{"id":"7863f1d8-cd05-4d10-8b29-24e084cd813d","arxiv_id":"2504.17223","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A spatial-frequency deepfake detector with hierarchical cross-modal fusion reports state-of-the-art accuracy on FaceForensics++ and improved cross-dataset AUC on Celeb-DF and DFDC.","lead":"A new deepfake detector combines pixel-level features, local frequency analysis, and global statistical checks to catch synthetic faces that fool other detectors. It reports higher accuracy and better generalization across three standard benchmarks, which matters for anyone building automated content moderation or forensic tools.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's own numbers contradict the 'outperforms SOTA in generalizability' claim: SFDG reports 75.83 AUC on Celeb-DF(v2) vs 74.68 for SFCL-HCMF, so the central claim is not supported without a unified re-run.","rationale":"The reader's verdict is CONDITIONAL, and my stress-test supports that verdict: the paper has a genuine architectural contribution but its headline claim is not fully supported by the reported experiments. The reader's weakest_assumption focused on the sufficiency of the hand-designed SIDA statistics. My concern is related but more direct: the paper's own Table 2 shows SFDG beating SFCL-HCMF on Celeb-DF(v2), which contradicts the abstract's claim of outperforming SOTA in generalizability. This is a concrete, falsifiable issue rather than a speculative assumption about internal feature statistics. I do not recommend REJECT because the contradiction can be resolved by a rigorous, protocol-matched comparison and by softening the claim to 'comparable or better.' The lack of code and error bars further supports keeping the CONDITIONAL verdict. I mark agreement as 'partial' because the reader identified a different weakest assumption, though both point to the same general concern that the generalization evidence is not robust.","tokens_in":12218,"tokens_out":5026,"duration_ms":52766,"concrete_test":"Re-run SFCL-HCMF and SFDG on the same FF++(c23) train/test split, with identical face cropping, frame sampling, frame-level AUC computation, and report bootstrap 95% confidence intervals over at least 5 training seeds on Celeb-DF(v2) and DFDC. Also add a SIDA on/off ablation under the same cross-dataset protocol. If SFDG significantly beats SFCL on Celeb-DF, the generalizability claim should be downgraded from 'outperforms' to 'comparable'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 4.5 claim that SFCL-HCMF outperforms state-of-the-art deepfake detection in both accuracy and generalizability. However, Table 2 directly undercuts this claim on cross-dataset generalization: SFDG [42] achieves 75.83 AUC on Celeb-DF(v2), while SFCL-HCMF achieves 74.68 AUC. On DFDC the gap is only 0.07 AUC (73.64 vs 73.71). Since the paper's own reported numbers show at least one SOTA baseline beating the proposed method on a cross-dataset benchmark, the blanket 'outperforms' claim is not supported by the presented evidence. Moreover, the baselines in Table 2 appear to be taken from their original papers with potentially different protocols (training splits, face cropping, frame sampling, video-level vs frame-level metrics), so the comparison may not be apples-to-apples. The ablation study in Table 3 validates SIDA only on FF++(c23); no cross-dataset ablation is provided for the component specifically claimed to improve scale-invariant generalization. Without a unified evaluation protocol, error bars, or multiple seeds, the central claim of generalizability superiority remains unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deepfake detection framework, SFCL-HCMF, that combines a spatial EfficientNet branch with a frequency branch built on block-wise DCT. The frequency branch has a local stream that uses a 3D-convolution SBCM and a CNN-F to model intra- and inter-block DCT correlations, and a global stream (SIDA) that computes row, column, and intra-block differential statistics over DCT sub-band features. A hierarchical cross-modal fusion mechanism (FAAE in shallow layers and HCMA in deep layers) integrates the two modalities. The method is evaluated on FaceForensics++ for in-dataset performance and on Celeb-DF(v2) and DFDC for cross-dataset generalization, with reported AUCs of 99.58% on FF++(c23), 94.21% on FF++(c40), 74.68% on Celeb-DF(v2), and 73.71% on DFDC. The central claim is that this architecture outperforms state-of-the-art methods in both accuracy and generalizability.","tokens_in":12486,"tokens_out":6130,"duration_ms":53355,"significance":"The architecture is coherent and the paper gives detailed equations for each module; the ablation studies in Tables 3 and 4 separate the contributions of SBCM, SIDA, FAAE, and HCMA, which is a strength. An effective local/global frequency representation combined with cross-modal fusion would be a useful addition to deepfake detection. However, the reported evidence is not yet sufficient to support the central claim. The cross-dataset comparison in Table 2 is not controlled, one baseline (SFDG) beats the proposed method on Celeb-DF(v2), and no uncertainty estimates or code/data are provided, so the reported superiority over state-of-the-art is unverified. The contribution would be significant if the claims were supported by a unified evaluation with error bars.","major_comments":[{"comment":"The statement in the abstract and Section 4.5 that the method outperforms state-of-the-art deepfake detection methods in both accuracy and generalizability is not supported by Table 2: SFDG reports 75.83 AUC on Celeb-DF(v2) while SFCL-HCMF reports 74.68, and on DFDC the reported gap is only 0.07 AUC (73.64 vs 73.71). The authors need either a unified re-run of all baselines under the same protocol or a revised, more qualified claim.","section":"Section 4.5, Table 2"},{"comment":"The cross-dataset numbers appear to be transcribed from the original baseline papers, with no evidence that training splits, face cropping, frame sampling, image-level versus video-level scoring, or post-processing were identical. Without a controlled protocol, the comparisons in Table 2 are not apples-to-apples and cannot support the generalizability claim. Please provide a protocol description and, ideally, re-run the baselines under the authors' evaluation setup.","section":"Section 4.5, Table 2"},{"comment":"The ablation study validates SIDA only on FF++(c23); there is no cross-dataset ablation that measures how much SIDA contributes to the Celeb-DF/DFDC generalization results. Since SIDA is the component presented in Section 3.2.1 as the scale-invariant generalization mechanism, the claimed benefit to generalization is currently unsupported.","section":"Section 4.6, Table 3"},{"comment":"The four hand-designed statistics (mean, std, skewness, kurtosis) are the entire global SIDA representation, but no justification or comparison to alternative aggregations is given, and there is no robustness test under resizing or re-compression. Because the cross-dataset gains are attributed to this branch, the choice of these particular statistics is load-bearing and needs empirical support.","section":"Section 3.2.1, Equations (4)-(8)"},{"comment":"There is no code, data, or model release, and all results are single-run image-level Acc/AUC without standard deviations or multiple seeds. Since several reported margins over baselines are below 0.1 AUC, the quantitative claims cannot be checked or assessed for statistical significance.","section":"Section 4.1, Tables 1-2"}],"minor_comments":[{"comment":"The method name is written SFCL-HCMF in the title and abstract but SPCL-HCMF in Tables 1 and 2; please make the notation consistent.","section":"Title and Tables 1-2"},{"comment":"The sentence comparing Two-branch and SFDG and claiming improved robustness is contradicted by Table 2 on Celeb-DF(v2) and should be corrected or qualified.","section":"Section 4.5"},{"comment":"The paragraph beginning with '2D CNNs exhibit inherent limitations' is repeated verbatim twice in Section 3.2.2; please remove the duplicate.","section":"Section 3.2.2"},{"comment":"The normalization uses H times W, but the feature maps in these equations have dimensions such as C by 64 by (H/8) by ((W/8)-1); please clarify the dimensions and normalization over the actual number of elements.","section":"Equations (4)-(7)"},{"comment":"References [5] and [6] appear to be the same paper; please consolidate and renumber. Also, the asterisks on MAD and M2TR in Tables 1 and 2 are not explained in the captions.","section":"References [5] and [6]"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope, and I do not see a fundamental flaw in the architecture; the main problems are empirical support and reproducibility. The authors should be asked to provide a controlled comparison and uncertainty estimates. I would not consider rejection unless the revised claims continue to be contradicted by their own table."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a real engineering effort: block-wise DCT with 3D spectral-band convolutions, a global differential-statistics branch (SIDA), and hierarchical cross-modal fusion. That combination is new, and the in-dataset numbers on FF++ (99.58 AUC on c23, 94.21 on c40) are strong. The ablations show each component contributes, which is good practice.\n\nBut the core generalizability claim does not survive contact with Table 2. SFDG gets 75.83 AUC on Celeb-DF(v2); your method gets 74.68. On DFDC the gap is 0.07. The abstract says \"outperforms state-of-the-art in both accuracy and generalizability,\" while Section 4.5 quietly downgrades to \"superior or comparable.\" Those are different claims, and the table supports only the weaker one.\n\nThe other major issue is verifiability. No code, no trained models, no error bars, no multiple seeds. The baselines in Table 2 are taken from their original papers, so training protocols, face crops, and frame sampling may differ. That makes the cross-dataset comparison unreliable even when the numbers look favorable. The SIDA ablation is only on FF++(c23), so the component explicitly designed for scale-invariant generalization is never tested cross-dataset. That is a gap.\n\nMinor but real: the method name appears as \"SPCL-HCMF\" in Table 1, and there is a \"Glocal\" typo in Section 4.6. References [5] and [6] are identical. These are not load-bearing, but they lower confidence in the manuscript's preparation.\n\nWho is this for? Someone working on deepfake detection who wants to see a plausible new frequency-domain architecture. This is a useful idea paper, not a verified result. I would send it to peer review, but with the expectation that the authors must release code and models, add error bars and multiple seeds, and either fix or explicitly temper the abstract claim. The claim problem is fixable; the missing artifacts are a demand, not a death sentence.","headline":"A plausible but unverified deepfake detection architecture whose own table contradicts the headline generalizability claim, and with no code or error bars to back the numbers.","tokens_in":13020,"tokens_out":1536,"would_cite":false,"duration_ms":15989,"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":"This paper claims that a deepfake detector which treats frequency evidence as a first-class modality, extracting block-wise DCT artifacts and global spectral statistics, reaches 99.58% AUC on FF++ (c23) and generalizes across datasets…","keywords":["deepfake detection","discrete cosine transform","frequency-domain analysis","cross-modal fusion","spatial-frequency learning","face forgery detection","generalization"],"falsifier":"Train SFCL-HCMF on FF++ (c23) and evaluate it on forgeries produced by a generator family absent from FF++, Celeb-DF, and DFDC, with compression or post-processing not used in training; if the frame-level AUC falls to the level of the spatial-only backbone, the generalization claim is falsified. A cheaper check: compute the 2304-dimensional SIDA descriptor on real and fake images after downscaling to 128x128 and see whether the real and fake distributions still separate cleanly.","tokens_in":11989,"feed_emoji":"🎭","tokens_out":6583,"duration_ms":57221,"temperature":0.7,"pith_summary":"This paper argues that deepfake detection should not choose between spatial and frequency evidence: a detector that extracts both and fuses them at multiple depths can recognize forgeries it has never seen. It proposes SFCL-HCMF, which runs an EfficientNet spatial stream alongside a frequency stream built on block-wise DCT, and reports 99.58% AUC on FF++ (c23), 94.21% on the heavily compressed FF++ (c40), and cross-dataset AUCs of 74.68% on Celeb-DF (v2) and 73.71% on DFDC after training only on FF++. The authors' point is that frequency-native artifacts, especially global distribution patterns, are the missing signal that keeps detectors from generalizing.","feed_headline":"Frequency-spatial fusion lifts deepfake detection to 99.58% AUC","feed_subtitle":"The same model also reaches 74.68% AUC on Celeb-DF and 73.71% on DFDC after training only on FF++.","key_machinery":"The load-bearing object is the block-wise DCT representation: an image is converted to YCbCr, divided into 8x8 blocks, DCT-transformed, and each block's coefficients are flattened by zigzag scanning into a 4D tensor that preserves both spatial block positions and frequency bands. On this tensor, the local branch (SBCM plus a modified CNN) uses 3D convolutions along the frequency dimension to model intra-block coefficient correlations and inter-block transitions; the global branch (SIDA) applies horizontal, vertical, and intra-block differentials and compresses them into a 2304-dimensional vector of mean, std, skewness, and kurtosis; and the hierarchical cross-modal fusion modules (FAAE and HCMA) inject frequency context into shallow spatial features and dynamically gate deep fused features.","core_discovery":"The paper's central claim is that a deepfake detector can be made generalizable by treating frequency information as a first-class modality rather than a filter applied to spatial features. SFCL-HCMF does this in three moves: block-wise 8x8 DCT with zigzag reordering so that DCT coefficients keep spatial-frequency structure; a local branch that convolves along the spectral-band dimension to catch intra-block and inter-block anomalies; and a global branch that computes row, column, and intra-block differences of DCT blocks at original resolution and summarizes them with four statistics (mean, std, skewness, kurtosis) into a 2304-dimensional descriptor. A hierarchical fusion then uses frequency-localized attention on shallow spatial features and dynamic gating on deep fused features. The reported result is that this architecture beats prior detectors on FF++ at both compression levels and, after training only on FF++, transfers to Celeb-DF (v2) and DFDC better than or comparably to the strongest baselines.","pith_inferences":["Beyond the paper: because the SIDA descriptor is computed from DCT blocks at the original resolution and does not depend on resizing, it may also serve as a lightweight, resolution-robust forensic fingerprint; a direct test would be to use it as a standalone detector on unseen generators.","Beyond the paper: the method's face cropping and fixed 380x380 input for the spatial and local-frequency streams leave open how much of the gain survives unaligned or very-low-resolution faces; that boundary is testable and not settled by the reported benchmarks.","Beyond the paper: the same block-wise DCT plus differential-statistics recipe is not face-specific beyond the crop, so it could plausibly be applied to synthetic-image detection for non-face content such as GAN-generated textures or documents; that is an extension, not a claim of the paper."],"forward_implications":["Frequency-domain clues should be treated as a native input stream in forensic detectors, not merely as augmentation of spatial features.","Detectors trained on one dataset can transfer to unseen manipulation types and compression levels, making real-world deployment more plausible.","The compact global statistical descriptor offers a resolution-preserving representation that could be used alongside spatial features for efficient screening.","The hierarchical fusion pattern, shallow attention plus deep gating, provides a reusable design for other multimodal forgery detection tasks."],"supporting_citations":[{"why":"Supplies the FF++ benchmark and compression levels c23/c40 used for in-dataset evaluation.","marker":"[36]"},{"why":"Supplies Celeb-DF v2 as the primary cross-dataset generalization target.","marker":"[25]"},{"why":"Supplies DFDC as the large-scale cross-dataset generalization target.","marker":"[10]"},{"why":"Supplies the YCbCr conversion and block-wise DCT preprocessing adopted for the frequency stream.","marker":"[44]"},{"why":"Serves as the frequency-aware baseline F3-Net whose cross-dataset limitations motivate the design.","marker":"[35]"},{"why":"Serves as the multi-attentional spatial baseline MAD in both in-dataset and cross-dataset tables.","marker":"[49]"},{"why":"Serves as the multi-modal transformer baseline M2TR compared in the evaluation.","marker":"[40]"},{"why":"Serves as the spatial-frequency relation baseline SFDG, the closest competitor on cross-dataset transfer.","marker":"[42]"},{"why":"Serves as the two-branch recurrent baseline that also fuses frequency and spatial evidence.","marker":"[30]"},{"why":"Supplies the EfficientNet backbone pre-trained on ImageNet used for the spatial stream.","marker":"[38]"}],"fun_headline_variants":["Frequency-first fusion beats deepfake baselines on unseen forgeries","DCT-based dual-branch detection generalizes deepfake spotting","Cross-modal spatial-frequency learning outperforms on Celeb-DF","Spatial-frequency fusion generalizes deepfake detection across datasets","Frequency-spatial fusion hits 99.58% AUC and generalizes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-picked four statistics (mean, std, skewness, kurtosis) computed on block-wise DCT differences at the original resolution capture enough of a forgery's global spectral fingerprint; if those statistics stop separating real from fake under resizing, recompression, or a new generator, the cross-dataset gains claimed here lose their support.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-first fusion beats deepfake baselines on unseen forgeries","DCT-based dual-branch detection generalizes deepfake spotting","Cross-modal spatial-frequency learning outperforms on Celeb-DF","Spatial-frequency fusion generalizes deepfake detection across datasets","Frequency-spatial fusion hits 99.58% AUC and generalizes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000726,"raw_usage":{"total_tokens":3252,"prompt_tokens":940,"completion_tokens":2312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2225}},"tokens_in":556,"tokens_out":2312,"duration_ms":13810,"temperature":1.0,"reasoning_tokens":2225,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:45:29.932285+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SFCL-HCMF on FF++ (c23) and evaluate it on forgeries produced by a generator family absent from FF++, Celeb-DF, and DFDC, with compression or post-processing not used in training; if the frame-level AUC falls to the level of the spatial-only backbone, the generalization claim is falsified. A cheaper check: compute the 2304-dimensional SIDA descriptor on real and fake images after downscaling to 128x128 and see whether the real and fake distributions still separate cleanly.","supporting_citations":[{"cited_title":"In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Supplies Celeb-DF v2 as the primary cross-dataset generalization target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the YCbCr conversion and block-wise DCT preprocessing adopted for the frequency stream."},{"cited_title":"InProceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Serves as the multi-attentional spatial baseline MAD in both in-dataset and cross-dataset tables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the multi-modal transformer baseline M2TR compared in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the spatial-frequency relation baseline SFDG, the closest competitor on cross-dataset transfer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the two-branch recurrent baseline that also fuses frequency and spatial evidence."}],"review_version":1}