{"id":"65c12f25-6e9c-419f-b0a3-1bd443a8706d","arxiv_id":"2508.00135","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A custom CNN trained on periocular eye images reaches 99% reported gender classification accuracy on CVBL and 96% on a Kaggle eye dataset.","lead":"The paper reports a convolutional neural network that classifies gender from color images of the eye region, reaching 99% accuracy on the CVBL dataset and 96% accuracy on a Kaggle eye dataset. It is relevant to security and surveillance applications that need soft biometric identification when the full face is unavailable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 99% CVBL result hinges on an identity-disjoint train/test split and a documented test set, neither of which appears in the visible text; since reference [34] identifies CVBL as an iris database with multiple images per subject, same-subject leakage could inflate the reported accuracy.","rationale":"The reader's conditional verdict is appropriate. The paper's strongest claim is an empirical accuracy number, and the visible text lacks the dataset and results sections, so the abstract alone cannot establish the evaluation protocol. My concern is not that the results are impossible; near-ceiling periocular gender classification has been reported on small datasets, and the proposed architecture is plausible. The issue is that the single most load-bearing condition, subject-disjoint evaluation, is undocumented and is made more salient by the dataset's own name in reference [34], which indicates a per-subject iris collection. This is an internal-evidence point, not a disagreement with consensus. The concrete test is feasible: a strict identity-disjoint split and re-training with the architecture from Table 2 would settle whether the reported numbers reflect generalization or subject memorization. I therefore keep the reader's CONDITIONAL verdict; if the split is provided and the accuracy holds, the paper should be accepted, but if the split is image-level, the reported numbers must be re-derived.","tokens_in":5392,"tokens_out":3198,"duration_ms":35205,"concrete_test":"Obtain from the authors, or reconstruct from CVBL metadata, subject IDs for every CVBL image and re-run the full pipeline under a strict identity-disjoint split (e.g., 80% of subjects for training, 20% for testing), with no early stopping or hyperparameter adjustment on the test partition, and report the test-set size and per-class accuracy with a confidence interval. If the 99% figure cannot be reproduced under this protocol, the central claim is unsupported; if it is reproduced, the concern is resolved. As a supplementary check, verify that the Kaggle 'Female and Male' split removes duplicate or near-duplicate eye images across partitions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption behind the headline claim is that the CVBL test set was identity-disjoint from training and that CVBL did not influence architecture or hyperparameter selection. The visible text contains no dataset section, no split description, no test-set size, no class balance, and no augmentation policy. Reference [34] identifies CVBL as an 'IRIS Gender Classification Database,' which means it is a per-subject ocular capture set with multiple images per eye. If a random image-level split was used, the same subject can appear in both training and test partitions; periocular images of the same eye share skin texture, eyelid shape, illumination, and possibly eyewear, so the network can memorize subjects rather than learn gender. With a reported 99% accuracy, only a handful of test errors are permitted, so a small or imbalanced test set would make the estimate fragile. The same protocol concern applies to the Kaggle 'Female and Male' dataset, though the harder-to-verify claim is the CVBL number.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a custom CNN for gender classification from color periocular images and reports 99% accuracy on the CVBL dataset and 96% accuracy on the Kaggle (Female and Male) eye dataset, claiming superiority over prior methods. The version of the manuscript provided to me contains the abstract, introduction and related work, a background section on CNNs, a description of the proposed architecture with a parameter table, and references, but it does not include the dataset description, experimental setup, results section, or comparison tables that would support the headline accuracy figures.","tokens_in":5588,"tokens_out":3339,"duration_ms":36286,"significance":"If the reported accuracies are reproducible under a properly designed identity-disjoint evaluation, the paper would provide useful evidence that periocular color images support near-ceiling gender classification on these two datasets, and the explicit architecture and hyperparameter table in Table 2 would aid replication. However, the central claim currently rests on an under-specified evaluation protocol, and the manuscript as presented contains no results section that would allow the reader to verify the 99% and 96% numbers or the claimed comparison with state-of-the-art methods. The paper does not provide code, data, or error bars, so its main contribution is an empirical accuracy claim that cannot yet be independently assessed.","major_comments":[{"comment":"The headline claim of 99% accuracy on the CVBL dataset cannot be evaluated because the manuscript does not state whether the train/test split is identity-disjoint, how many images were used for training and testing, whether classes are balanced, or what preprocessing and augmentation were applied. Reference [34] identifies CVBL as an iris gender classification database with multiple images per subject; if the split was done at the image level rather than the subject level, the same subject could appear in both partitions, and the network could memorize subjects rather than learn gender. The term 'previously unused' also leaves unclear whether CVBL was used for architecture or hyperparameter selection.","section":"Abstract and experimental protocol (missing)"},{"comment":"The abstract and introduction state that the model was evaluated with an extensive range of metrics and compared with state-of-the-art approaches, but the provided manuscript contains no results section, no comparison table, no confusion matrix, no ROC or precision/recall values, and no confidence intervals. Without these elements, the claimed 99% and 96% accuracies and the stated superiority over prior methods are unsupported in the visible text. If this material exists in the full submission, it must be included in the reviewed version; as it stands, the central empirical claim is unverifiable.","section":"Results and comparison (absent from provided text)"},{"comment":"Table 2 reports the optimizer, loss, number of epochs, and batch size, but it omits the learning rate, momentum, weight decay, weight initialization scheme, image resolution, and number of repeated runs. With 200 epochs and no stated repetition strategy, the reported accuracy values have no variance estimate; a single favorable run could produce a 99% figure even if the average performance is lower. The manuscript should report mean and standard deviation over multiple runs with different seeds, or otherwise justify the stability of the headline numbers.","section":"Table 2"},{"comment":"The introduction's claim that the paper provides 'a comprehensive comparison with other state-of-the-art approaches' is not backed by any visible evaluation against the cited prior works, such as the PCANET method [14] or the pretrained-CNN approach [13]. A proper comparison needs matched train/test protocols, identical preprocessing, and preferably the same backbone or at least reported parameter counts and training costs; none of these are presented in the visible text.","section":"Introduction and related work"}],"minor_comments":[{"comment":"Table 2 lists 'Batch Normalization Mini-batch size 64' while the Optimizer row lists 'Batch size 128'; please clarify whether the mini-batch size is 64 or 128 during training.","section":"Table 2"},{"comment":"The phrase 'small number of learnable parameters (7,235,089)' is misleading: 7.2 million parameters is not small in absolute terms, and the paper does not compare this count with the parameter counts of the prior models it discusses.","section":"Abstract"},{"comment":"The provided text jumps from the CNN background section directly to the proposed architecture and then to references, with no visible section for the datasets or experiments; if this is an extraction artifact, the submitted version should be checked for completeness, and if not, the manuscript is structurally incomplete.","section":"General presentation"},{"comment":"Reference formatting is inconsistent, for example [13] and [25] contain stray brackets, and some entries use inconsistent journal-name capitalization; a final proofreading pass is needed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core concern is that the paper's central claim is an empirical accuracy figure, and the version under review provides no way to verify the evaluation protocol: no identity-disjoint split, no test-set size, no class balance, no results tables, and no repeated-run variance. I am not rejecting because the claim is implausible on its face; periocular CNNs have reached high accuracy in prior work. But the missing protocol is load-bearing, not cosmetic. The authors should be asked to supply the full experimental section, clarify how CVBL was used in model selection, and report subject-disjoint splits and confidence intervals. If the full submission already contains these, the review should proceed on that version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The paper is a straightforward CNN application to periocular gender classification, and the empirical headline (99% on CVBL, 96% on a Kaggle dataset) is plausible but unsupported because the experimental protocol is missing. The visible text contains no dataset section, no split description, no test-set size, no error bars, and no code.\n\nWhat's new: a custom CNN with 7.2M parameters and a previously unreported evaluation on the CVBL dataset. The authors do cite the relevant literature—Cimtay, Kuehlkamp, Manyala—and the architecture table is concrete. That is real, if modest, value.\n\nThe soft spots are load-bearing. The 99% claim depends on the CVBL test set being identity-disjoint from training. Reference [34] describes CVBL as an iris gender classification database, which means multiple images per subject. If the split is at the image level, the same eye can appear in both training and test partitions, and the network can memorize skin texture, eyelid shape, illumination, and eyewear rather than gender. With 99% accuracy, only a handful of errors are permitted, so even a small or imbalanced test set would make the estimate fragile. The paper does not say. The same concern applies to the Kaggle set, though that dataset is more familiar. Also, the abstract boasts comparisons, but the results tables are not in the visible text, and there are no repeated runs or confidence intervals.\n\nThe paper does not appear to be formally verified or reproducible at this stage. There is no code, data split, or training protocol beyond optimizer and epochs. The claim is plausible—periocular gender classification has been done at high accuracy before—and a 99% number on a fresh database is not shocking, but it is not evidence without the protocol.\n\nWho it is for: a reader tracking periocular soft-biometric benchmarks might want to watch for the final version, but the current draft should not be relied on. If the authors release code and a clear identity-disjoint split description, then a serious referee would be worth the time. As it stands, I would not send it out; I would ask the authors to supply the missing methods and then revisit.","headline":"A plausible but under-documented CNN result on periocular gender classification; the 99% CVBL claim hinges on an identity-disjoint split that the paper never describes.","tokens_in":6117,"tokens_out":2901,"would_cite":false,"duration_ms":28348,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A purpose-built convolutional neural network classifies gender from color periocular images at 99% accuracy on the CVBL dataset and 96% on a second dataset, with about 7.2 million parameters.","keywords":["gender classification","deep learning","periocular biometrics","convolutional neural network","eye images","soft biometrics","CVBL dataset","image classification"],"falsifier":"Hold out the entire CVBL dataset until every architecture and hyperparameter decision is frozen, then run an identity-disjoint train/test split and measure accuracy on that untouched data; if the 99% figure depended on tuning or on same-person images appearing in both sets, the number would drop.","tokens_in":5195,"feed_emoji":"👁","tokens_out":11154,"duration_ms":88574,"temperature":0.7,"pith_summary":"The area around the eye, including the eyelids, eyebrows, and the skin between them, can reveal a person's gender even when the rest of the face is hidden or altered, and this paper sets out to show that a purpose-built convolutional neural network can perform that reading automatically from color images. The authors report 99% accuracy on the CVBL dataset of such periocular images, which they call previously unused, and 96% on a second dataset named Female and Male, with a model of roughly 7.2 million parameters. If the numbers hold, periocular color images become a practical soft-biometric signal for security, surveillance, and human-machine interaction where full faces are not available. The paper further claims that the proposed architecture outperforms earlier periocular and iris-based gender classifiers across the metrics it examines.","feed_headline":"Custom CNN reads gender from eye-region images at 99% accuracy","feed_subtitle":"The same compact model scores 96% on a second eye dataset with 7.2 million parameters.","key_machinery":"The load-bearing object is the proposed CNN architecture: 3-by-3 convolutional filters with same padding and stride 1, batch normalization, exponential linear unit (ELU) activations, 2-by-2 max pooling, dropout at probability 0.15, one fully-connected output with a sigmoid, and binary cross-entropy loss trained with stochastic gradient descent for 200 epochs at batch size 128. With roughly 7.2 million total parameters, the architecture carries the argument by showing that high accuracy can be achieved without a large pretrained backbone and that the color periocular region supplies enough signal for the task.","core_discovery":"The central claim is that the color periocular region alone carries enough gender-discriminative information for a compact custom CNN to reach near-ceiling accuracy: 99% on CVBL and 96% on Female and Male. The network is trained end-to-end with binary cross-entropy and a sigmoid output, using convolutional blocks with batch normalization, exponential linear unit activations, max pooling, and dropout. The authors interpret these results as showing that the periocular region is a reliable and efficient basis for gender classification, and they compare the model against earlier methods using iris texture, pretrained CNNs, and near-infrared images, reporting better accuracy in their test setup.","pith_inferences":["A direct extension the paper does not run is to vary cosmetics, disguise, and occlusion systematically on the same datasets, even though those factors motivate the study.","The same architecture could be transferred to near-infrared periocular images to test whether the color-based advantage persists in the modality used by several comparison baselines.","Breaking down accuracy by left and right eye, image resolution, and lighting or image-quality subgroups would reveal whether the headline numbers are uniform or driven by a subset of easy images."],"forward_implications":["If the reported accuracies generalize, periocular color images are sufficient for automated gender classification even when full facial images are unavailable.","The compact parameter count implies that a task-specific CNN, rather than a large pretrained network, can serve as an efficient gender classifier in resource-constrained settings.","The favorable comparison with iris-based and earlier periocular methods indicates that color periocular cues carry stronger gender information than iris texture alone.","The CVBL dataset, described as previously unused, would become a natural benchmark for future periocular gender classification research."],"supporting_citations":[{"why":"Supplies the CVBL periocular dataset on which the 99% accuracy is reported.","marker":"[34]"},{"why":"Supplies the Female and Male eye-image dataset on which the 96% accuracy is reported.","marker":"[35]"},{"why":"Provides the near-infrared periocular gender classification baseline of 89.22% that the proposed model is compared against.","marker":"[16]"},{"why":"Establishes that the periocular area rather than iris texture carries the most gender information, the comparison context for the paper's claim.","marker":"[17]"},{"why":"Provides smartphone periocular image results of 90.15% for the right eye and 87.15% for the left eye used as state-of-the-art comparison.","marker":"[18]"},{"why":"Provides earlier CNN-based near-infrared periocular gender classification methods whose performance the paper compares with and extends.","marker":"[19]"},{"why":"Gives the pretrained-CNN eye-image gender classification results that motivate the paper's small task-specific architecture.","marker":"[13]"}],"fun_headline_variants":["Eye-region only: custom CNN hits 99% gender accuracy","99% gender accuracy from eye-area images with compact CNN","Periocular region alone: 99% gender classification via CNN","Eye-region CNN: 99% on one dataset, 96% on another","Gender from eyes: 99% with a 7.2M-param CNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 99% CVBL result rests on the premise that the CVBL images were not used, directly or through model selection, to choose the architecture or hyperparameters, and that the train/test split keeps distinct identities apart.","fun_headline_variants_meta":{"raw":{"variants":["Eye-region only: custom CNN hits 99% gender accuracy","99% gender accuracy from eye-area images with compact CNN","Periocular region alone: 99% gender classification via CNN","Eye-region CNN: 99% on one dataset, 96% on another","Gender from eyes: 99% with a 7.2M-param CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3333,"prompt_tokens":929,"completion_tokens":2404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":2307}},"tokens_in":545,"tokens_out":2404,"duration_ms":16976,"temperature":1.0,"reasoning_tokens":2307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:20:31.213393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out the entire CVBL dataset until every architecture and hyperparameter decision is frozen, then run an identity-disjoint train/test split and measure accuracy on that untouched data; if the 99% figure depended on tuning or on same-person images appearing in both sets, the number would drop.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CVBL periocular dataset on which the 99% accuracy is reported."},{"cited_title":"Female and Male Eyes","cited_arxiv_id":null,"evidence_quote":"Supplies the Female and Male eye-image dataset on which the 96% accuracy is reported."},{"cited_title":"Relevant Features for Gender Classification in NIR Periocular Images","cited_arxiv_id":null,"evidence_quote":"Provides the near-infrared periocular gender classification baseline of 89.22% that the proposed model is compared against."},{"cited_title":"Sex-Classification from Cellphones Periocular Iris Images","cited_arxiv_id":null,"evidence_quote":"Provides smartphone periocular image results of 90.15% for the right eye and 87.15% for the left eye used as state-of-the-art comparison."},{"cited_title":"CNN -Based Gender Classification in near - Infrared Periocular Images","cited_arxiv_id":null,"evidence_quote":"Provides earlier CNN-based near-infrared periocular gender classification methods whose performance the paper compares with and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the pretrained-CNN eye-image gender classification results that motivate the paper's small task-specific architecture."}],"review_version":1}