DSD self-distillation mechanism is not the weakest point: even if soft labels generalize exactly as hoped, the resulting model is still not state-of-the-art on Celeb-DF. Secondary issues such as the mismatch between the implementation (B=2 experts) and the ablation recommending B=5, the lack of error bars, and the selection of student 3 using cross-dataset test AUC compound the problem but need not be resolved for the headline concern to land.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DTN, a hybrid CNN-transformer network for face forgery detection. The architecture consists of a noisy mixture-of-experts (MoE) module, a locally-enhanced vision transformer (LEVT) with a multi-attention scaling (MAS) module, and a deepfake self-distillation (DSD) scheme in which a frozen teacher generates soft labels for same-architecture student models. Experiments report within-dataset results on FF++, cross-dataset results on Celeb-DF, DFDC, DF-1.0, and DFD, cross-manipulation generalization, robustness to image corruptions, and component ablations. The authors claim that the method surpasses the state of the art on five deepfake datasets and generalizes more broadly than existing detectors.","tokens_in":19412,"tokens_out":6398,"duration_ms":65399,"significance":"If substantiated, the paper would offer useful components: MAS is plug-and-play and is shown to improve several transformer backbones, DSD produces consistent internal gains on both CNN and transformer models, and the cross-manipulation experiments are informative. The ablation tables are a genuine strength because they isolate each module. However, the central state-of-the-art claim is not supported by the paper's own Table 1, and the baseline comparison protocol mixes reproduced and literature-reported numbers. The lack of error bars and the selection of the final student using test-set AUC further weaken the empirical claims. The result is potentially salvageable, but the current evidence does not justify the headline claims.","major_comments":[{"comment":"The claim that DTN 'surpasses the state of the arts on five deepfake datasets' is contradicted by the paper's own cross-dataset table. On Celeb-DF (after training on FF++), DTN obtains AUC 75.32, while Guan et al. [36] reports 95.14, ResNet34 [35] reports 86.40, and FoCus [34] reports 76.13. The text's statement that DTN 'exceeds the existing state-of-the-art approach, RECCE, by 6.6% AUC' selects RECCE (68.71) as the reference point even though three other rows in the same table are higher. As written, the state-of-the-art claim and the 'more general to various deepfake datasets' claim are false; the authors must either revise these claims or provide a dataset-by-dataset comparison under a common protocol.","section":"Abstract, Section 5.3, Table 1"},{"comment":"The comparison mixes numbers reproduced by the authors (marked with †) with numbers taken from other papers, without a common evaluation protocol. Face cropping, frame sampling, image size, compression level, and training-set composition differ across published baselines. For example, DTN uses dlib crops at 256x256 on FF++ HQ, while the cited numbers for Guan et al., ResNet34, and FoCus come from their original papers and may use different preprocessing. Cross-paper AUC differences, including the 20-point gap on Celeb-DF, are not meaningful unless all methods are evaluated under the same protocol. The authors should re-run the baselines under one protocol or explicitly restrict the comparison to the † rows.","section":"Table 1, Section 5.2"},{"comment":"All results are single-run ACC/AUC values with no standard deviations, error bars, or significance tests. Several reported differences are small; for example, on FF++ test AUC, DTN's 99.70 differs from Yu et al.'s 99.55 by 0.15, and many ablation increments are 0.5–1.0 AUC. Without repeated trials (at least 3–5 seeds), these differences cannot be distinguished from training noise. The authors should report mean ± std and, where appropriate, paired tests or confidence intervals.","section":"All experimental tables (Tables 1–10)"},{"comment":"The pseudocode returns the original teacher rather than the trained student. Line 3 stores a copy of the teacher in DTNcopy_tea, and line 22 sets DTNsota_stu to this unchanged copy. This contradicts the text and Table 10, where DTN-Student 3 is the final model. The algorithm should return the best saved student (e.g., the model saved at line 11), and the loop conditions involving Lstu, Ltea, z, and t should be clarified.","section":"Algorithm 1"},{"comment":"The number of DSD generations is selected using cross-dataset test AUC: the text states 'so we chose student 3 as our model' based on Table 10, which reports AUC on Celeb-DF, DFDC, DF-1.0, and DFD test sets. Selecting the generation by test-set performance leaks test information into model selection and makes the reported cross-dataset numbers optimistic. The stopping criterion should be defined on a validation set, or the hard-label loss rule in Algorithm 1 should be followed without consulting the test set.","section":"Section 5.4, Table 10"},{"comment":"The final model uses B=2 experts in the MoE module (§5.2), but Figure 6(d) shows that AUC is maximized with B=5 when Gaussian noise is used. The paper should either adopt the empirically optimal B, or justify fixing B=2 (for example, by a computational budget) and state the corresponding trade-off.","section":"Section 5.4, Figure 6(d), Section 5.2"}],"minor_comments":[{"comment":"The text refers to the 'SFG module' in the paragraph after Eq. (5), but the module is called STG everywhere else; this should be corrected.","section":"Section 3.2"},{"comment":"The notation is inconsistent: δ is described in Eq. (6) as softmax with temperature, but no temperature parameter appears, and Eq. (20) also uses δ for softmax. Please define δ, σ, and the temperature explicitly.","section":"Eqs. (6), (20)"},{"comment":"The text reports a '+3.4%' gain from introducing STG, but Table 4 shows smaller absolute differences (e.g., +1.93 ACC on FF++). Please specify which metric and which dataset the 3.4% refers to.","section":"Section 5.4, Table 4"},{"comment":"The second method is labeled 'Ous'; this should be 'Ours'.","section":"Table 7"},{"comment":"The first column header reads 'Loss function', but the rows contain transformer depth L; the header should be 'Depth L'.","section":"Table 9"},{"comment":"The class centers c_y and c_{1-y} are not defined; please specify their dimensionality and how they are updated during training.","section":"Eq. (3)"},{"comment":"The curves in Figures 6(a) and 6(b) lack numerical axis values and error bars, which makes the claimed differences difficult to verify quantitatively.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper's architectural ideas may be useful, but the central state-of-the-art claim is directly contradicted by its own Table 1 on Celeb-DF, and the evaluation protocol does not support direct baseline comparisons. This is publication-blocking but fixable: the authors can reposition the claims, provide a controlled comparison or re-run published baselines, add repeated trials, fix Algorithm 1, and state a valid model-selection rule. I therefore recommend major revision rather than rejection, because the internal ablations and the plug-and-play MAS evidence suggest the method has merit. The editor may also wish to ask for a clear statement of the evaluation protocol and, if possible, code release."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2412.20156, DTN for face forgery detection. The paper is a mix of genuinely useful components and an overblown headline. The one thing to know first: the abstract and conclusion claim the method 'surpasses the state of the arts on five deepfake datasets,' but Table 1 of the paper itself shows DTN at 75.32 AUC on Celeb-DF cross-dataset versus 95.14 for Guan et al. and 86.40 for ResNet34. That's a twenty-point gap. The claim as written is false. Everything else is secondary.\n\nWhat is actually new and what the paper does well: the MAS module - per-head learnable scaling factors on attention logits - is simple, appears new, and they show it transfers to several transformers (ViT, CViT, EfficientViT, CrossEfficientViT) with consistent gains. That's the most reusable piece. The deepfake self-distillation scheme is also a new application, and the ablation shows it helps across datasets. The corruption-robustness experiments are thoughtful; they even replace the Gaussian noise in MoE with uniform noise to avoid leakage.\n\nSoft spots, in proportion. The SOTA claim is the load-bearing one and it doesn't hold on Celeb-DF. The text in Section 5.3 says 'surpasses most methods,' which is closer to true, but the abstract and conclusion overstate. The implementation uses B=2 experts, while their own Figure 6(d) shows the maximum at five experts; if they used five, the numbers would likely be better, but the choice needs explanation. No error bars, single runs. They select student 3 based on cross-dataset test AUC, which is test-set selection. Table 3's final row matches the full model with DSD loss from Table 4, so it's unclear whether the component ablation is confounded with self-distillation. No code is released.\n\nThis is a paper that deserves a serious referee, not a desk reject. The components are plausible, the internal ablations are consistent, and the flaws are addressable in revision. The main fix is to tone down the SOTA language and add disclosure about selection and the expert count. I'd send it to review.\n\nFor my own work, I wouldn't cite the SOTA numbers, but the MAS module is worth a mention as related work.","headline":"The method pieces are worth building on, but the SOTA claim is contradicted by the paper's own Table 1 on Celeb-DF.","tokens_in":19909,"tokens_out":5544,"would_cite":true,"duration_ms":54282,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A distilled CNN-transformer hybrid with per-head attention scaling and self-distilled soft labels is claimed to outperform prior face-forgery detectors on five deepfake datasets.","keywords":["face forgery detection","deepfake detection","vision transformer","knowledge distillation","mixture of experts","attention collapse","cross-dataset generalization","soft labels"],"falsifier":"Train the full DTN on FF++ and also train an identical DTN with the DSD loss replaced by hard-label cross-entropy only, then evaluate both on a deepfake test set generated by a diffusion-based face-swapping method not included in the five datasets; if the self-distilled model is not more accurate on that unseen set, the paper's generalizability claim is falsified.","tokens_in":18937,"feed_emoji":"🎭","tokens_out":6663,"duration_ms":63192,"temperature":0.7,"pith_summary":"The paper sets out to show that face-forgery detection improves when a detector combines local and global forgery cues instead of relying on either alone. Its proposed distilled transformer network (DTN) uses a mixture-of-experts module to collect diverse local artifacts, a locally-enhanced vision transformer to model global relations refined by local context, a multi-attention scaling module to stop transformer attention from collapsing in deeper layers, and a deepfake self-distillation scheme that generates soft labels automatically. The authors report that DTN surpasses state-of-the-art detectors on FF++, Celeb-DF, DFDC, DeeperForensics-1.0, and DFD, and that it generalizes to unseen manipulations and image corruptions better than prior methods. A reader should care because the practical bottleneck in deepfake detection is cross-dataset and cross-manipulation generalization, which is exactly what the paper claims to address.","feed_headline":"Deepfake detector beats state of the art on five datasets","feed_subtitle":"Locally enhanced transformer with self-distilled soft labels generalizes across unseen forgeries.","key_machinery":"The argument is carried by four interacting components. The MoE module produces several expert attention maps from noisy copies of the backbone features and combines them via learnable scale factors and max-pooling, which is meant to capture diverse local forgery patterns. The LEVT module partitions tokens into heads, applies self-attention, then passes the aggregated global features through a local convolutional connection so that global relations are refined by local context. The MAS module is the remedy for attention collapse: each attention head's pre-softmax scores are multiplied by its own sigmoid-bounded learnable weight, so different heads are encouraged to focus on different regions as the network deepens. The DSD scheme wraps the whole model in an iterative teacher-student loop where soft labels are generated automatically by the frozen teacher and combined with cross-entropy and contrastive-center losses; each student becomes the next teacher until performance stops improving.","core_discovery":"The central claim is that a face-forgery detector can be made both more accurate and more general by combining four mechanisms in one network: a mixture-of-experts module that adds random Gaussian noise and integrates several expert attention maps to collect diverse local forgery cues; a locally-enhanced vision transformer (LEVT) that refines global attention features with a local convolutional connection; a multi-attention scaling (MAS) module that assigns each attention head its own learnable scale so deeper transformer blocks do not collapse onto identical attention patterns; and a deepfake self-distillation (DSD) scheme in which a frozen teacher generates soft labels and the student, trained to match those labels plus the hard labels, becomes the teacher of the next generation. The authors report that the resulting DTN reaches 99.70% AUC on FF++ and 75.32%, 80.01%, and 78.77% AUC on Celeb-DF, DFDC, and DeeperForensics-1.0 when trained only on FF++, and that it generally beats prior detectors across all five datasets and across unseen manipulation types.","pith_inferences":["Beyond the paper, MAS acts only on attention computation, so it is a candidate drop-in fix for attention collapse in transformer models for other fine-grained visual tasks, though the paper only tests it on face-forgery backbones.","Beyond the paper, the paper's own Limitations section admits that some teacher soft labels are wrong or ambiguous; filtering or reweighting low-confidence soft labels is a natural extension that could make the distillation chain more reliable and is testable on the same five datasets.","Beyond the paper, the paper names diffusion-model forgeries as future work; a fair test of the generality claim would be training on FF++ only and evaluating on a diffusion-based face-swap test set, where the soft-label chain might help or might amplify dataset-specific bias."],"forward_implications":["If DTN is right, a single model trained on FF++ should keep high detection accuracy on Celeb-DF, DFDC, DeeperForensics-1.0, and DFD, making it usable without per-dataset retraining.","MAS can be added to existing transformer detectors such as ViT, CViT, EfficientViT, and CrossEfficientViT at negligible parameter and FLOP cost and still raise their cross-dataset AUC, so attention-collapse control transfers across architectures.","The DSD scheme improves CNN-based Xception and transformer-based EfficientViT as well as the full DTN, implying soft-label self-distillation is a detector-agnostic training strategy.","Training on one FF++ manipulation and testing on the other three, DTN achieves the highest AUC when trained on Face2Face or NeuralTextures, so it should detect manipulation types not seen during training."],"supporting_citations":[{"why":"Defines the CViT hybrid CNN-transformer baseline that DTN extends and that motivates the attention-collapse problem.","marker":"[1]"},{"why":"Supplies the vision transformer architecture and global-dependency modeling that LEVT is built on.","marker":"[5]"},{"why":"Provides the FF++ dataset used for training and within-dataset evaluation, plus the Xception baseline.","marker":"[29]"},{"why":"Provides DeeperForensics-1.0 as a cross-dataset target and the corruption protocol used for robustness tests.","marker":"[8]"},{"why":"Provides Celeb-DF as a cross-dataset evaluation target.","marker":"[31]"},{"why":"Provides DFDC as a cross-dataset evaluation target.","marker":"[30]"},{"why":"Defines RECCE, the state-of-the-art baseline DTN claims to beat by 6.6% AUC on Celeb-DF.","marker":"[10]"},{"why":"Defines the self-blended image baseline for generic forgery representation that DTN claims to surpass.","marker":"[12]"},{"why":"Motivates the local connection layer in LEVT, which refines global attention with local neighborhood information.","marker":"[28]"},{"why":"Provides the strongest cross-manipulation baseline on DF and FS training settings that DTN compares against in generalization experiments.","marker":"[38]"}],"fun_headline_variants":["Fusing local and global cues with self-distillation catches more deepfakes","Self-distilled soft labels help transformer detect unseen forgery types","Multi-attention scaling prevents attention collapse in deepfake transformers","Mixture-of-experts gathers robust cues for face forgery detection","Distilled transformer with local attention beats state-of-the-art deepfake detectors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cross-dataset gain rests on the assumption that soft labels produced by a teacher trained on the same data encode transferable inter-class similarity knowledge, and that each distillation generation passes on better knowledge rather than the teacher's mistakes; if the teacher's soft labels are biased toward the training distribution, the self-distillation chain could reinforce rather than remove that bias.","fun_headline_variants_meta":{"raw":{"variants":["Fusing local and global cues with self-distillation catches more deepfakes","Self-distilled soft labels help transformer detect unseen forgery types","Multi-attention scaling prevents attention collapse in deepfake transformers","Mixture-of-experts gathers robust cues for face forgery detection","Distilled transformer with local attention beats state-of-the-art deepfake detectors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001072,"raw_usage":{"total_tokens":4517,"prompt_tokens":997,"completion_tokens":3520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":3429}},"tokens_in":613,"tokens_out":3520,"duration_ms":24776,"temperature":1.0,"reasoning_tokens":3429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:30:32.334862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full DTN on FF++ and also train an identical DTN with the DSD loss replaced by hard-label cross-entropy only, then evaluate both on a deepfake test set generated by a diffusion-based face-swapping method not included in the five datasets; if the self-distilled model is not more accurate on that unseen set, the paper's generalizability claim is falsified.","supporting_citations":[{"cited_title":"Dosovitskiy, L","cited_arxiv_id":null,"evidence_quote":"Supplies the vision transformer architecture and global-dependency modeling that LEVT is built on."},{"cited_title":"Rössler, D","cited_arxiv_id":null,"evidence_quote":"Provides the FF++ dataset used for training and within-dataset evaluation, plus the Xception baseline."},{"cited_title":"Jiang, R","cited_arxiv_id":null,"evidence_quote":"Provides DeeperForensics-1.0 as a cross-dataset target and the corruption protocol used for robustness tests."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Celeb-DF as a cross-dataset evaluation target."},{"cited_title":"Dolhansky, J","cited_arxiv_id":null,"evidence_quote":"Provides DFDC as a cross-dataset evaluation target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines RECCE, the state-of-the-art baseline DTN claims to beat by 6.6% AUC on Celeb-DF."},{"cited_title":"Shiohara, T","cited_arxiv_id":null,"evidence_quote":"Defines the self-blended image baseline for generic forgery representation that DTN claims to surpass."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the local connection layer in LEVT, which refines global attention with local neighborhood information."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the strongest cross-manipulation baseline on DF and FS training settings that DTN compares against in generalization experiments."}],"review_version":1}