{"id":"9822df5c-8f3d-4b5d-81f1-ddc61a6dc60b","arxiv_id":"2507.01712","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Comparing camera fingerprints directly in the wavelet domain, without reconstructing the image, improves source-camera identification accuracy and speed.","lead":"This paper speeds up camera identification by comparing sensor fingerprints directly in the wavelet domain, skipping a costly image reconstruction step. The method reports slightly better matching accuracy and much faster processing on a standard image database, which could help large-scale forensic searches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains of gray-WDLAW may stem from grayscale-first pre-processing rather than the wavelet-domain representation; the paper lacks the decisive gray-LAW baseline.","rationale":"The paper's strongest claim is that the proposed method is both more accurate and faster because comparisons are done in the wavelet domain. The evidence for this is a single small experiment, and the main method bundles the wavelet-domain change with a grayscale-first change. Since grayscale-first is already documented in the PRNU literature and reduces the number of wavelet decompositions by a factor of three, the reported runtime improvement is largely explainable without any wavelet-domain innovation. The accuracy improvement could likewise be a property of grayscale-first rather than of the wavelet representation. The paper does include WDLAW-gray, which keeps three-channel processing and only varies the wavelet-domain aspect; however, even that variant changes where grayscaling occurs and omits zero-mean normalization, so it is not a clean control. A gray-LAW baseline would settle the attribution. This is a more direct threat to the central claim than the statistical issues noted by the reader, because it questions whether the proposed modification has any causal effect at all. The reader partly recognized the confound in the rationale (mentioning grayscale-first as an explanation for the speedup) but did not make it the main concern. I therefore recommend keeping the conditional verdict, with the required control added to the conditions.","tokens_in":13553,"tokens_out":10028,"duration_ms":112984,"concrete_test":"Implement gray-LAW: convert each RGB image to grayscale first, then run the original LAW pipeline (DWT + Mihcak filtering + IDWT, no color combination needed, then zero-mean normalization and Fourier-domain Wiener filtering in the image domain) using the same parameters (db4, J=4, sigma_n=1.82). Compare gray-LAW, LAW, and gray-WDLAW on the same 26-camera subset using camera-level cross-validation for threshold selection. If gray-LAW matches gray-WDLAW in accuracy and runtime, the wavelet-domain representation provides no benefit. Also run WDLAW-gray against a corresponding image-domain 'LAW-gray' baseline (per-channel filtering, grayscale after wavelet-domain filtering, then image-domain artifact removal) to isolate the effect of omitting the inverse transform.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim attributes higher accuracy and lower runtime to performing fingerprint comparison in the wavelet domain. However, the headline method gray-WDLAW differs from the LAW baseline in two ways at once: it converts to grayscale before decomposition, and it omits the inverse wavelet transform. Grayscale-first is a known pre-processing choice (the paper cites [2] for it) and reduces extraction from three DWT passes to one; this alone can explain most of the runtime drop in Table 2 (1.127s to 0.502s). No comparison is made against a 'gray-LAW' baseline that applies grayscale-first but keeps the original image-domain reconstruction and Fourier Wiener filtering. Without this control, the accuracy and speed advantages cannot be attributed to wavelet-domain fingerprints, which is the paper's core claim. The included WDLAW-gray variant is closer to a control but still changes two factors relative to LAW; the residual gain (86s vs 96s total) is small and could be noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes representing camera sensor pattern noise fingerprints directly in the wavelet domain, avoiding the inverse DWT and image-domain reconstruction, and comparing the resulting wavelet coefficients with cosine similarity. Three DWT-based variants (gray-WDLAW, rgb-WDLAW, WDLAW-gray) and one DTCWT-based variant (gray-WDLAW-DTCWT) are introduced. The methods are evaluated on a subset of the Dresden Image Database (130 images from 26 cameras, 8385 pairwise comparisons) against the LAW and DTCWT baselines, reporting TPR/TNR, TPR at TNR=0.99, AUC, and runtimes. The paper claims that wavelet-domain fingerprints give higher or comparable accuracy while substantially improving processing speed.","tokens_in":13727,"tokens_out":11186,"duration_ms":128339,"significance":"If the empirical claims were established, the idea would be practically useful: removing the inversion step and comparing directly in the wavelet domain could reduce runtime in large-scale forensic searches while preserving the discriminative PRNU signal. The algorithm descriptions are clear and the methods build on publicly available implementations, which aids reproducibility. However, the current experimental evidence does not isolate the wavelet-domain mechanism from the grayscale-first pre-processing change, the evaluation protocol is self-referential because thresholds are chosen on the test set, and the stated dimensionality-reduction argument is quantitatively wrong for the J=4 setting used in the experiments. The contribution is therefore plausible but not yet demonstrated.","major_comments":[{"comment":"The headline comparison is confounded. gray-WDLAW differs from LAW in two simultaneous changes: grayscale conversion is moved before the DWT, and the inverse DWT/Fourier-domain Wiener processing is replaced by wavelet-domain filtering without reconstruction. The paper itself cites [2] for the grayscale-first idea, so this change alone can account for much of the extraction-time drop (1.127 s to 0.502 s). No gray-LAW baseline (grayscale-first but with image-domain reconstruction and Fourier-domain Wiener filtering) is reported. Without this control, the claimed accuracy and speed advantages cannot be attributed to the wavelet-domain fingerprint representation, which is the paper's central claim.","section":"§3.2, §4.2.1, Table 2"},{"comment":"The thresholds λ_YI and the TNR=0.99 threshold are selected from the ROC curve of the same 8385 pairwise comparisons on which TPR, TNR, and AUC are then reported. This makes the reported accuracy metrics optimistic by construction; no held-out validation, cross-validation, or per-camera/image-level bootstrap is provided. With only 260 same-source pairs (26 cameras × C(5,2)), the TPR differences of 0.03–0.07 in Table 1 are also within plausible sampling variability, and no error bars are given. Please report validation on held-out data, for example splitting by camera, and provide confidence intervals for all accuracy metrics.","section":"§2.3 and §4.1"},{"comment":"The first displayed expression for l is correct, but the simplification and the surrounding claim are not. The sum evaluates to m²(1−4^{−J}); for the J=4 setting used in the experiments this is 255/256 m² ≈ 0.996 m², not 'about 75%'. Thus the retained wavelet-domain fingerprint is only about 0.4% smaller than the image-domain fingerprint, and the 10× comparison-time reduction in Table 2 (0.010 s vs 0.001 s) cannot be explained by the claimed dimensionality reduction. Please correct the formula and reconcile the timing figures with the actual coefficient count.","section":"§3.1, Eq. (6)"},{"comment":"The timing columns are internally inconsistent. If extraction is 1.127 s per pair, 8385 pairs alone would require about 9,450 s, not the reported 96.7 s total; if it is per image, 130 images require about 146.5 s, still exceeding the reported total. Please clarify whether the extraction time is per image, per pair, or averaged differently, and report total extraction time and total comparison time separately so that the speed claims can be evaluated.","section":"Table 2"}],"minor_comments":[{"comment":"The sentence 'since TNR(λ) = 1 − TPR(λ)' is incorrect; TNR = 1 − FPR, and the threshold for TNR = R should set FPR(λ) = 1 − R. The same paragraph also defines TNR as the proportion of actual positives, which should be actual negatives.","section":"§2.3"},{"comment":"The paper states that omitting the row-/columnwise zero-mean operation 'does not impact the final results,' but no supporting ablation is shown. Please include the comparison with and without this step.","section":"§3.1"},{"comment":"The output notation f4l is used in Algorithm 4 before being defined; please introduce it explicitly, for example next to Eq. (6), and clarify the relationship between the six complex-valued DTCWT subimages and the real vector of length 4l.","section":"§3.3, Algorithm 4"},{"comment":"The paper cites the base PRNU and DTCWT implementations but does not state whether the modified WDLAW/WDLAW-DTCWT code will be made available; please add a data and code availability statement.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is worth pursuing, but the abstract's accuracy and speed claims currently outrun the evidence. The decisive missing experiment is the gray-LAW control; without it, the reader cannot tell whether the improvements come from wavelet-domain fingerprints or from the known grayscale-first pre-processing choice. The evaluation also needs a proper validation strategy, since thresholds are selected on the same comparisons used for reporting. These issues are fixable within the scope of a revision, so I would not reject the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a correct and useful idea — because cosine similarity is invariant to orthonormal transforms, you can keep the PRNU fingerprint as wavelet detail coefficients and skip the inverse DWT. That is a sensible efficiency trick for large-scale source camera identification, and I don't recall seeing it applied this way in the PRNU literature. The reported numbers go in the right direction and the DTCWT extension is a nice consistency check.\n\nBut the empirical core is less solid than the narrative. The headline comparison is gray-WDLAW vs. LAW, and these differ in two ways at once: gray-WDLAW converts to grayscale before the DWT, and it omits the inverse wavelet transform. The paper cites [2] for grayscale-first, which is a known preprocessing choice that reduces extraction from three DWT passes to one. That alone explains most of the runtime drop (1.127s to 0.502s per pair) and may well explain the small accuracy gain. There is no gray-LAW baseline — grayscale-first but keeping image-domain reconstruction and the second Wiener filter in the image domain — so the specific claim that wavelet-domain comparison is what helps is not actually tested. The included WDLAW-gray variant is closer to a control, but it still keeps the no-inversion change, and its total runtime (86s vs 96s) is close enough to LAW that the difference could be noise.\n\nA second issue: the decision thresholds are selected on the same 8385 pairwise comparisons used to report TPR, TNR, and AUC. No held-out test or cross-validation; no error bars or per-camera variability. That makes the accuracy advantages look overfit. Also, Equation (6) is wrong for J=4, which is the value used in the experiments: the sum gives about 0.996 m^2, not ~75% of the pixels, so the claimed dimension reduction for comparison is negligible. The comparison time drop from 0.010s to 0.001s therefore cannot be attributed to fewer coefficients. It might be an implementation artifact or noise.\n\nThe paper also asserts without evidence that zero-mean filtering has no impact (\"our analysis ... shows that this step does not impact the final results\"), but no ablation is shown.\n\nNone of this kills the central idea. The orthonormal-invariance argument is sound, the method is clearly described, and the code is based on open-source PRNU implementations. A serious revision that adds the gray-LAW control, uses held-out thresholds, reports variability, and fixes the dimension analysis would make this a useful contribution. As it stands, the claimed speed and accuracy advantages are plausible but not properly attributed.\n\nI'd send it to peer review, with the expectation of substantial revision. The right readers are forensic image analysts who care about PRNU extraction pipelines and large-scale comparison.","headline":"Useful efficiency idea, but the main experiment conflates grayscale-first with wavelet-domain comparison; needs a gray-LAW control before the gains can be attributed as claimed.","tokens_in":14261,"tokens_out":3446,"would_cite":false,"duration_ms":36379,"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":"By comparing camera fingerprints directly in the wavelet domain and skipping the inverse transform, the proposed WDLAW method improves source-camera identification accuracy and speed on tested real-world images.","keywords":["sensor pattern noise","source camera identification","wavelet domain fingerprint","discrete wavelet transform","PRNU","dual-tree complex wavelet transform","cosine similarity","image forensics"],"falsifier":"Run gray-WDLAW and the standard LAW method on the full Dresden Image Database (or another multi-camera corpus) under the same settings (db4 wavelet, decomposition level J=4, noise variance sigma_n=1.82) and compare TPR at TNR=0.99, AUC, and wall-clock time. If gray-WDLAW does not match or beat LAW on accuracy and speed there, the claim that the wavelet-domain pipeline preserves discriminative SPN information would be refuted.","tokens_in":13311,"feed_emoji":"📷","tokens_out":3291,"duration_ms":36387,"temperature":0.7,"pith_summary":"The paper claims that a camera's sensor pattern noise (SPN) can be compared as a wavelet-domain fingerprint, without ever rebuilding the full image-domain fingerprint. This removes the inverse wavelet transform from the extraction pipeline, shrinking the fingerprint and lowering comparison cost. On a subset of the Dresden Image Database, the proposed gray-WDLAW method improved both identification accuracy and processing speed over the standard LAW method. The same idea also transfers to a dual-tree complex wavelet version, offering comparable accuracy with faster extraction. The practical goal is large-scale image-to-database camera matching, where computational efficiency matters.","feed_headline":"Wavelet-domain fingerprints speed up camera ID and boost accuracy","feed_subtitle":"Keeping sensor fingerprints as wavelet coefficients avoids an inversion step, cutting runtime and improving accuracy in tests.","key_machinery":"The central object is the wavelet-domain fingerprint: a vector formed by concatenating the filtered detail coefficients (horizontal, vertical, diagonal at each level) from a discrete wavelet transform, or the six complex coefficient subimages from a dual-tree complex wavelet transform. The argument rests on two facts: the DWT is an orthonormal linear transform, so by Parseval's theorem signal energy is preserved, and cosine similarity is invariant under orthonormal transformations. These make the inverse wavelet transform informationally redundant for comparison purposes, allowing the pipeline to skip reconstruction and apply the second Wiener filter directly to the wavelet coefficients per subimage.","core_discovery":"The paper's central claim is that the final inversion step of wavelet-based PRNU extraction is unnecessary for source camera identification: an orthonormal wavelet transform preserves the cosine similarity used for fingerprint comparison, so the filtered detail coefficients can serve directly as the fingerprint. The proposed wavelet-domain fingerprint, called WDLAW, applies locally adaptive Mihcak filtering and then Fourier-domain Wiener filtering to each wavelet subimage, concatenates the coefficients, and compares them with cosine similarity. This reduces the fingerprint dimension to less than the full image size (about 75% for small decomposition levels) and skips the inverse DWT and the zero-mean artifact-removal step. In experiments on 26 cameras from the Dresden Image Database, gray-WDLAW achieved higher TPR/TNR and AUC than the standard LAW method while reducing average extraction time from 1.127 to 0.502 seconds per pair and total runtime from 96.7 to 37.8 seconds. A DTCWT-based variant, gray-WDLAW-DTCWT, also matched or exceeded the accuracy of its image-domain counterparts while cutting extraction time.","pith_inferences":["A testable extension is to replace the Daubechies-4 wavelet with other orthonormal or biorthogonal wavelets; if the cosine-invariance argument is the true driver, the speed benefit should persist while accuracy may vary with the basis.","The reported accuracy gain may partly come from discarding the approximation coefficients, which acts as an aggressive high-pass operation; comparing against an image-domain high-passed PRNU could isolate that effect.","The speed advantage is most pronounced when comparison volume is large; on very small databases, the removed inversion step may be negligible, so the method's practical benefit is workload-dependent.","Storing fingerprints as wavelet coefficients also reduces storage, which is an implicit benefit for large-scale forensic databases that the paper notes but does not quantify."],"forward_implications":["Source-camera identification pipelines can store and compare fingerprints that are only about 75% of the image size, reducing memory and per-comparison cost.","The gray-WDLAW method reported TPR/TNR of 0.90/0.97 versus 0.87/0.95 for LAW, with TPR at TNR=0.99 improving from 0.80 to 0.87, showing a concrete accuracy gain on the tested data.","Total runtime for the DWT-based comparison dropped from about 96.7 seconds to 37.8 seconds on the 8385-pair test set, a speed gain that should scale to larger databases.","The same skip-inversion principle extends to the dual-tree complex wavelet transform, where the gray-WDLAW-DTCWT variant reached an AUC of 0.97 with faster extraction than the original DTCWT method.","Because the modification is structural rather than tied to a specific denoising filter, other wavelet-domain filters could be plugged into the same fingerprint framework."],"supporting_citations":[{"why":"Supplies the foundational PRNU extraction method and the LAW wavelet-denoising pipeline that the proposed WDLAW modifies.","marker":"[29]"},{"why":"Establishes sensor noise for image origin and the artifact-removal steps (zero-mean normalization and Wiener filtering) whose necessity the paper re-examines.","marker":"[11]"},{"why":"Provides the low-complexity locally adaptive Mihcak denoising filter used on the wavelet coefficients.","marker":"[33]"},{"why":"Supplies the spatially adaptive statistical model and Wiener-like shrinkage equation that the Mihcak filter is based on.","marker":"[32]"},{"why":"Provides the DTCWT-based SPN extraction method that the paper extends into the wavelet-domain gray-WDLAW-DTCWT.","marker":"[45]"},{"why":"Supplies the Dresden Image Database used for the experiments and the comparison workload.","marker":"[19]"},{"why":"Supports the grayscale-first processing choice and the discussion of shared-component artifacts in SPN fingerprints.","marker":"[2]"},{"why":"Provides the Parseval/orthonormality background that justifies comparing fingerprints in the wavelet domain rather than after inversion.","marker":"[13]"}],"fun_headline_variants":["Wavelet-domain fingerprints skip inversion, speed up camera ID","No inversion needed: wavelet fingerprints cut time, boost accuracy","Direct wavelet fingerprints make camera ID faster and sharper","Camera fingerprinting gets a wavelet shortcut: faster, more accurate","Wavelet-domain SPN: bypass inversion for quicker, better camera ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central bet is that the nonlinear processing sequence still works when moved into the wavelet domain: discarding the approximation coefficients, applying Fourier-domain Wiener filtering per wavelet subimage, and converting color to grayscale before decomposition do not destroy the camera-specific signal. Only one small empirical comparison supports this, so if the sequence fails on other data the reported accuracy and speed advantages would not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet-domain fingerprints skip inversion, speed up camera ID","No inversion needed: wavelet fingerprints cut time, boost accuracy","Direct wavelet fingerprints make camera ID faster and sharper","Camera fingerprinting gets a wavelet shortcut: faster, more accurate","Wavelet-domain SPN: bypass inversion for quicker, better camera ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3674,"prompt_tokens":864,"completion_tokens":2810,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":2727}},"tokens_in":480,"tokens_out":2810,"duration_ms":25161,"temperature":1.0,"reasoning_tokens":2727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:44:55.008371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run gray-WDLAW and the standard LAW method on the full Dresden Image Database (or another multi-camera corpus) under the same settings (db4 wavelet, decomposition level J=4, noise variance sigma_n=1.82) and compare TPR at TNR=0.99, AUC, and wall-clock time. If gray-WDLAW does not match or beat LAW on accuracy and speed there, the claim that the wavelet-domain pipeline preserves discriminative SPN information would be refuted.","supporting_citations":[{"cited_title":"Digital camera identification from sensor pattern noise","cited_arxiv_id":null,"evidence_quote":"Supplies the foundational PRNU extraction method and the LAW wavelet-denoising pipeline that the proposed WDLAW modifies."},{"cited_title":"Determining image origin and integrity using sensor noise","cited_arxiv_id":null,"evidence_quote":"Establishes sensor noise for image origin and the artifact-removal steps (zero-mean normalization and Wiener filtering) whose necessity the paper re-examines."},{"cited_title":"Low- complexity image denoising based on statistical modeling of wavelet coeffi- cients","cited_arxiv_id":null,"evidence_quote":"Provides the low-complexity locally adaptive Mihcak denoising filter used on the wavelet coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the spatially adaptive statistical model and Wiener-like shrinkage equation that the Mihcak filter is based on."},{"cited_title":"Source camera identification with dual-tree complex wavelet transform","cited_arxiv_id":null,"evidence_quote":"Provides the DTCWT-based SPN extraction method that the paper extends into the wavelet-domain gray-WDLAW-DTCWT."},{"cited_title":"The’Dresden Image Database’for benchmarking digital image forensics, in: Proceedings of the 2010 ACM Symposium on Applied Computing, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the Dresden Image Database used for the experiments and the comparison workload."},{"cited_title":"On the SPN estimation in image forensics: A systematic empirical evaluation","cited_arxiv_id":null,"evidence_quote":"Supports the grayscale-first processing choice and the discussion of shared-component artifacts in SPN fingerprints."},{"cited_title":"An Introduction to Wavelets","cited_arxiv_id":null,"evidence_quote":"Provides the Parseval/orthonormality background that justifies comparing fingerprints in the wavelet domain rather than after inversion."}],"review_version":1}