{"id":"ff33787a-43ba-4bc1-93fe-ce4d290b03c0","arxiv_id":"2504.16364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A progressive multi-scale convolutional block with growing dilation rates, placed in an encoder-decoder-critic steganography network, is claimed to improve image quality metrics; no steganalysis scores are reported.","lead":"CLPSTNet is a steganography network that hides secret data in images using a new block, combining Inception-style branches with progressively larger dilated convolutions. It reports higher PSNR and SSIM than several prior models on three public datasets, but the claimed low steganalysis detectability is not backed by reported measurements.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Steganalysis-resistance is asserted but never measured; no experiment reports detection accuracy or steganalysis scores for CLPSTNet or baselines, despite the abstract and conclusion claiming low scores.","rationale":"The strongest claim in the abstract bundles visual quality, decoding accuracy, and low steganalysis scores. The visual quality part is at least tabulated, and the decoding accuracy part is ambiguous but reported. The steganalysis-resistance part, however, is the distinctive security promise of the paper and is entirely absent from the experimental section. Section 4.3 has no steganalysis experiment; Table 8, the comparison with SteganoGAN and HCISNet, contains no security metric. The Critic is a training component, not an external validator. This is not a disagreement with consensus; it is missing evidence for a stated result. I therefore identify it as the most load-bearing concern. The reader's weakest_assumption focused on the undefined \"Accuracy\" metric; that concern is real and remains, but the missing steganalysis measurement is the broader gap because it leaves the abstract's central security claim unsupported. Since the reader already returned CONDITIONAL, the additional concern does not change the verdict; it reinforces the need for major revisions.","tokens_in":14662,"tokens_out":4701,"duration_ms":44348,"concrete_test":"Using the released code, generate stego images on ALASKA2 at 1, 3, and 6 bpp for CLPSTNet and the Table 8 baselines (e.g., SteganoGAN, HCISNet). Train an independent steganalyzer (SRNet or XuNet) on the corresponding cover/stego pairs and report detection accuracy/AUC, plus the steganalysis score distributions, for each model under an identical protocol. If CLPSTNet does not achieve detection accuracy near chance or does not beat the baselines, the \"low steganalysis scores\" claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is the security/undetectability component of the central claim. The abstract states that stego images generated by CLPSTNet \"have low steganalysis scores\", and Section 5 concludes that the model \"realizes the invisibility and undetectability of information\". Yet no quantitative steganalysis evaluation appears anywhere. Tables 3-8 report PSNR, SSIM, MSSSIM, RMSE, and an undefined \"Accuracy\", but never a detection accuracy, ROC/AUC, or steganalysis score for CLPSTNet or any baseline against an external steganalyzer. The Critic in Section 3.5 is a training-time adversary (with loss Lstehsis in Eq. 5), not an independent measurement; its outputs are never reported. The ALASKA2 dataset is itself a steganalysis benchmark, but only cover images are used, with no stego-detection test. Because the \"low steganalysis scores\" claim is explicitly part of the abstract's central claim and is never measured, the most important promised result is currently unsupported. A secondary compounding issue: \"Accuracy\" is undefined and, with BCE in Eq. (9), appears to be per-pixel bit accuracy; at 1 bpp, 0.98 leaves about 2% of hidden bits wrong, and exact payload recovery is not demonstrated. Both gaps need closure before the claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CLPSTNet, a convolutional steganography model built from progressive multi-scale convolution blocks (PMCBs) that combine Inception-style branches with dilated convolutions and dense connectivity. The model is trained with an encoder-decoder-critic architecture using a composite loss of embedding quality (MSE, SSIM, MSSSIM), binary cross-entropy for message recovery, and an adversarial steganalysis loss. Experiments on ALASKA2, VOC2012, and ImageNet report PSNR, SSIM, MSSSIM, RMSE, and an 'Accuracy' column for capacities of 1-6 bpp, plus ablations and comparisons with ResNet, SteganoGAN, HCISNet, DenseASPP, and attention-based baselines. The abstract and conclusion claim that the generated stego images have low steganalysis scores, in addition to high visual quality and decoding accuracy.","tokens_in":14929,"tokens_out":2511,"duration_ms":25326,"significance":"If fully supported, the proposed progressive multi-scale architecture could be a useful contribution to deep image steganography, particularly for improving the visual quality of stego images at higher embedding capacities. The paper provides extensive tables and ablations showing large PSNR/SSIM improvements over several baselines, and it ships a public code link. However, the most distinctive claimed advantage—steganalysis resistance—is never measured, and the 'Accuracy' metric is undefined and appears to weaken at high capacities. The manuscript therefore currently establishes an architectural claim about image quality much more convincingly than it establishes the security claim that appears in the abstract.","major_comments":[{"comment":"The abstract and conclusion assert that CLPSTNet generates stego images with 'low steganalysis scores' and achieves 'undetectability of information', yet no steganalysis evaluation appears anywhere in the paper. The Critic described in Section 3.5 is a training-time adversary whose loss is part of the objective; its outputs are never reported, and it is not an independent steganalysis detector. Tables 3-8 contain only image-quality metrics and an undefined Accuracy column. To support the central security claim, the authors must add an experiment with an external steganalysis detector (e.g., detection accuracy, AUC, or false-positive rate on ALASKA2 or another benchmark) comparing CLPSTNet against at least the main baselines; alternatively, the claim must be removed or substantially weakened.","section":"Abstract and Section 5"},{"comment":"The metric labeled 'Accuracy' is never defined. Since the recovery loss in Eq. (9) is binary cross-entropy, Accuracy appears to be per-pixel bit accuracy. Under that interpretation, Table 3 shows roughly 2% of hidden bits are wrong at 1 bpp on ALASKA2, and at 3-6 bpp the accuracy falls to 0.60-0.67, meaning a large fraction of the payload is not recovered bit-exactly. This contradicts the paper's language about 'high decoding accuracy'. The authors must define the metric precisely and, for a steganography claim, report exact-recovery rates or bit-error rates per capacity; otherwise the decoding claim is not interpretable.","section":"Section 3.6 and Tables 3-8"},{"comment":"The baseline numbers are internally inconsistent. In Table 8, the 'Conv' row reports SSIM=0.98351, MSSSIM=0.99771, PSNR=33.788, RMSE=0.020, Accuracy=0.99, which are exactly the values reported for 'ProgressiveNet' in Table 4, not for the conv-only baseline (Table 4 lists Conv at 1 bpp as SSIM=0.99027, MSSSIM=0.99806, PSNR=34.260, Accuracy=0.74). The same duplicated values appear for ProgressiveNet and Conv in Table 8. This inconsistency undermines the comparison experiment, since the baseline identities are not trustworthy as printed. Please correct the tables and re-run or re-report the comparison.","section":"Table 8 and Table 4"}],"minor_comments":[{"comment":"Equation (5) uses coefficients 'a' and 'b', while the text refers to 'α,β'; please make the notation consistent in both places.","section":"Section 3.6 and Section 4.2"},{"comment":"The sentence 'Critic chooses the XuNet-based steganalysis auxiliary network proposed by Zhang et al [5]' is duplicated verbatim in the same paragraph; one occurrence should be removed.","section":"Section 3.3"},{"comment":"The text says 'test results are shown in Figure 3' but later refers to 'Figure 11'; the figure numbering should be corrected.","section":"Section 4.1 and Figure 3"},{"comment":"In Table 3, the ALASKA2 row for D=4 has Accuracy 0.66 while D=5 shows 0.67; such non-monotone values are not explained. A brief comment on variance or training details would help readers assess the reliability of the reported numbers.","section":"Tables 3 and 5"},{"comment":"The paper states that the encoding loss coefficients are in the ratio 0.5:0.5:0.3, but Eq. (6) suggests λ1, λ2, λ3 are weights; please clarify whether these are normalized weights or absolute coefficients.","section":"Section 4.2"},{"comment":"The GitHub link is provided but no license or repository snapshot is cited; consider adding a versioned release or stating the exact commit used for the experiments to support reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper's core architectural contribution may have merit, but the current version's abstract and conclusions overstate the evidence: steganalysis resistance is asserted without a single measurement, and the decoding metric is undefined yet central to the evaluation. The table inconsistency in Table 8 also needs correction. These are fixable with additional experiments and careful re-reporting, so I recommend major revision rather than rejection. I would also encourage the editor to check whether the released code reproduces the reported numbers before final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a modest architecture contribution, and its central security claim is unsupported as written. The useful part is the PMCB module: Inception-style branches plus two dilated convolutions, with the dilation rates scheduled [3,6,12,18] across the encoder. That is genuinely new in this combination, though each ingredient is known. The paper also does a real service by reporting extensive ablations on ALASKA2, VOC2012, and ImageNet, including comparisons against SteganoGAN, HCISNet, DenseASPP, and attention variants. The visual-quality numbers look plausible: PSNR above 50 dB at 1 bpp, SSIM close to 1, and the ablation tables consistently show the full model beating stripped-down versions. The citation pattern is fine, including the one self-citation, which is background only.\n\nThe soft spots are serious. First, the abstract and conclusion claim low steganalysis scores, but no steganalysis experiment appears anywhere. The Critic is a training-time adversary, not an independent detector, and its outputs are never reported. ALASKA2 is a steganalysis benchmark, but only the cover images are used. The undetectability claim is load-bearing and currently has zero measured support. Second, the 'Accuracy' metric is never defined, and with BCE in Eq. (9) it is almost certainly per-pixel bit accuracy. At 1 bpp, 0.98 still leaves 2% of hidden bits wrong, and at 3+ bpp accuracy falls to 0.6-0.7. For steganography, exact recovery is usually the requirement, so this metric is too weak and the capacity claim collapses above 2 bpp. Third, the tables have internal inconsistencies: nonmonotonic accuracy across capacities, and identical Conv and ProgressiveNet rows in Table 8, which looks like a copy-paste error. These need correction before the numbers can be trusted. The 'curriculum learning' framing is also mostly a label: there is no curriculum schedule, just a static dilation progression. That is a minor overclaim, but worth fixing.\n\nWho is this for? Practitioners who want a concrete incremental encoder architecture and a baseline for comparison in deep image steganography. It is not a breakthrough, and the security claim should not be cited as established. But the architecture is described clearly, the image-quality evidence is substantial, and the ablations are useful. My recommendation: do not desk reject. Send it to peer review with a strong request for major revision, requiring a real steganalysis evaluation against an external detector, a defined recovery metric, and corrected tables. If those are added, this could be a solid incremental contribution.","headline":"Incremental architecture paper with plausible image-quality tables, but the headline steganalysis-resistance claim is never measured; the security component is asserted, not demonstrated.","tokens_in":15495,"tokens_out":2769,"would_cite":false,"duration_ms":28496,"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":"This paper claims that a progressive multi-scale convolutional block, whose kernel sizes and dilation rates grow stage by stage, lets an image steganography network hide secret bits with higher visual quality and stronger steganalysis…","keywords":["image steganography","CLPSTNet","progressive multi-scale convolution","curriculum learning","dilated convolution","Inception module","steganalysis resistance","generative adversarial network"],"falsifier":"Run CLPSTNet on the ALASKA2 test set at each capacity, extract the recovered 128x128xD bits, and count how many differ from the original secret bits; the count will show whether the reported Accuracy scores certify exact recovery or instead allow a percentage of wrong decoded bits.","tokens_in":14418,"feed_emoji":"🖼️","tokens_out":11705,"duration_ms":95193,"temperature":0.7,"pith_summary":"The paper is trying to establish that a steganography network can hide secret bits inside natural images without visibly degrading them and without leaving traces that a steganalysis network can detect. The proposed CLPSTNet builds both its encoder and decoder from PMCB, a five-branch convolution module that pairs Inception-style multi-kernel pathways with dilated convolutions whose kernel and dilation rate grow progressively through the network. On ALASKA2, VOC2012, and ImageNet, at capacities from 1 to 6 bits per pixel, the authors report PSNR around 50 dB at 1 bpp, SSIM above 0.999, and decoding accuracy of 0.94 to 0.98 at 1 bpp, along with low steganalysis scores for the generated images. If these claims hold, they would mean that curriculum-like progressive receptive-field growth is a practical way to improve both visual imperceptibility and resistance to detection in high-capacity image steganography.","feed_headline":"Progressive multi-scale block lifts stego quality, resists detection","feed_subtitle":"CLPSTNet grows its receptive field stage by stage, lifting PSNR/SSIM and steganalysis resistance on three large datasets.","key_machinery":"The load-bearing object is PMCB, the Progressive Multi-scale Convolution Block: five parallel branches, with 1x1, 3x3, and 5x5 convolutions plus two 3x3 dilated convolutions, each large-kernel branch preceded by a 1x1 bottleneck, and all branch outputs concatenated at the end. In CLPSTNet the dilation rates are scheduled progressively, growing from (3,6) to (6,12) to (12,18) through the encoder and decoder, so the effective receptive field expands layer by layer. Dense connections carry shallow edge and detail features forward, and a XuNet-style critic provides the adversarial loss that pushes generated stego images toward low steganalysis scores.","core_discovery":"The central claim is architectural: progressively increasing the receptive field of the embedding and recovery networks, in the same spirit as curriculum learning, is what lets CLPSTNet outperform simpler steganography networks. Each PMCB has five parallel branches, namely 1x1, 3x3, 5x5, and two dilated 3x3 convolutions, and the dilation rates grow from (3,6) to (6,12) to (12,18) across the network, so the shallow layers capture fine local details while deeper layers perceive larger image regions. The authors report that this progressive schedule beats fixed dilation rates, that PMCB plus dense connectivity beats either module alone, and that CLPSTNet surpasses SteganoGAN, HCISNet, and several attention-based baselines on SSIM, MSSSIM, PSNR, and RMSE at 1 bpp on ALASKA2 while keeping decoding accuracy comparable.","pith_inferences":["In my reading, the 'curriculum' in CLPSTNet is structural rather than data-level: training samples are never sorted by difficulty, so the paper is effectively proposing that curriculum-like architecture can substitute for curriculum scheduling; a natural test is whether adding data-level curriculum on top of the progressive dilation schedule adds further gains.","Because the reported decoding accuracy is an average per-pixel score, exact message recovery is not demonstrated; readers who need bit-perfect hidden communication should treat the capacity figures as upper bounds and consider wrapping an error-correcting code around the hidden payload.","The steganalysis-resistance claim is tied to the XuNet-style critic used during training; applying a stronger or unseen steganalysis detector to the released model's outputs would show whether the low scores transfer beyond the training adversary."],"forward_implications":["At 1 bpp on all three datasets, CLPSTNet reports PSNR above 49 dB and SSIM above 0.999, which the paper reads as visual near-identity between cover and stego images.","Decoding accuracy at 1 bpp is 0.98 on ALASKA2, 0.94 on VOC2012, and 0.95 on ImageNet, but it drops as capacity grows to roughly 0.60 to 0.87 at 3 to 6 bpp, so usable capacity depends on the acceptable error rate.","Ablation experiments in the paper show that the PMCB module and the dense connectivity module each contribute to image quality, and that the progressive dilation schedule outperforms fixed dilation rates.","In the reported comparison on ALASKA2 at 1 bpp, CLPSTNet beats SteganoGAN, HCISNet, DenseASPP, FC-DenseNet, SENet, ECANet, and CBAM on SSIM, MSSSIM, and PSNR, which is the paper's evidence for its main claim."],"supporting_citations":[{"why":"Supplies the tripartite encoder-decoder-critic scheme and the SteganoGAN baseline that CLPSTNet compares against.","marker":"[4]"},{"why":"Provides the dilation-plus-dense-connectivity multi-scale steganography baseline (HCISNet/HidingGAN) that motivates PMCB.","marker":"[8]"},{"why":"Introduces curriculum learning, the training principle the progressive dilation schedule is meant to embody.","marker":"[23]"},{"why":"Source of progressive growing in GANs, which inspires the gradual expansion of receptive field in CLPSTNet.","marker":"[9]"},{"why":"Defines the Inception multi-branch convolution structure that PMCB extends with dilated convolution branches.","marker":"[33]"},{"why":"Gives the XuNet-based steganalysis network used as the Critic in CLPSTNet.","marker":"[5]"},{"why":"Provides the FC-DenseNet baseline representing dense-connectivity image steganography.","marker":"[6]"},{"why":"Provides the ResNet baseline used in the comparison experiments.","marker":"[34]"}],"fun_headline_variants":["Progressive receptive field growth boosts stego invisibility","Curriculum-style dilation sharpens stego quality and security","Five-branch progressive blocks improve stego invisibility","Growing receptive field stepwise lifts stego performance","CLPSTNet: multi-scale progressive modules resist detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported decoding success relies on an undefined per-pixel bit-accuracy score; if the intended use is exact recovery of the secret message, a reported 0.98 at 1 bpp does not by itself show that the message comes out intact.","fun_headline_variants_meta":{"raw":{"variants":["Progressive receptive field growth boosts stego invisibility","Curriculum-style dilation sharpens stego quality and security","Five-branch progressive blocks improve stego invisibility","Growing receptive field stepwise lifts stego performance","CLPSTNet: multi-scale progressive modules resist detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000155,"raw_usage":{"total_tokens":1259,"prompt_tokens":1032,"completion_tokens":227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":151}},"tokens_in":648,"tokens_out":227,"duration_ms":2644,"temperature":1.0,"reasoning_tokens":151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:05:07.568525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CLPSTNet on the ALASKA2 test set at each capacity, extract the recovered 128x128xD bits, and count how many differ from the original secret bits; the count will show whether the reported Accuracy scores certify exact recovery or instead allow a percentage of wrong decoded bits.","supporting_citations":[{"cited_title":"Hidinggan: High capacity information hiding with generative adversarial network","cited_arxiv_id":null,"evidence_quote":"Provides the dilation-plus-dense-connectivity multi-scale steganography baseline (HCISNet/HidingGAN) that motivates PMCB."},{"cited_title":"High-capacity image steganography based on improved fc-densenet","cited_arxiv_id":null,"evidence_quote":"Provides the FC-DenseNet baseline representing dense-connectivity image steganography."}],"review_version":1}