{"id":"44476c67-ceb6-47d3-838e-435f0af94d44","arxiv_id":"2505.23586","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A pipeline that combines geometric-mean Grad-CAM maps with pre-trained segmentation masks to locate manipulations without pixel labels, tested on a small biased subset of CASIA2.0.","lead":"This paper proposes a weakly-supervised method that fuses multi-scale activation maps from an image-level manipulation detector with segmentation masks from pre-trained segmenters to localize edited regions. The authors report gains on a 40-image manually selected subset of CASIA2.0, which makes the claims hard to generalize.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's headline numbers come from a manually selected 40-image subset chosen because activation maps already localize well; the comparison to ManTraNet and the 5-11% F1 gain are therefore not valid evidence for the central feasibility claim.","rationale":"I agree with the reader that Eq. 7 is seriously underspecified: the probability terms are undefined and the equation is a mathematical identity rather than an implementable procedure, so the claimed refinement stage cannot be reproduced or independently validated as written. However, I regard the manual selection in Section 3.5 as the more load-bearing weakness, because it invalidates the quantitative comparison that the abstract, strongest claim, and conclusion all rely on. Even if Eq. 7 were fully formalized tomorrow, the reported 0.704 AUC and 0.682 F1 would still not support the headline conclusion, since those numbers were computed only on images for which the activation maps already approximately highlighted the target region. The paper itself admits this selection criterion, so this is not an external assumption but an internal limitation. A controlled evaluation on an unbiased subset of CASIA2.0, or on all correctly classified manipulated images, is the single check that would settle whether the central feasibility claim survives. Additional weaknesses, such as the absence of comparison to the cited weakly-supervised baselines [23, 25, 26], reinforce the rejection but are secondary. Given both the sampling bias and the non-algorithmic refinement step, the current manuscript does not establish the claimed result; the reader's REJECT verdict remains appropriate.","tokens_in":7529,"tokens_out":3357,"duration_ms":33615,"concrete_test":"Ask the authors to run the full pipeline on a fixed random held-out split of CASIA2.0, or at least on every manipulated image that the image-level classifier identifies correctly, without manual filtering by activation-map quality, and to report AUC and F1 using a fixed pixel threshold. If the AUC or F1 falls below ManTraNet's values (0.653 AUC, 0.238 F1), or if the F1 improvement over the backbone disappears, the central feasibility claim fails. In parallel, require a line-by-line definition of the probability terms in Eq. 7 or runnable code; without either, the proposed refinement stage remains non-verifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the paper, that weakly-supervised manipulation localization is feasible and outperforms fully-supervised ManTraNet, rests on Table 1. Yet Section 3.5 states that the 40 test images were 'manually selected ... for which the activation maps approximately highlight the target region.' This is selection on the outcome: only images whose coarse maps already localize well are scored. The reported AUC of 0.704 and F1 of 0.682 are therefore not estimates for any well-defined population, and the comparison against ManTraNet on a disjoint, non-random sample is not meaningful. Even setting this aside, the refinement step in Eq. 7 is not an operational algorithm: P(A), P(Mi*), and P(Mi*|A) are never defined or estimated from data, so the 'Bayesian inference' is just Bayes' rule rewritten and cannot be implemented, reproduced, or validated. Consequently, the conclusion's claimed 5% to 11% F1 improvement over the backbone is unsupported both statistically and operationally.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weakly-supervised method for localizing manipulated image regions without pixel-level labels. The pipeline uses the authors' image-level detector WCBnet to compute Grad-CAM activation maps at multiple feature scales, aggregates them by geometric mean, and then combines the resulting coarse map with masks from pre-trained segmentation models (DeepLab, SAM, PSPnet) through a step described as Bayesian inference. The method is evaluated on 40 manually selected images from CASIA2.0; Table 1 reports an AUC of 0.704 and an F1-score of 0.682 for WCBnet combined with DeepLab, which the authors compare favorably against the fully-supervised ManTraNet (AUC 0.653, F1 0.238). The paper's central claim is that accurate manipulation localization is feasible without pixel-level annotations.","tokens_in":7770,"tokens_out":6681,"duration_ms":60024,"significance":"The motivation is relevant and the direction of combining class activation maps with off-the-shelf segmentation models is a sensible one for weakly-supervised localization. The paper also provides qualitative evidence that multi-scale activation maps can improve localization over a single backbone, and it honestly acknowledges the limitation of small-region manipulation. However, the quantitative evidence is not currently convincing: the 40-image test set is selected on the outcome, the Bayesian refinement is not an operational algorithm, and the conclusion's stated F1 improvement does not match the numbers in Table 1. These issues directly affect the paper's central feasibility claim. If the authors replace the biased evaluation with a random or full-dataset protocol and specify an implementable refinement step, the approach could become publishable; as it stands, the empirical contribution is unverified and the method is not reproducible.","major_comments":[{"comment":"The quantitative evaluation is invalid for the paper's central claim because Section 3.5 states that the 40 test images were \"manually selected ... for which the activation maps approximately highlight the target region.\" This is selection on the outcome: only images whose coarse activation maps already localize the manipulation enter the test set, so the reported AUC of 0.704 and F1 of 0.682 are not estimates for any definable population. The comparison to ManTraNet on the same cherry-picked subset is therefore not a meaningful benchmark; the authors should report results on the full CASIA2.0 test split or a random subset, and if the selection is retained, it should be presented as an illustrative case study rather than as evidence for the feasibility claim.","section":"Section 3.5, Table 1"},{"comment":"Equation (7) is Bayes' rule, not an algorithm. The quantities P(A), P(Mi*) and P(Mi*|A) are never defined in terms of the available data, never estimated, and no procedure is given for computing the \"enhanced activation heatmap\" A* from Eq. (7). Without an operational definition of these probabilities, the refinement step cannot be implemented, reproduced, or validated, and the claimed Bayesian inference is a mathematical identity that does no computational work. The authors need to specify how these probabilities are computed from A and Mi*, or replace Eq. (7) with a concrete algorithmic step.","section":"Section 2.4, Eq. (7)"},{"comment":"Because the selected mask is the one that maximizes the similarity to A (Eq. (6)) and P(Mi*|A) is never estimated, the final output A* is essentially a reweighting or smoothing of A by the most similar segmentation mask. The paper does not demonstrate that the segmentation maps contribute information beyond a monotone transform of the activation map; a simple test would be to compare A* against the original A on a random set using the same metrics. As written, the combination step risks being circular and does not substantiate the claim of adding independent region information.","section":"Section 2.4, Eqs. (5)-(6)"},{"comment":"The conclusion's claim of a \"5% to 11%\" F1 improvement over the backbone model is not supported by Table 1. The table shows WCBnet+DeepLab at 0.682 F1 versus ResNet+DeepLab at 0.296 (a much larger relative improvement), WCBnet+SAM at 0.310 versus ResNet+SAM at 0.257, and WCBnet+PSPnet at 0.365 versus ResNet+PSPnet at 0.259; no column in the table reports the backbone model alone without segmentation, and no value in the table falls in the claimed 5-11% range. The authors should specify the baseline and the computation used for this claim and report the corresponding numbers.","section":"Conclusion and Table 1"}],"minor_comments":[{"comment":"The sentence \"To claim, we only use the segmentation masks from DeepLab although it could be a semantic segmentation model\" is incomplete and inconsistent with Table 1, which lists SAM and PSPnet; please clarify which segmentation models are used for quantitative evaluation and which are used only for visualization.","section":"Section 3.2"},{"comment":"The paper reports training on approximately 1,800 images per class but does not state how the 40 test images are split from or related to the training set; please describe the validation/test protocol and the manipulation types (splicing versus copy-move) in the 40-image subset.","section":"Section 3.1"},{"comment":"The notation D(Mi) is used without a precise definition of the distance transform (for example, Euclidean distance to the mask boundary) and its normalization; please define it explicitly so that Eq. (5) is reproducible.","section":"Equation (5)"},{"comment":"The caption uses \"The features maps\" (typo) and the symbols WCBnet2, WCBnet3, WCBnet4, and WCBnetm are not defined in the caption; please add a clear legend or explanation.","section":"Figure 2"},{"comment":"The paper reports AUC and F1 with a \"fixed threshold\" but never states the threshold value or how it was chosen; this should be specified for reproducibility.","section":"Section 3.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's heavy reliance on the authors' own WCBnet as the sole image-level backbone, together with the selection-biased evaluation, makes the claimed advantage over fully-supervised methods unpersuasive. I would advise the editor that the paper needs a substantially revised experimental design and an actual algorithmic specification of the refinement step before it can be considered further. The comparison with ManTraNet should also clarify the training data and protocol for the fully-supervised baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the idea is reasonable, but the evidence is not. The authors combine geometric-mean multi-view Grad-CAM from their WCBnet with a simple mask-selection heuristic: pick the pre-trained segmentation mask most similar to the coarse activation map and call that the refined output. The qualitative figures suggest that this can clean up edges, and the multi-view averaging over CBAM block sizes is a cheap, plausible way to stabilize Grad-CAM. Comparing DeepLab, SAM, and PSPNet is also a reasonable sanity check. That part is worth a look.\n\nThe problem is the evaluation and the fusion step. Section 3.5 says explicitly that the 40 test images were manually selected because their activation maps already highlight the target region. So the reported AUC of 0.704 and F1 of 0.682 are conditional on the outcome, not estimates for any defined population. The comparison to ManTraNet on a different, non-random sample is meaningless, and the conclusion's 5-11% F1 gain over the backbone inherits the same bias. The authors are transparent about the manual selection, which is honest, but it does not make the numbers usable.\n\nThe deeper structural gap is Eq. 7. Writing Bayes' rule with P(A), P(Mi*), and P(A|Mi*) does not become an inference algorithm until you say what these quantities are and how they are estimated. As written, A* is not computable. The core refinement step is therefore underspecified and unreproducible. This is not a minor omission; it is the entire method after the mask selection. If the authors meant a pixelwise weighted average, they should just say that.\n\nThere is also no comparison to the weakly-supervised baselines they themselves cite, which makes it hard to judge whether the combination adds anything beyond known ingredients.\n\nWho this is for: someone working on weakly-supervised forensics might borrow the multi-view averaging idea, but the paper as a whole is not a reliable source for the feasibility claim. I would not cite it for the numbers. If the authors redo the evaluation on a random or full test split and specify the fusion operation, it might be worth another look. As submitted, I would desk-reject it.","headline":"A sensible segmentation-refinement idea is sunk by evaluation on a manually selected 40-image subset and a Bayesian fusion step that is never actually defined.","tokens_in":8239,"tokens_out":3324,"would_cite":false,"duration_ms":32614,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Weakly-supervised localization of manipulated image regions is feasible without pixel-level labels, using fused activation and segmentation maps.","keywords":["weakly-supervised learning","image manipulation detection","tampering localization","Grad-CAM","segmentation masks","Bayesian fusion","CASIA2.0"],"falsifier":"Evaluate the described pipeline (with any fixed instantiation of the undefined probabilities) on all tampered CASIA2.0 images rather than the manually selected 40, and check whether the AUC/F1 advantage over ManTraNet survives; if it vanishes on the unselected set, the reported superiority is an artifact of hand-picking images where the activation maps already point at the tampered region.","tokens_in":7356,"feed_emoji":"🔍","tokens_out":5807,"duration_ms":48933,"temperature":0.7,"pith_summary":"This paper tries to establish that image manipulation can be localized without any pixel-level ground truth, using only image-level manipulation labels plus off-the-shelf segmentation. It combines multi-scale Grad-CAM activation maps from an image-wise detector (WCBnet) with region masks from pre-trained segmentation models, then refines the result with a Bayesian fusion step. On a 40-image manually selected subset of CASIA2.0, the method reports an AUC of 0.704 and F1 of 0.682, exceeding the fully-supervised ManTraNet baseline (AUC 0.653, F1 0.238). The authors conclude that weak supervision is sufficient for accurate manipulation localization.","feed_headline":"Weak supervision localizes tampered regions with no pixel labels","feed_subtitle":"Fusing multi-scale activation maps with segmentation masks beats a fully-supervised baseline.","key_machinery":"The method's engine is a three-part fusion. First, the image-wise detector WCBnet, built on ResNet-50 with Cross-block Attention Modules, weights and fuses hierarchical features; Grad-CAM over three CBAM configurations yields activation maps $A_i$ that are aggregated by geometric mean $A = (\\prod_{i\\in\\{2,3,4\\}} A_i)^{1/3}$. Second, pre-trained segmentation models (DeepLab, SAM, PSPnet) partition the image into masks $M_i$. Third, a distance-transform similarity $S(M_i,A)$ selects the best mask $M_{i^*}$, and a Bayesian expression $P(A|M_{i^*}) = P(M_{i^*}|A)P(A) / P(M_{i^*})$ is used to refine the heatmap. The distance transform ties the coarse activation to object boundaries; the Bayesian step is the formal mechanism by which the selected mask sharpens the map, though the paper does not specify how the probability terms are estimated.","core_discovery":"The central claim is that fusing activation maps from multiple receptive fields of an image-level manipulation classifier with a segmentation mask selected by boundary proximity, then applying Bayes' rule, produces accurate pixel-level localization despite having no pixel labels. The paper argues that the geometric mean of Grad-CAM maps from CBAM variants at scales 2, 3, and 4 captures both fine detail and global context, and that the pre-trained segmentation mask (especially DeepLab) supplies the precise object boundaries the activation maps lack. The reported improvement over both the ResNet backbone and the fully-supervised ManTraNet is presented as evidence that this weak-supervision route is viable.","pith_inferences":["One extension the authors do not pursue is a full-dataset evaluation: if the AUC/F1 advantage over ManTraNet persists when the 40-image hand-picked subset is replaced by all tampered CASIA2.0 images, the method's practical reach is much wider than the paper demonstrates.","The Bayesian fusion in equation (7) is formally an identity until the probabilities are given operational definitions; specifying them empirically (e.g., activation value histograms over training images) would turn the refinement step into a concrete, reproducible algorithm, and its benefit could be measured by ablation.","The distance-transform similarity could serve as a general selection rule for choosing which segmentation mask to trust for a given image, a component that other weakly-supervised localization pipelines might reuse.","If this recipe transfers to video or deepfake localization, it would inherit the same advantage of needing only clip-level labels."],"forward_implications":["If the reported results hold, forensic tools can locate spliced or copy-moved regions in images that carry only an image-level 'manipulated' flag, removing the need for expensive pixel-wise annotation.","The geometric-mean fusion of activation maps across receptive fields is a generally applicable ingredient: any image-level classifier with multi-branch feature weighting can be turned into a coarse localizer this way.","Combining a coarse activation map with a pre-trained segmentation mask yields better localization than either source alone, with DeepLab's larger regions giving the best trade-off in this paper.","A fully-supervised model such as ManTraNet is not an insurmountable upper bound in localization quality when the evidence is weak supervision plus generic segmentation."],"supporting_citations":[{"why":"Supplies the WCBnet image-wise manipulation detector whose cross-block attention features produce the activation maps used for localization.","marker":"[18]"},{"why":"Grad-CAM is the method used to convert the classifier's weighted features into the multi-view activation maps.","marker":"[14]"},{"why":"DeepLab2 is one of the pre-trained segmentation models used to generate region masks; its masks give the best reported results.","marker":"[19]"},{"why":"SAM is a pre-trained segmentation model used to generate fine-grained region masks.","marker":"[12]"},{"why":"PSPnet is a pre-trained segmentation model used to generate region masks and compared against DeepLab and SAM.","marker":"[8]"},{"why":"CASIA2.0 is the dataset used for training the image-wise model and for the 40-image localization evaluation.","marker":"[5]"},{"why":"ManTraNet is the fully-supervised baseline that the proposed weakly-supervised method is compared against.","marker":"[20]"},{"why":"Signed-value error levels provide the input preprocessing that WCBnet uses to expose JPEG compression artifacts.","marker":"[6]"}],"fun_headline_variants":["No pixel labels: multi-res maps pinpoint tampering","Bayesian fusion of activation and segmentation maps localizes edits","Fusing activation maps with segmentation finds edits without labels","Weakly-supervised tampering localization via map fusion and Bayes","Multi-scale activation plus segmentation detects edited pixels weakly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that applying Bayes' rule to activation values and a segmentation mask genuinely improves localization, even though the probability terms in the fusion equation are never defined or estimated from data.","fun_headline_variants_meta":{"raw":{"variants":["No pixel labels: multi-res maps pinpoint tampering","Bayesian fusion of activation and segmentation maps localizes edits","Fusing activation maps with segmentation finds edits without labels","Weakly-supervised tampering localization via map fusion and Bayes","Multi-scale activation plus segmentation detects edited pixels weakly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00103,"raw_usage":{"total_tokens":4286,"prompt_tokens":840,"completion_tokens":3446,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":3368}},"tokens_in":456,"tokens_out":3446,"duration_ms":24114,"temperature":1.0,"reasoning_tokens":3368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:42:17.975270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the described pipeline (with any fixed instantiation of the undefined probabilities) on all tampered CASIA2.0 images rather than the manually selected 40, and check whether the AUC/F1 advantage over ManTraNet survives; if it vanishes on the unselected set, the reported superiority is an artifact of hand-picking images where the activation maps already point at the tampered region.","supporting_citations":[{"cited_title":"WCBnet: Weighted convolutional block mod- elling of signed-value error levels for image-wise copy-move and splicing detection","cited_arxiv_id":null,"evidence_quote":"Supplies the WCBnet image-wise manipulation detector whose cross-block attention features produce the activation maps used for localization."},{"cited_title":"Grad-CAM: visual explanations from deep networks via gradient-based localization","cited_arxiv_id":null,"evidence_quote":"Grad-CAM is the method used to convert the classifier's weighted features into the multi-view activation maps."},{"cited_title":"Image Segmentation Keras : Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras","cited_arxiv_id":"2307.13215","evidence_quote":"PSPnet is a pre-trained segmentation model used to generate region masks and compared against DeepLab and SAM."},{"cited_title":"ManTra-Net: Manipulation tracing network for detection and localization of image forgeries with anomalous fea- tures","cited_arxiv_id":null,"evidence_quote":"ManTraNet is the fully-supervised baseline that the proposed weakly-supervised method is compared against."},{"cited_title":"Development of photo forensics algorithm by detecting photoshop manipulation using error level analysis","cited_arxiv_id":null,"evidence_quote":"Signed-value error levels provide the input preprocessing that WCBnet uses to expose JPEG compression artifacts."}],"review_version":1}