{"id":"a8cb6145-ec40-4ff6-93c1-c31c99387030","arxiv_id":"2608.01737","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"IDraw improves image-only artist verification by training on tablet-pen sensor signals from separate artists and subtracting object content before comparing drawings.","lead":"A new system called IDraw verifies who made a digital drawing by learning, during training, how pen pressure and movement show up in finished images, then comparing a disputed drawing with an artist's known works. It also removes what same-object drawings share, so comparisons reflect the artist rather than the subject.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Device heterogeneity in the collected sensor signals may confound the behavior supervision in Eq. (2), so the reported gains could reflect device fingerprints rather than artist behavior; a cross-device evaluation would settle this.","rationale":"The paper's strongest claim is that IDraw consistently beats image-only verification and that this comes from inferring drawing behavior and suppressing content. The data collection design--participants use their own iPads/Pencils--creates a natural experiment in which sensor features and image renderings are both device-dependent. If sensor features vary by device, the behavior loss in Eq. (2) can be minimized without recovering artist motor behavior: the image encoder only needs to predict device-identifying sensor deltas. Because every drawing by one artist comes from the same device, artist and device are perfectly confounded in training. This would not invalidate the paper's engineering result on the collected data, but it would invalidate the behavioral interpretation and raise doubt about generalization to other hardware. The fact that M3's largest gains are on ResNet-50, a model that relies on local texture, is consistent with the model picking up rendering fingerprints. Cross-device evaluation is therefore the decisive experiment. The object-mean concern (Eq. (5)) is secondary: the mean is estimated from training artists, so it cannot leak unseen-artist identity, though it may be noisy; the device confound is more fundamental. Single-seed training and overlapping splits are also secondary; they affect error bars but not the mechanism. I therefore agree with the reader's conditional verdict: the paper is promising and the ablation structure is reasonable, but a device-control analysis is needed before the behavioral claim and the 40% headline can be accepted.","tokens_in":12207,"tokens_out":7184,"duration_ms":72220,"concrete_test":"Partition the 37 artists by iPad model / Apple Pencil generation (the supplementary material should record device metadata). Train IDraw on artists in device group A and evaluate on artists in device group B, and vice versa, reusing the 15-split protocol. If the mean IDraw AUC gain over baseline at N=1 drops from the reported +0.138 (or from the per-encoder gains in Table 2) to near zero, the gains are substantially driven by device fingerprints rather than transferable artist behavior. As a complementary check, train a linear classifier on the standardized 50 behavior features to predict device identity with artist-disjoint folds; chance-level accuracy would weaken the concern, while near-perfect accuracy would confirm it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"IDraw's central claim is that M3's behavior loss (Eq. (2)) -- regressing standardized sensor-feature differences Δb from normalized image-embedding differences Δz -- teaches the image encoder to recover artist-specific drawing behavior from completed images. This requires the 50 behavior features in M2 to be informative about the artist's motor behavior. But the sensor features are collected on each participant's own iPad and Apple Pencil (Data Collection). Different iPad models and Pencil generations have different pressure curves, tilt reporting, event sampling, and screen properties, so features such as average pressure, timing, and stroke geometry are plausibly device-identifying. The completed drawing images are also rendered and captured on those same devices, so the image encoder can learn visual fingerprints of the device (resolution, aspect ratio, line rasterization, color profile) that correlate with the sensor features. Because each artist uses one device throughout, artist identity and device identity are confounded in the training pairs. M3 can then satisfy the regression by encoding device-specific rendering differences, not motor behavior. This would inflate the Table 2 gains and the Figure 4(b) 'sensor upper bound' comparison, and it would break the paper's behavioral interpretation and its generalization to new artists' devices. The paper reports no device-type control or per-participant device metadata, and the selected feature set explicitly includes pressure and timing, the most device-sensitive families. The secondary concerns about single-seed training and overlapping splits are real but less load-bearing than this confound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IDraw, a framework for verifying the authorship of digital drawings from completed images only. During training, IDraw uses tablet-pen sensor signals recorded from separate training artists as privileged supervision: an image encoder is trained with a pairwise behavior loss (Eq. 2) that regresses differences in standardized sensor-derived behavior features from differences in normalized image embeddings, and a triplet loss (Eq. 3) that pulls same-artist different-object drawings together and pushes same-object different-artist drawings apart. At verification time, the method subtracts per-object mean embeddings estimated from training artists (Eq. 5), aggregates reference drawings, and uses cosine distance to the query. The authors collect a fully crossed dataset of 1,110 drawings from 37 artists drawing 30 objects, each paired with 14 types of tablet-pen sensor signals. They evaluate on unseen artists across nine image encoders and six reference-set sizes, reporting consistent AUC/EER gains over a cross-entropy baseline, including a reduction in verification error of up to 40% and an average AUC gain of 0.138 at N=1. Ablations show that behavior-guided supervision, triplet loss, and content suppression each contribute to the gains.","tokens_in":12510,"tokens_out":4059,"duration_ms":42696,"significance":"If the results hold, the paper makes a useful contribution: it demonstrates a learning-using-privileged-information approach to offline drawing verification and introduces a dataset that is, to my knowledge, the first fully crossed artist-object drawing dataset with tablet-pen sensor signals. The experimental sweep over nine encoders and multiple reference-set sizes is unusually thorough, and the module-level ablations (Tables 3 and 4) provide a clear picture of where the gains come from. The main contributions are potentially valuable for practical authorship disputes where only completed images are available. However, the central behavioral interpretation rests on the assumption that the sensor-signal supervision is indicative of artist-specific drawing behavior rather than device-specific artifacts; the current manuscript does not provide the control experiments needed to establish this. The statistical significance claims are also overstated because the 15 splits are not independent. The work is significant but needs substantial strengthening before its core claim can be accepted.","major_comments":[{"comment":"The sensor signals are recorded from each participant's own iPad and Apple Pencil, so device identity is completely confounded with artist identity in the training data. The 50 selected behavior features include pressure, timing, and stroke geometry, all of which are plausibly affected by device model, Pencil generation, pressure-curve calibration, sampling rate, and rendering pipeline. Because the completed drawing images are rendered and captured on the same device, the image encoder can learn visual fingerprints of the device (e.g., resolution, line rasterization, color profile) that correlate with the sensor features. The M3 regression in Eq. (2) can then be satisfied by encoding device-specific rendering differences rather than motor behavior. The paper reports no device-type metadata, no per-device analysis, and no cross-device evaluation. This is load-bearing because the paper's central claim is that IDraw infers behavior-related artist cues and will generalize to previously unseen artists; if the gains instead reflect device fingerprints, the behavioral interpretation and the generalization claim collapse. A minimum fix would be a leave-one-device-out or cross-device experiment, or at least a demonstration that the gains persist when device information is controlled.","section":"Data Collection; Eq. (2)"},{"comment":"The significance testing is not valid as reported. The 15 artist-disjoint splits are not independent: each split selects 23 fit, 6 validation, and 8 verification artists from the same fixed pool of 37 artists, so the same artists and drawings appear across splits. Paired t-tests across these 15 splits therefore treat 15 dependent measurements as independent, and the reported p-values below 10^-6 (for all 54 combinations) overstate the statistical evidence. In addition, all experiments use a single training seed. The central claim that IDraw consistently outperforms the baseline would be much better supported by reporting variance across multiple seeds or by using an artist-level bootstrap that respects the resampling structure. This is a load-bearing point because the paper's 'consistent' improvement is asserted largely through these significance claims.","section":"Evaluation Protocol; Overall Results"},{"comment":"The object mean mu_o in Eq. (5) is estimated from the 23 training artists and is then subtracted from embeddings of unseen verification artists. This presumes that the per-object mean of the training artists is a good estimate of the artist-independent content component for the verification artists. If the training artists have systematic style correlations, or if object identity and artist identity are not fully separable, subtracting this mean may either fail to remove content or may remove artist-specific identity cues. The paper reports only a single instantiation of the dataset split, so there is no evidence about the stability of mu_o across different training-artist subsets. A simple experiment would estimate mu_o from random halves of the training artists and compare verification performance; the current ablation (Table 4, 'Content Suppress.') is not sufficient to establish the robustness of this central mechanism.","section":"Content Suppression, Eq. (5)"}],"minor_comments":[{"comment":"The table reports five of the nine encoders, yet the text says that all nine encoders improve at every N and cites 'supplementary material' for full results. The main text should either include all nine encoders or clearly state that the five shown are representative and provide a complete table in an appendix that is accessible to the reader.","section":"Table 2"},{"comment":"The 'Fully crossed' column header is ambiguous: the check mark in the IDraw row should be defined explicitly in the caption as 'every artist draws every object' to avoid confusion with the 'crossed' design terminology used in statistics.","section":"Table 1"},{"comment":"The improvement in behavior-feature prediction correlations is small (mean gain +0.016 over 50 features) and the figure shows overlapping error bars. The text claims 'consistent' improvement, but it would be helpful to state the effect size in standard deviations and to show per-feature confidence intervals for the key families (timing, pressure, revision).","section":"Figure 4(a)"},{"comment":"The dataset and code are not released. Given that the paper introduces a new dataset and a new training procedure, releasing these artifacts would be important for reproducibility and for verifying that the device-confounding concern is not material. At minimum, the authors should state a release plan or provide a detailed description of the sensor-signal preprocessing and feature definitions in the main text.","section":"Dataset Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is well executed internally, but the device confound is the central risk: if the sensor supervision and the rendered images share device-specific artifacts, the reported gains may not reflect drawing behavior at all. This is not a circularity issue in the narrow sense, but it is a validity threat to the core claim. I would want to see either a cross-device/leave-device-out experiment (recognizing that the current dataset may not support one) or a substantially more cautious interpretation of the results. The statistical issues with the paired t-tests over non-independent splits are also important and should be fixed, ideally by multi-seed results. If the authors can address the device-confound concern and redo the significance analysis, the paper could be a solid contribution; without those fixes, the behavioral claim is not established. There is also a scope/fit question: the paper is placed in cs.CV but has strong ties to biometrics and security; this seems acceptable, but the editor may want to consider whether the implications for verification in the presence of device diversity are sufficiently addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead IDraw. Short version: it is a genuine contribution with a real dataset, and the empirical result is probably real, but there is an unaddressed confound between artist behavior and device type that should be resolved before publication, and the statistics are a bit sloppier than they look.\n\nWhat is new: the fully crossed 37-artist x 30-object drawing–sensor dataset is a solid resource; as far as I know, nothing like it exists. The training scheme is also a neat application of privileged information: regress differences in sensor-derived behavior features from differences in image embeddings (Eq. 2) and subtract object means at test time. The module ablations are informative, and the fact that the method improves all nine encoders at all six reference-set sizes is strong evidence that the pipeline does something useful. The paper does not oversell the gains; they are modest but consistent.\n\nThe soft spots, in order of importance. First, the device confound: participants drew on their own iPads and Apple Pencils, and the sensor features include pressure and timing; the completed images are rendered on the same devices. So artist and device are perfectly confounded within the training data. The M3 regression could be learning device-specific rendering and response patterns rather than motor behavior. That would still produce better verification among artists who use different devices, but it would undercut the behavioral interpretation and could fail when an artist switches devices. The paper reports no device metadata and no cross-device control. This is the main thing I would want fixed. Second, the statistical protocol: 15 artist-disjoint splits reuse the same 37 artists, so the paired t-tests treat dependent splits as independent; the effective sample size is much smaller than 15. Single seed also means no measure of training variance. Neither is fatal, but both should be addressed. Third, no code or dataset release. For a paper whose main contributions include a new dataset, that is a serious reproducibility gap.\n\nThe circularity analysis is clean: object means come from training artists only, and no target labels enter the construction. So that is not a concern.\n\nWho is this for? People working on digital art attribution, biometrics, and privileged sensor supervision. It deserves serious refereeing; I would accept it with major revision. The authors should release the data and code, add a cross-device or per-device analysis, and redo the significance testing with properly non-overlapping splits. If the device confound turns out to be real, the paper still has value as a verification method, but it needs to be reframed accordingly.\n\nI would bring it to reading group and would probably cite the dataset once it is available. Whether the behavior claim survives contact with a device-control experiment, I genuinely do not know.","headline":"A real dataset and a sensible privileged-supervision pipeline with consistent gains, but the artist-device confound and loose statistics need work before I'd trust the behavioral interpretation.","tokens_in":13023,"tokens_out":3689,"would_cite":true,"duration_ms":33390,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"IDraw verifies digital-drawing authorship from images alone, cutting verification error by up to 40%.","keywords":["drawing authorship verification","artist verification","tablet-pen sensor signals","behavior-guided learning","content suppression","multimodal dataset","digital drawings","learning using privileged information"],"falsifier":"A cross-device experiment: recruit the same artists to draw the same objects on two different tablet models, train IDraw on artists using device A, and verify on artists using device B; if AUC falls to near baseline while same-device verification stays high, device-specific pen artifacts rather than artist behavior are carrying the signal. A quicker check is whether the 50 behavior features for one artist change more across devices than across artists.","tokens_in":12049,"feed_emoji":"✍️","tokens_out":6625,"duration_ms":55853,"temperature":0.7,"pith_summary":"This paper claims that artist-specific drawing behavior—pen pressure, stroke timing, editing habits—leaves recoverable traces in a finished digital drawing, and that a verification system can be trained to read those traces even when the claimed artist's sensor data is never available. The proposed framework, IDraw, trains an image encoder on drawings paired with 14 tablet-pen sensor signals collected from separate artists, teaching the encoder to make embedding differences mirror behavior differences. At verification time only the disputed drawing and trusted reference drawings are used, with an added step that subtracts an estimated object mean to suppress what the drawing depicts. On a newly collected dataset of 1,110 drawings by 37 artists, each drawing all 30 objects, IDraw consistently outperforms standard image-based verification across nine encoder backbones and all six reference-set sizes, reducing verification error by up to 40 percent.","feed_headline":"Sensor-trained IDraw cuts artist-verification error by 40%","feed_subtitle":"Training on pen-pressure and timing signals lets a model verify artists from finished drawings alone, with no sensors at test time.","key_machinery":"The machinery has two load-bearing parts. First, behavior-guided supervision (M3): a linear head $A$ maps the difference $\\Delta \\hat{z}$ between two normalized image embeddings to the difference $\\Delta \\tilde{b}$ between their standardized behavior feature vectors, trained by the mean-squared-error loss $L_{\\text{beh}} = \\frac{1}{K}\\|A\\Delta\\hat{z} - \\Delta\\tilde{b}\\|_2^2$; this is what teaches the image encoder to read behavior from pixels. Second, content suppression (M5): for each object $o$, the mean embedding $\\mu_o$ over training artists is subtracted from every drawing's standardized embedding before $\\ell^2$ normalization, removing the component shared by same-object drawings. A triplet loss (M4) with batch-hard mining pulls same-artist different-object drawings together and pushes same-object different-artist drawings apart. The verification score is the cosine distance between the content-suppressed query representation and the $\\ell^2$-normalized average of content-suppressed reference representations.","core_discovery":"IDraw's central claim is that a completed digital drawing retains enough of the artist's production behavior—pressure, stroke timing, revision patterns—that a model trained with tablet-pen sensor data from other artists can verify authorship from images alone. The key design choice is to supervise the difference between normalized image embeddings of two same-object drawings by different artists with the difference between their standardized 50-dimensional behavior feature vectors, so the embedding geometry is pushed to track behavior rather than object appearance. At verification, the method subtracts an estimated object mean from each embedding and averages the reference drawings to form an artist representation; the cosine distance to the query drawing is the verification score. In the authors' evaluation on 37 artists and 30 objects, this recipe outperforms the cross-entropy baseline on all nine encoders and all six reference-set sizes, with relative equal-error-rate reductions up to 40 percent, and it closes 58.1 percent of the gap between baseline and the sensor-signal upper bound on one representative encoder.","pith_inferences":["A natural extension the paper leaves implicit: resizing, cropping, or JPEG compression may destroy some behavior traces, so testing IDraw on redistributed images would reveal whether the gains survive real online sharing.","Because the object mean is estimated from only 23 training artists, the method could be made more robust by estimating content statistics from a larger artist pool or from unlabeled same-object drawings available at verification time.","If generative models imitate visual style but not pen-pressure and timing statistics, behavior-guided verification could serve as a complementary defense against style imitation; the paper names this as an open question rather than a result."],"forward_implications":["Reference drawings need not depict the same object as the disputed drawing, so verification can work across an artist's varied output.","No sensor data is needed from the claimed artist at verification time; the sensor guidance is baked into the image encoder during training, so the method applies wherever finished images are shared.","The gains appear across nine image encoders spanning different pretraining paradigms, suggesting the behavior signal is broadly available rather than tied to one architecture.","The fully crossed dataset (37 artists, 30 objects, 14 sensor signals) enables controlled study of artist versus object variation that existing drawing datasets cannot support."],"supporting_citations":[{"why":"Establishes that pen pressure, azimuth, and altitude features capture identity in online signature verification, motivating the sensor-feature choices.","marker":"Muramatsu and Matsumoto 2007"},{"why":"Deep online signature verification demonstrates that sensor signals encode writer-specific behavior, the premise IDraw distills into images.","marker":"Tolosana et al. 2021"},{"why":"Defines learning using privileged information, the paradigm under which sensor signals are available at training but not verification.","marker":"Vapnik and Vashist 2009"},{"why":"Provides the cross-entropy artist-classification baseline on fine-art paintings that IDraw is compared against.","marker":"Tan et al. 2016"},{"why":"Supplies the batch-hard triplet mining used by M4 for content-aware metric learning.","marker":"Hermans, Beyer, and Leibe 2017"},{"why":"Shows stroke-level attributes can support attribution but that content can dominate image-based comparisons, the problem IDraw targets.","marker":"Elgammal, Kang, and Den Leeuw 2018"},{"why":"Introduces score-normalization ideas that IDraw adapts when suppressing shared object content.","marker":"Auckenthaler, Carey, and Lloyd-Thomas 2000"},{"why":"The open-source tool used to record the 14 tablet-pen sensor signals in the IDraw dataset.","marker":"Ding 2016"}],"fun_headline_variants":["IDraw verifies artists from drawings alone, cutting error by 40%","Sensor-trained IDraw verifies from finished images, 40% less error","Inferring pen behavior from images improves artist verification by 40%","No sensors needed: IDraw verifies drawings with 40% lower error","IDraw's content suppression slashes artist-verification errors 40%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pen-pressure, timing, and tilt signals collected on each artist's own tablet and digital pen are comparable across devices and reflect drawing behavior, not device-specific artifacts; if that fails, the sensor supervision teaches device fingerprints and the measured gains would not mean what the paper claims.","fun_headline_variants_meta":{"raw":{"variants":["IDraw verifies artists from drawings alone, cutting error by 40%","Sensor-trained IDraw verifies from finished images, 40% less error","Inferring pen behavior from images improves artist verification by 40%","No sensors needed: IDraw verifies drawings with 40% lower error","IDraw's content suppression slashes artist-verification errors 40%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2254,"prompt_tokens":987,"completion_tokens":1267,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":1167}},"tokens_in":603,"tokens_out":1267,"duration_ms":9167,"temperature":1.0,"reasoning_tokens":1167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:04:38.703157+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A cross-device experiment: recruit the same artists to draw the same objects on two different tablet models, train IDraw on artists using device A, and verify on artists using device B; if AUC falls to near baseline while same-device verification stays high, device-specific pen artifacts rather than artist behavior are carrying the signal. A quicker check is whether the 50 behavior features for one artist change more across devices than across artists.","supporting_citations":[],"review_version":3}