{"id":"f2dc0c89-a3f3-4ec4-a6dd-e723a33f4911","arxiv_id":"2412.09981","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"SUMI-IFL applies sufficiency and minimality information-theoretic constraints to forgery feature learning and reports state-of-the-art image forgery localization on multiple benchmarks.","lead":"This paper proposes SUMI-IFL, a neural network for locating tampered regions in forged images, adding two information-theoretic loss terms intended to keep the learned features comprehensive and compact. Reported results beat seven prior methods on several forgery benchmarks, but the theoretical derivations behind the losses contain serious mathematical gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sufficiency-view derivation is not mathematically sound: Eq. (3)'s monotonicity premise is false, and the appendix's joint-KL bound does not match the conditional-KL loss in Eq. (5), so the claimed theoretical grounding for SUMI-IFL is unsupported.","rationale":"I agree with the reader's conditional assessment. The concern is load-bearing because the paper's contribution bullets explicitly claim the constraints are 'supported by rigorous theoretical analysis,' and the sufficiency loss is the first of the two new constraints. The reported F1 gains and the ablation study are suggestive, but without a correct derivation the performance cannot be attributed to the information-theoretic objective. I would not move the verdict to reject, because a corrected proof and a released implementation could validate the empirical direction; the experiments are extensive and the ablation shows both losses matter. The conditional verdict should require fixing or removing the invalid derivations, stating the exact form of the implemented sufficiency loss, and releasing code or error bars. The reader's weakest assumption points to the same equations; my only nuance is that the decisive flaw is the unproved bridge and the joint-versus-conditional KL mismatch, not necessarily the scalar inequality itself, which may be recoverable under an expectation interpretation. Hence the verdict remains conditional.","tokens_in":14479,"tokens_out":18963,"duration_ms":215808,"concrete_test":"Run a direct numerical check on the paper's own definitions with binary M, f1, f2 and F = (f1, f2). Enumerate all joint distributions (or at least test the identical-copy distribution M=f1=f2) and compute: (a) Eq. (3)'s LHS I(M;F) and RHS sum_i I(f_i;M|F\\f_i); the identical-copy case gives 1 bit vs 0 bits, falsifying the asserted monotonicity. (b) The appendix's Eq. (22) lower bound sum_i D_KL[p(M,F_i)||p(M,F\\f_i)] versus the implemented loss sum_i E[D_KL[p(M|F)||p(M|F\\f_i)]]; show whether these quantities are equal. If they are unequal, the appendix proof does not cover the loss used in Eq. (5), regardless of whether the scalar inequality in Eq. (4) can be salvaged.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the derivation of the sufficiency constraint, Eqs. (3)-(5). Equation (3) replaces the chain-rule conditioning sets f1,...,f_{i-1} by the leave-one-out sets F\\f_i and asserts that expanding the known conditions can only increase the conditional mutual information. This monotonicity claim is false: if M, f1, and f2 are identical fair bits, I(M; f1, f2) = 1 bit, while I(f1; M | f2) + I(f2; M | f1) = 0. The paper states no independence or diversity assumption that would rule out such redundancy among the multi-view features, and the three attention backbones process the same input image. The appendix does not repair the gap: Eq. (22) expresses the conditional MI as a conditional entropy term plus D_KL[p(M,F_i) || p(M,F\\f_i)], and Eq. (23) keeps only the joint KL. That joint KL is not the conditional KL D_KL[p(M|F) || p(M|F\\f_i)] used in Eq. (5)/(24); the two KLs differ by marginal terms, and Eq. (4), the bridge from the derived bound to the implemented loss, is asserted without a valid derivation. Consequently, the paper does not show that minimizing exp(-D_KL[P_F || P_F\\f_i]) maximizes I(M;F), so the advertised 'rigorous theoretical analysis' behind the sufficiency constraint is unsupported. This leaves the empirical results as the main evidence, but the central contribution as stated is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SUMI-IFL, an image forgery localization framework with three attention U-Nets that extract channel, spatial, and pixel features, followed by a learned fusion layer and a feature reasoning network. Two information-theoretic losses are introduced: a sufficiency-view constraint, derived from maximizing the mutual information between the latent forgery feature and the ground-truth mask, and a minimality-view constraint, derived from the conditional entropy bottleneck. The model is trained on a compound dataset and evaluated with in-dataset, cross-dataset, and robustness experiments, reporting substantial gains over several state-of-the-art baselines.","tokens_in":14828,"tokens_out":7045,"duration_ms":76342,"significance":"If the theoretical derivation were valid, the paper would offer a principled way to constrain multi-view forgery features for localization, and the empirical results are indeed strong: Table 2 shows F1 improvements of roughly 8-15 percentage points over the best baselines on DEFACTO-12 and SSRGFD, Table 3 shows a 9.7-point F1 gain on NIST16, and the ablation study in Table 4 indicates that both proposed losses contribute to the final performance. The paper also evaluates generalization and robustness, which is a useful strength. However, the central theoretical claim is not supported as written: the derivation of the sufficiency loss rests on a false inequality and on an appendix proof that proves a different statement. The significance of the contribution is therefore conditional on either correcting the derivation or repositioning the losses as heuristic components within an empirically strong system.","major_comments":[{"comment":"The inequality in Eq. (3), I(M; f1,...,fn) = sum_i I(fi; M | f1,...,f_{i-1}) <= sum_i I(fi; M | F\\fi), is not a consequence of mutual information monotonicity. Conditional mutual information is not monotone under arbitrary enlargement of the conditioning set; for example, if M, f1, and f2 are identical fair bits, the left-hand side is 1 bit while the right-hand side is 0. Since the three backbones process the same input image and no independence or diversity assumption is stated, this step is unsubstantiated and it is the load-bearing bridge from the mutual information objective to the sufficiency loss.","section":"Method, 'Sufficiency-view constraint', Eq. (3)"},{"comment":"Eq. (4) is asserted with the proof deferred to the appendix, but the appendix does not prove it. The appendix's Eq. (23) bounds the conditional mutual information by the joint KL divergence DKL[p(M, Fi) || p(M, F\\fi)] plus a nonnegative conditional-entropy term, whereas Eq. (5) and Eq. (24) implement exp(-DKL[PF || PF\\fi]) with PF = p(y|F) and PF\\fi = p(y|F\\fi). The joint KL and the conditional KL are different quantities, and no argument shows that the joint-KL bound implies the conditional-KL inequality in Eq. (4). Thus the main text does not establish that minimizing exp(-DKL[PF || PF\\fi]) maximizes I(M; F).","section":"Method, Eq. (4); Appendix, Eqs. (21)-(24)"},{"comment":"The appendix derivation contains algebra and notation errors that prevent it from repairing the main text. Eq. (18) writes the conditional mutual information as a sum over p(M, Fi) while retaining p(F) and p(M, F\\fi), conflating distinct marginals; the derivation of Q1 in Eqs. (19)-(20) attaches H(fi | F\\fi) to a sum over M without the required joint distribution; and Eq. (21) defines Q2 as a joint KL that is not the conditional KL appearing in the implemented loss. As written, the supplementary proof is not internally consistent.","section":"Appendix, Eqs. (18)-(22)"},{"comment":"Even if Eq. (4) were established, the loss LSU = min exp(-DKL[PF || PF\\fi]) encourages each individual feature fi to alter the posterior label distribution, which is a non-redundancy or necessity property rather than an obvious comprehensiveness property of the fused feature F. The paper does not explain why maximizing the posterior shift caused by removing fi guarantees that F contains maximal task-related information; this conceptual gap should be addressed or the claims should be tempered.","section":"Method, Eq. (5)"}],"minor_comments":[{"comment":"The text reads 'The sufficient-view constraint'; this should be 'sufficiency-view' for consistency with the rest of the paper.","section":"Method, Overview"},{"comment":"The claim that SUMI-IFL 'consistently ranks among the top two' is not supported on the CIMD AUC row, where WSCL achieves 0.6273 and SUMI-IFL achieves 0.5513, placing SUMI-IFL third on that metric.","section":"Table 3 and 'Cross-dataset comparisons'"},{"comment":"The sentence 'resulting in an F1 increase of 9.8% and 5.1% on DEFACTO-12 and SSRGFD, respectively' does not match the table: removing LSU drops F1 from 0.9249 to 0.8335 on DEFACTO-12 and from 0.7995 to 0.6783 on SSRGFD, while removing LMI drops F1 from 0.9249 to 0.8774 on DEFACTO-12 and from 0.7995 to 0.6784 on SSRGFD; the reported percentages should be reconciled.","section":"Ablation study, Table 4"},{"comment":"The expectation notation E_{p(f,m)p(z|m)} in Eq. (9) appears to be a typo; the subsequent equations use E_{p(f)p(z|f)}, which is presumably the intended measure.","section":"Method, Eq. (9)"},{"comment":"The email address for the second affiliation appears garbled as 'luoxy ieu@sina.comthird'.","section":"Author block"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are strong and the ablation study gives evidence that the proposed losses help, but the paper's advertised 'rigorous theoretical analysis' is not presently sound. The authors should either supply a valid derivation of Eq. (4) under explicit assumptions, or substantially reframe the theoretical discussion as a heuristic motivation. Given the centrality of the information-theoretic claim to the paper's framing, I cannot recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the empirical IFL work here is strong and likely worth building on, but the information-theoretic scaffolding that gives the paper its title does not hold up. The framework is new, the results are large, and the ablations support the losses. But the derivation of the sufficiency constraint contains a false inequality, an unproven bridge, and an appendix that proves a different bound than the one used. That is a load-bearing flaw, because the paper advertises the theory as rigorous.\n\nWhat's new: three attention U-Nets (channel, spatial, pixel), a learned fusion layer, and two KL-based losses, one on the extraction side and one on the reasoning side. On DEFACTO-12, SSRGFD, and CASIAv2, SUMI-IFL beats the prior best by 8.5, 15.7, and 2.6 points in F1; on NIST16 cross-dataset it leads by 9.7. The ablation shows both losses contribute several F1 points. That is a solid empirical package, and the authors compare against relevant baselines.\n\nWhere it breaks: Eq. (3) replaces the chain-rule conditioning sets with leave-one-out sets and claims expanding the known conditions can only increase conditional mutual information. That is false in general; the stress-test counterexample (identical fair bits) is valid. The paper offers no independence or diversity assumption that would save it. Eq. (4) is asserted as a variational bound but is not proven; the appendix instead derives a lower bound involving D_KL[p(M,Fi) || p(M,F\\fi)], which is not the conditional KL D_KL[p(M|F) || p(M|F\\fi)] that appears in the implemented loss. So minimizing exp(-KL) has not been shown to maximize I(M;F). The minimality-view constraint is more standard (a CEB-style bound), though the paper doesn't fully address the log p(z|f) term.\n\nAlso minor: no code, no error bars, and the noise mask details and hyperparameters are under-specified. The ablation would be more informative with a baseline that removes the theory and just uses the architectural pieces.\n\nNet: the empirical contribution deserves a serious referee; the theoretical contribution as stated does not. I'd send it to review, but the referee should require either a corrected derivation or an honest reframing of the losses as heuristics. If the authors can do that, the paper could be solid. If they can't, the results still stand on their own as empirical findings, but the advertised theory needs to be removed or demoted.","headline":"Strong empirical gains undercut by a broken sufficiency-view derivation; the advertised theory doesn't support the losses, but the experiments deserve a referee.","tokens_in":15347,"tokens_out":2649,"would_cite":false,"duration_ms":29040,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SUMI-IFL claims that sufficiency and minimality constraints on forgery features—derived from mutual information and the information bottleneck—produce more comprehensive and accurate tamper localization than existing methods.","keywords":["image forgery localization","mutual information","information bottleneck","sufficiency constraint","minimality constraint","multi-view feature fusion","attention U-Net","tampering detection"],"falsifier":"Train two models identical except that one includes the sufficiency loss and one does not, then estimate $I(M;F)$ directly on a small test set using a neural mutual information estimator; if the model with the loss does not show higher estimated mutual information, the loss is not doing what the theory claims. Alternatively, search for a simple counterexample to the inequality in Eq. (4) by sampling from a toy distribution where the KL term is large but the conditional mutual information is small, which would invalidate the theoretical link.","tokens_in":14262,"feed_emoji":"🔍","tokens_out":11051,"duration_ms":91233,"temperature":0.7,"pith_summary":"The paper introduces SUMI-IFL, an image forgery localization framework that constrains the learned feature representation from two information-theoretic directions. A sufficiency constraint, derived from maximizing mutual information between the latent feature and the ground-truth mask, aims to keep all forgery-relevant clues; a minimality constraint, derived from the conditional entropy bottleneck, aims to discard task-unrelated information. The authors argue that this combination addresses the persistent incompleteness and redundancy of forgery features, and they report state-of-the-art F1 and AUC scores on DEFACTO-12, SSRGFD, and CASIAv2, with competitive cross-dataset results on CIMD and NIST16. The potential significance is a principled, theory-driven recipe for building detectors that generalize across datasets and resist post-processing distortions.","feed_headline":"Two constraints push forgery localization past prior art","feed_subtitle":"Sufficiency and minimality constraints sharpen tampered-region localization on benchmarks.","key_machinery":"The central mechanism is a pair of loss terms placed on two stages of the network. For sufficiency, since direct estimation of $I(M;F)$ is intractable, the paper optimizes a variational lower bound that compares the prediction distribution with and without each individual view: $L_{SU} = \\min \\exp(-D_{KL}[P_F \\| P_{F \\setminus f_i}])$, where $P_F = p(y|F)$ and $F \\setminus f_i$ is the fusion of all views except $f_i$. For minimality, the paper adapts the conditional entropy bottleneck and uses variational inference to obtain $L_{MI} = E_{p(f)p(z|f)}[KL(p(z|f) \\| q(z|m))]$, where $q(z|m)$ is obtained by encoding the ground-truth mask into the forgery feature space. These constraints are applied around a feature extractor of three attention U-Nets (channel, spatial, and pixel views) fused by a learnable layer $B_\\phi$, and a mask-guided encoder-decoder that produces the final concise feature and an auxiliary noise-perturbed mask prediction.","core_discovery":"The central claim is that imposing two losses on the forgery feature pipeline makes the learned representation both comprehensive and concise. The sufficiency loss, $L_{SU} = \\min \\exp(-D_{KL}[P_F \\| P_{F \\setminus f_i}])$, is applied to the feature extraction network and is claimed to lower-bound the task-related mutual information $I(M;F)$, ensuring that no forgery clue is missed. The minimality loss, $L_{MI} = E_{p(f)p(z|f)}[KL(p(z|f) \\| q(z|m))]$, is applied to the feature reasoning network and is claimed to remove task-unrelated information by pulling the feature distribution toward a mask-encoded variational distribution. The paper asserts that this information-theoretic grounding yields a latent forgery feature that is both sufficient and minimal, and it reports that the resulting model outperforms seven state-of-the-art methods in in-dataset, cross-dataset, and robustness evaluations.","pith_inferences":["If the sufficiency loss is replaced by a direct mutual-information estimator on small-scale problems, the performance gap could reveal whether the theoretical bound or the multi-view architecture drives the gain.","The same two-stage sufficiency/minimality scheme could transfer to other dense prediction tasks, such as anomaly localization or medical image segmentation, where multi-view features are fused and label-guided compression is desirable.","Because the fusion layer uses a single scalar weight per view, the framework implicitly assumes that one global weighting is enough; per-pixel or per-channel fusion might further improve comprehensiveness.","The noise-perturbed auxiliary mask training suggests a broader principle: injecting structured noise into the label path can make the variational encoder more robust, a technique that may generalize beyond forgery localization."],"forward_implications":["Localization models trained with both constraints should be more robust to JPEG compression and Gaussian blur, because task-unrelated artifacts are explicitly suppressed.","The sufficiency constraint offers a direct way to combine multiple forensic views without requiring orthogonality, so complementary clues can be pooled.","The minimality constraint shows how ground-truth masks can guide feature compression through a learned variational mapping, reducing false alarms from unrelated image content.","The reported cross-dataset results suggest that the constraints improve generalization to unseen tampering types and unseen dataset distributions.","The two losses are modular and could be added to other forgery localization architectures as auxiliary training objectives."],"supporting_citations":[{"why":"Supplies the information bottleneck objective from which the minimality constraint is derived.","marker":"(Tishby, Pereira, and Bialek 2000)"},{"why":"Introduces the conditional entropy bottleneck that the minimality loss adapts.","marker":"(Fischer 2020)"},{"why":"Provides the variational lower bound used to make the bottleneck objectives tractable.","marker":"(Alemi et al. 2016)"},{"why":"Cites challenges in direct mutual information estimation, motivating a variational surrogate.","marker":"(Poole et al. 2019)"},{"why":"U-Net architecture is the backbone for the three attention-based feature extractors.","marker":"(Ronneberger, Fischer, and Brox 2015)"},{"why":"DN-DETR's box denoising training inspires the noise-perturbed auxiliary mask design.","marker":"(Li et al. 2022)"},{"why":"Provides the DEFACTO-12 dataset used in training and in-dataset evaluation.","marker":"(Mahfoudi et al. 2019)"},{"why":"Provides the SSRGFD stereo forgery dataset used in training and evaluation.","marker":"(Yin et al. 2023)"},{"why":"Provides the CASIAv2 dataset used in training and in-dataset evaluation.","marker":"(Dong, Wang, and Tan 2013)"},{"why":"MVSS-Net is a key multi-stream baseline whose edge/noise supervision the paper compares against.","marker":"(Dong et al. 2022)"}],"fun_headline_variants":["Information-theoretic constraints sharpen forgery localization","Sufficiency and minimality: Key to better forgery detection","Two losses make forgery localization more precise","SUMI-IFL: A new info-theoretic approach to spotting fakes","Minimal yet sufficient: Better forgery localization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the sufficiency loss really does maximize how much the latent feature tells us about the forgery mask; the paper's derivation relies on an inequality about statistical distances that is not generally true.","fun_headline_variants_meta":{"raw":{"variants":["Information-theoretic constraints sharpen forgery localization","Sufficiency and minimality: Key to better forgery detection","Two losses make forgery localization more precise","SUMI-IFL: A new info-theoretic approach to spotting fakes","Minimal yet sufficient: Better forgery localization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1665,"prompt_tokens":948,"completion_tokens":717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":638}},"tokens_in":564,"tokens_out":717,"duration_ms":6960,"temperature":1.0,"reasoning_tokens":638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:29:02.050646+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two models identical except that one includes the sufficiency loss and one does not, then estimate $I(M;F)$ directly on a small test set using a neural mutual information estimator; if the model with the loss does not show higher estimated mutual information, the loss is not doing what the theory claims. Alternatively, search for a simple counterexample to the inequality in Eq. (4) by sampling from a toy distribution where the KL term is large but the conditional mutual information is small, which would invalidate the theoretical link.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the conditional entropy bottleneck that the minimality loss adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cites challenges in direct mutual information estimation, motivating a variational surrogate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"U-Net architecture is the backbone for the three attention-based feature extractors."},{"cited_title":"L.; and Pic, M","cited_arxiv_id":null,"evidence_quote":"Provides the DEFACTO-12 dataset used in training and in-dataset evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SSRGFD stereo forgery dataset used in training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CASIAv2 dataset used in training and in-dataset evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MVSS-Net is a key multi-stream baseline whose edge/noise supervision the paper compares against."}],"review_version":1}