{"id":"8334b99b-d467-4303-bc34-2f4628c1a3f7","arxiv_id":"2507.16397","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ADCD-Net adaptively weights JPEG-DCT features by a predicted alignment score and disentangles document content from editing traces to localize forgeries robustly under resizing, cropping, compression, blur, and noise.","lead":"ADCD-Net is a new computer vision model that pinpoints tampered regions in document images, such as edited contracts or receipts, and stays accurate when images are resized, cropped, compressed, blurred, or noisy. A learned switch decides how much to trust JPEG frequency traces in each image, which is the key to surviving everyday image distortions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 20.79% robustness gain is the central claim, but the paper provides no numeric table, error bars, or averaging protocol behind Fig. 4, leaving the headline result unverifiable.","rationale":"I identify as more load-bearing than the PPE/OCR assumption the absence of a numeric table and error bars for the central 20.79% robustness claim. The PPE module contributes only a small increment in the ablation (Table 4, row #8 vs. row #6: 0.750 vs. 0.741), so even a failure of the pristine-background assumption would not overturn the headline robustness result. By contrast, the 20.79% figure is the paper's primary claimed contribution and is presented only as curves in Fig. 4. Without per-condition F1 values, standard deviations, and the exact averaging protocol, the claim cannot be checked from the manuscript; it could be driven by a single favorable condition or by the choice of a weak baseline. The reader's CONDITIONAL verdict already flags the missing table and error bars, so I do not change the verdict, but I shift the main concern from the PPE assumption to the verifiability of the central quantitative claim.","tokens_in":13786,"tokens_out":6869,"duration_ms":73004,"concrete_test":"Run the released checkpoint on Test/FCD/SCD under each of the five distortions at the severity levels used for Fig. 4; produce per-condition F1 and standard deviations; recompute the 20.79% average gain over TruFor according to the paper's averaging rule and report gains per subset and excluding the most extreme severity level. If the gain is not reproduced or the advantage is not consistent across subsets and severities, the headline robustness claim is overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is the 20.79% average improvement in robustness over state-of-the-art methods (Abstract, Section 1, Fig. 4). This number is load-bearing because the entire stated contribution is the robustness gain. However, the paper does not provide a numeric table of F1 scores behind Fig. 4, no error bars or confidence intervals, and no specification of how the average over '5 types of distortions' is computed (number of severity levels, inclusion of the clean condition, relative vs. absolute percentage, and which method is used as the baseline at each condition). The only tabulated results (Table 1) are for the standard DocTamper benchmark, where ADCD-Net's average F1 (0.787) is essentially tied with PS-Net (0.785) and not best on every subset (e.g., on FCD, ADCD-Net 0.815 vs. PS-Net 0.827 and DTD 0.816). Thus the 'consistently outperforming' language is not supported by the data shown in the paper. Without the underlying numeric results and variance estimates, a reader cannot determine whether the 20.79% gain is consistent across subsets and severity levels or dominated by a few conditions, nor whether it is statistically meaningful. This is a reportability and reproducibility gap in the central evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADCD-Net, a document image forgery localization network with three main components: an adaptive RGB-DCT encoder that modulates DCT features with a learned alignment score, a Hierarchical Content Decoupling (HCD) module that separates content from forgery features via reconstruction with shuffled forgery features, and a Pristine Prototype Estimation (PPE) module that constructs a pristine prototype from OCR-background pixels. The model is trained on DocTamper and evaluated on cross-domain test sets under five distortion types, with the central claim of a 20.79% average F1 improvement over TruFor, as well as a competitive result on the standard DocTamper benchmark (average F1 0.787). The paper also presents ablations in Table 4 and analyses of the alignment score, feature gradients, and text-background bias.","tokens_in":14103,"tokens_out":7773,"duration_ms":81653,"significance":"If the robustness claim is substantiated, the work is significant: existing DCT-based document forgery localizers degrade sharply under resizing, cropping, and block misalignment, and ADCD-Net's adaptive gating of DCT features offers a plausible and practical remedy. The architecture is described in enough detail to be re-implemented, the code is released, public benchmarks are used, and the ablation in Table 4 broadly supports the contribution of each module. The analysis in Table 3 and Fig. 6 showing that the predicted score tracks distortion types is a useful diagnostic. However, the headline robustness number is not currently verifiable from the reported data, and the PPE module's core assumption about pristine backgrounds is not stress-tested. These gaps prevent the paper from being accepted in its present form.","major_comments":[{"comment":"The headline robustness claim that ADCD-Net outperforms TruFor by 20.79% on average over five distortions is not supported by a numerical table, per-distortion F1 values, error bars or confidence intervals, or a definition of the averaging protocol (how severity levels are aggregated, whether the clean condition is included, and whether the percentage is absolute or relative). Please add a table with per-distortion, per-severity F1 values in the format of Table 1, state the exact aggregation formula, and provide variance estimates; without these, the central claim cannot be independently checked.","section":"§4.3, Fig. 4"},{"comment":"The PPE prototype is computed by averaging forgery features over OCR-background pixels, which assumes that background regions are pristine and that the CRAFT mask is correct. Neither assumption is tested. If a tampered region lies in the background or if the OCR mask mislabels text, contaminated prototypes directly enter Eq. (4) and the modulation in Eq. (5). Please add a sensitivity analysis with synthetic OCR-mask corruptions and with a set of forged-background test cases, and report the impact on the PPE rows of Table 4.","section":"§3.3, Eq. (3)"},{"comment":"The statement that ADCD-Net \"consistently outperforms\" state-of-the-art methods is not supported by the standard benchmark in Table 1: the average F1 margin over PS-Net is 0.002, and on the FCD subset ADCD-Net (0.815) is below both PS-Net (0.827) and DTD (0.816). Please either qualify the wording to \"competitive on DocTamper\" or provide a statistical test of the differences; also specify clearly whether the 20.79% figure refers only to the robustness protocol of Fig. 4.","section":"§4.3, Table 1"},{"comment":"The ablation attributes performance gains to individual modules, but every row reports a single F1 value on a 3000-sample subset with no standard deviation or repeated-run information. Several adjacent rows differ by less than 0.02 (e.g., rows #5 and #6 differ by 0.026, rows #7 and #8 by 0.041), so without variance estimates the module-attribution claims are not statistically supported. Please report means and standard deviations over at least three training runs or over bootstrap resamples of the test subset.","section":"§4.5, Table 4"},{"comment":"The paper repeatedly refers to \"the supplementary\" for implementation details, more severe degradations, and additional PPE and HCD analysis, but the arXiv submission contains no supplementary material. Since part of the evidence for robustness and module design may reside there, please include the supplementary in this submission or state explicitly which details are omitted.","section":"Supplementary material"}],"minor_comments":[{"comment":"The acronym is inconsistent: \"ADCD-Net\" appears in the title and abstract, while \"ACDC-Net\" appears in Fig. 3, Tables 1-2, and the GitHub URL in the abstract; please unify the naming.","section":"Throughout"},{"comment":"Equation (1) uses F^{i+1}_{fuse} on the left-hand side but defines the feature set as {F^i_fuse}_{i=1}^L; please clarify the indexing so that the multi-scale fusion is unambiguous.","section":"§3.1, Eq. (1)"},{"comment":"The reported gains of \"8.37%\" and \"7.06%\" do not state whether they are absolute F1 differences or relative percentages; please specify the calculation.","section":"§4.4, Table 3"},{"comment":"\"TurFor\" in the table header should be \"TruFor\".","section":"Table 2"},{"comment":"The sentence \"This phenomenon can be easily verified in many datasets, e.g. [27, 29]\" cites FaceForensics++ [29], which contains face images rather than documents; please replace with a document-specific dataset or remove the citation.","section":"§3.3, first paragraph"},{"comment":"The caption describes \"Clean\" as no distortion, but it is unclear whether the clean condition is included in the \"5 types of distortions\" average; please clarify in the caption or text.","section":"Fig. 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The most important issue is the missing numeric support for the 20.79% robustness claim; if the authors provide the underlying table, aggregation details, and variance estimates, plus a sensitivity study for the PPE prototype, the paper could become acceptable. I do not see evidence of a deeper methodological flaw, and the architecture description is sufficiently detailed to re-implement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honestly, this is a credible piece of document forensics work. The genuinely new piece is the learned alignment score that gates DCT feature contribution, which is a simple answer to a real problem: DTD-style DCT fusion collapses under resizing, cropping, and pixel shifts. The hierarchical content decoupling with shuffled forgery features is also sensible, and the ablation in Table 4 supports the contribution of each module to the average F1. The writing is clear and the related work is properly covered.\n\nThe soft spot is the headline claim. The abstract and introduction promise a 20.79% average gain over state of the art under five distortions, but the only evidence is Fig. 4, a line chart with no numeric table, no error bars, and no statement of how the average is computed or which conditions are included. That is a load-bearing number and it is unverifiable from the paper. The stress-test note is right about this. Also, on the standard DocTamper benchmark (Table 1), ADCD-Net is essentially tied with PS-Net (0.787 vs 0.785) and is not best on the FCD or SCD subsets, so the phrase 'consistently outperforming' overstates the tabulated results.\n\nThe PPE module assumes background pixels are mostly pristine and that the CRAFT OCR mask is accurate. The paper states the assumption but does not test sensitivity to OCR errors or forgeries placed in the background. That is a real limitation, since the prototype is aggregated from OCR-background pixels. The hyperparameters (loss weights, contrastive sample size, augmentation magnitudes) are deferred to the supplementary, which is not accessible. Minor point: the paper alternates between ADCD-Net and ACDC-Net, and the code link uses the latter.\n\nNone of these concerns are fatal. The architecture is coherent, the ablation supports the internal logic, and the adaptive alignment score is a genuinely citable idea. The missing numeric table is a reportability gap that a referee can reasonably request.\n\nMy take: send it to peer review, but the authors should be asked to provide the exact F1 values and variance estimates behind Fig. 4, and to state the averaging protocol. If those numbers hold, this is a useful contribution.","headline":"Solid architecture with a genuinely new adaptive DCT-gating idea, but the headline 20.79% robustness gain rests on a figure with no supporting numbers, so the paper needs revision before its central claim can be trusted.","tokens_in":14596,"tokens_out":4904,"would_cite":true,"duration_ms":46723,"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":"ADCD-Net claims to make document forgery localization robust to cropping, resizing, and recompression by adaptively weighting DCT forensic features.","keywords":["document image forgery","forgery localization","DCT forensics","adaptive feature fusion","content disentanglement","pristine prototype estimation","JPEG artifacts","robustness to distortions"],"falsifier":"Take a set of forged documents where tampering is inserted into uniform background areas (or where the OCR mask is systematically wrong), run ADCD-Net with and without the Pristine Prototype Estimation module, and compare F1; if the assumption holds, disabling that module should barely matter on foreground-only forgeries, while background-embedded forgeries should show a measurable drop in the full model relative to the ablated one.","tokens_in":13584,"feed_emoji":"📄","tokens_out":6573,"duration_ms":59925,"temperature":0.7,"pith_summary":"The paper argues that the main barrier to deploying document forgery detectors in practice is not the absence of forensic traces but their fragility: DCT-domain compression artifacts are informative when the image is block-aligned but misleading after cropping, resizing, or pixel shifts. To fix this, ADCD-Net learns a per-image alignment score that decides how much weight the DCT features should receive, fusing them with RGB features only when they are trustworthy. On top of this, the method disentangles content from forgery features at multiple scales and uses the mostly pristine background of documents to estimate a 'pristine prototype' that sharpens the final tampering map. The paper reports that the full model outperforms state-of-the-art methods by an average of 20.79% across five distortion types, while remaining competitive under standard JPEG recompression with a modest 23M parameters. If true, this would make document forgery localization robust enough for real-world images that have been resized, cropped, or recompressed.","feed_headline":"Adaptive DCT weighting boosts document forgery localization by 20.79%","feed_subtitle":"It keeps finding tampered regions even after cropping, resizing, or recompression.","key_machinery":"The load-bearing mechanism is the adaptive DCT alignment score: a small classification head on the deepest DCT feature predicts $\\hat{s}_{aln}$, which multiplies the DCT feature at every encoder scale before fusion with RGB features (Eq. 1), trained with a cross-entropy loss against labels derived from whether the input was block-aligned. Two document-specific modules carry the rest of the argument: the Hierarchical Content Decoupling module splits each fused feature into content and forgery channels and enforces the split by reconstructing the image and DCT coefficients from content features plus a spatially shuffled forgery stream, and the Pristine Prototype Estimation module computes $\\mathbf{p}^i_{prs}$ by averaging forgery features over background pixels identified by an OCR model (Eq. 3), then uses per-pixel cosine similarity to that prototype (Eq. 4) to scale and bias the penultimate feature (Eq. 5).","core_discovery":"The central claim is that DCT-based forensic features can be made robust rather than discarded. The authors propose to modulate the DCT branch by a predicted alignment score $\\hat{s}_{aln}=f_{asp}(F^L_{dct})\\in(0,1)$, so that the fused features are $F^i_{fuse}=f^i_{fuse}(F^i_{rgb}+\\hat{s}_{aln}\\,F^i_{dct})$ at each scale. A Hierarchical Content Decoupling module separates content from forgery features across scales, and a Pristine Prototype Estimation module averages forgery features over OCR-marked background pixels (Eq. 3) to build a pristine reference that re-weights the penultimate features. With these components, the network is claimed to consistently outperform state-of-the-art methods by 20.79% averaged over 5 types of distortions while attaining an average F1 of 0.787 on the standard recompression benchmark and the lowest false alarm rate (1.2%) among compared methods on pristine documents.","pith_inferences":["Inference: the alignment-score idea is a general recipe: any forensic trace whose reliability depends on a measurable precondition (block alignment, JPEG grid, camera noise consistency) could be gated by a learned scalar, so the same architecture could be adapted to other fragile features.","Inference: because the pristine prototype is built from OCR-marked background, forgeries deliberately placed in uniform background regions—or OCR errors that label text as background—should degrade the Pristine Prototype Estimation module; a natural test is to compare ADCD-Net with and without that module on background-embedded tampering.","Inference: the multi-scale content/forgery disentanglement might transfer to other document forensics tasks such as detecting synthetic text or AI-generated documents, where the separation of what the content is from how it was produced is the same underlying distinction."],"forward_implications":["DCT-based detectors no longer need to choose between accuracy on recompressed images and robustness to cropping, resizing, and pixel shifts.","A single model with 23M parameters can hold its own against much larger general-purpose detectors on standard recompression while far exceeding them under five distortion types.","The OCR-driven pristine prototype gives the model a very low false alarm rate (1.2%) on pristine documents, which is important for practical screening.","Content disentanglement with reconstruction from shuffled forgery features provides a mechanism to suppress text-background bias, improving localization of small tampered text."],"supporting_citations":[{"why":"Supplies the DocTamper dataset, the FPH DCT feature extractor reused as $E_{dct}$, and the principal DCT-fusion baseline (DTD) that ADCD-Net must beat.","marker":"[27]"},{"why":"Establishes JPEG double-quantization artifacts from DCT coefficients for manipulation localization, motivating fusion of RGB and DCT features.","marker":"[16]"},{"why":"Provides the strongest robustness competitor (TruFor) and the self-supervised noise-fingerprint approach that ADCD-Net is compared against across distortions.","marker":"[12]"},{"why":"Supplies the within-image contrastive loss (FOCAL) that the paper adopts as $L_{con}$ to separate pristine and forged pixels.","marker":"[40]"},{"why":"Provides the CRAFT OCR model that produces the background mask $X_{bg}$ used by Pristine Prototype Estimation.","marker":"[1]"},{"why":"Documents the fragility of DCT compression traces to block-misalignment operations, which motivates the adaptive alignment score.","marker":"[6]"},{"why":"Provides the Restormer encoder/decoder backbone that ADCD-Net builds on and initializes with DocRes.","marker":"[45]"}],"fun_headline_variants":["Adaptive DCT beats crop, resize, recompress forgeries","20.79% better doc forgery localization via adaptive DCT","Forgery locator stays sharp after crop, resize, recompress"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that background regions in a document are predominantly pristine, so averaging forgery features over OCR-marked background yields a clean pristine prototype; if a forgery sits in the background or the OCR mask mislabels text, that prototype is contaminated and the contrast it creates is distorted.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive DCT beats crop, resize, recompress forgeries","20.79% better doc forgery localization via adaptive DCT","Forgery locator stays sharp after crop, resize, recompress"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2690,"prompt_tokens":1017,"completion_tokens":1673,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":1614}},"tokens_in":633,"tokens_out":1673,"duration_ms":12642,"temperature":1.0,"reasoning_tokens":1614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:11:00.578446+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of forged documents where tampering is inserted into uniform background areas (or where the OCR mask is systematically wrong), run ADCD-Net with and without the Pristine Prototype Estimation module, and compare F1; if the assumption holds, disabling that module should barely matter on foreground-only forgeries, while background-embedded forgeries should show a measurable drop in the full model relative to the ablated one.","supporting_citations":[{"cited_title":"Towards robust tampered text detection in document image: New dataset and new solution","cited_arxiv_id":null,"evidence_quote":"Supplies the DocTamper dataset, the FPH DCT feature extractor reused as $E_{dct}$, and the principal DCT-fusion baseline (DTD) that ADCD-Net must beat."},{"cited_title":"Learning jpeg compression artifacts for image manipulation detection and localization","cited_arxiv_id":null,"evidence_quote":"Establishes JPEG double-quantization artifacts from DCT coefficients for manipulation localization, motivating fusion of RGB and DCT features."},{"cited_title":"Trufor: Leveraging all-round clues for trustworthy image forgery detection and localiza- tion","cited_arxiv_id":null,"evidence_quote":"Provides the strongest robustness competitor (TruFor) and the self-supervised noise-fingerprint approach that ADCD-Net is compared against across distortions."},{"cited_title":"Rethinking im- age forgery detection via contrastive learning and unsuper- vised clustering","cited_arxiv_id":null,"evidence_quote":"Supplies the within-image contrastive loss (FOCAL) that the paper adopts as $L_{con}$ to separate pristine and forged pixels."},{"cited_title":"Character region awareness for text de- tection","cited_arxiv_id":null,"evidence_quote":"Provides the CRAFT OCR model that produces the background mask $X_{bg}$ used by Pristine Prototype Estimation."},{"cited_title":"Self-supervised image local forgery detection by jpeg compression trace","cited_arxiv_id":null,"evidence_quote":"Documents the fragility of DCT compression traces to block-misalignment operations, which motivates the adaptive alignment score."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Provides the Restormer encoder/decoder backbone that ADCD-Net builds on and initializes with DocRes."}],"review_version":1}