{"id":"75e31726-d3c2-4839-a68d-870d1dd49d89","arxiv_id":"2504.17609","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A teacher-model-based curriculum training strategy improves steganographic image quality metrics while sometimes reducing bit decoding accuracy at high capacity.","lead":"This paper trains image steganography models by ordering images from easy to hard, using teacher models to score difficulty, rather than presenting all images at once. The authors report better stego-image quality and lower steganalysis detection on three benchmark datasets, with some loss of decoding accuracy at high embedding rates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Teacher-consistency difficulty score is the load-bearing element but is never validated as intrinsic or transferable; Table 6 shows hard subsets with baseline PSNR comparable to or higher than easy subsets, so the easy-to-hard premise is not established.","rationale":"The reader's weakest-assumption diagnosis matches my own reading. The difficulty score is the only component of STCL that genuinely distinguishes it from random sample ordering; the knee-point rule just decides when to switch. If the score does not transfer across models, the entire method reduces to tuning the order in which fixed subsets are presented, and the reported gains could be explained by subset imbalance, thresholds, or random variation. The paper supplies no validation of this transfer, and Table 6 contains direct counter-evidence for the PSNR metric on two of three datasets. I do not think the paper should be rejected outright; the experiments are extensive, and the mechanism is plausible. But the central claim is conditional on a validation that is absent. The encoding/decoding accuracy and steganalysis inconsistencies in Tables 1 and 8 are additional reasons not to strengthen the claim, but they are secondary. I would keep the reader's CONDITIONAL verdict, with the condition made explicit: demonstrate that the difficulty ordering is stable and predictive for a freshly initialized student, and clarify Eq. (4)'s quantifier over teachers.","tokens_in":15567,"tokens_out":5994,"duration_ms":62265,"concrete_test":"Run a fresh random-sampled student on the full dataset, then compute Spearman correlations between the teacher-consistency difficulty score (Eqs. 2-4) and that student's per-image PSNR and SSIM on the test set. If the correlation is near zero or negative on ImageNet or VOC2012, the teacher-based ordering does not transfer to a new student, so the claimed easy-to-hard curriculum is not the cause of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core of STCL is the difficulty score in Eqs. (2)-(4): three teacher models trained for C1<C2<C3 epochs produce SSIM/PSNR values per sample, and thresholds alpha1, alpha2, mu1, mu2 sort images into Easy/Medium/Hard. The paper never establishes that this ordering is a property of the images rather than of the specific teachers, metrics, or thresholds. Eq. (4) is also under-specified: the condition is written with indices i,j without stating whether Easy/Hard must hold for every teacher or for any teacher, so the assignment rule is not reproducible as written. Table 6 provides internal evidence against the premise: on ImageNet the Hard subset has baseline PSNR 35.389 versus Easy 34.126, and on VOC2012 Hard has PSNR 34.337 versus Easy 32.108, so the 'hard' images are not harder under one of the two headline metrics for a randomly trained student. If the ordering is teacher-specific, the reported gains could come from subset composition, hand-tuned thresholds, or initialization luck rather than from easy-to-hard curriculum. Secondary weaknesses (Table 1 decoding accuracy drops e.g. ImageNet 2bpp 0.99->0.73; Table 8 steganalysis scores sometimes increase) do not rescue the central claim, but the difficulty score is the load-bearing element.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STCL, a curriculum-learning training strategy for deep-learning image steganography. The method first trains several teacher models of the same architecture for different numbers of epochs, then scores each training image by the consistency of the SSIM and PSNR values achieved by those teachers, and finally divides the training set into Easy, Medium, and Hard subsets. The student model is trained in three stages, starting with the easy subset and adding medium and then hard subsets, with each early stage stopped at a so-called knee point. Experiments on ALASKA2, VOC2012, and ImageNet at 1-3 bpp compare STCL with random-order baseline training on PSNR, SSIM, MSSSIM, RMSE, decoding accuracy, and XuNet steganalysis scores, and additionally report generalization experiments on SteganoGAN and FC-DenseNet. The central claim is that STCL improves stego-image quality and security while maintaining decoding accuracy, and that it generalizes across datasets and model frameworks.","tokens_in":15964,"tokens_out":5352,"duration_ms":48907,"significance":"If the claims were fully supported, the paper would make a useful training-time contribution to deep-learning steganography, since curriculum learning has not been systematically explored for this task. The manuscript has concrete strengths: it reports experiments on three public datasets, includes a security evaluation with XuNet, tests generalization to two other architectures, and releases code. However, the current evidence does not establish the headline claims. The proposed difficulty score is not validated as a transferable measure of intrinsic image difficulty, several reported numbers contradict the text's qualitative claims, and there is no statistical support for the comparisons. These issues affect the core contribution, so the significance is currently conditional on substantial revision.","major_comments":[{"comment":"The claim that STCL maintains or improves decoding accuracy is contradicted by multiple cells. For example, ALASKA2 D=2 accuracy drops from 0.99 (Baseline) to 0.92 (STCL); ImageNet D=2 drops from 0.99 to 0.73; and VOC2012 D=3 drops from 0.92 to 0.83. The text in §4.3 states that accuracy is \"slightly higher than the baseline scheme at 1-2bpp steganographic capacity,\" which is not supported by these rows. The authors need to correct the overclaim and either explain the trade-off or restrict the claim to the settings where it holds.","section":"§4.3, Table 1"},{"comment":"Equation (4) is not reproducible as written. The condition for \"Easy\" is \"(Sij(xi,x_i) ≥ α1 and Pij(xi,x_i) ≥ μ1)\", but Sij and Pij are defined as scalar scores for sample xi under teacher Tj, so the notation with two arguments is undefined; more importantly, the formula does not state whether the inequality must hold for all teachers j, for at least one teacher, or on average. Since this rule determines the Easy/Medium/Hard partition that drives the entire method, the ambiguity directly affects the validity of all subsequent experiments.","section":"§3.1, Eq. (4)"},{"comment":"The internal evidence does not support the premise that the Hard subset is actually harder. In Table 6, the NoCL baseline PSNR on the ImageNet Hard subset is 35.389, which is higher than the Easy subset's 34.126; for VOC2012, the Hard subset PSNR is 34.337 versus 32.108 on Easy. The text in §4.3.5 claims that \"the model without the STCL strategy performs poorly on the difficult subset, and each of the hidden writing metrics is lower than the test results on the simple and medium subsets,\" which is directly contradicted by these rows. This undermines the teacher-consistency difficulty score as a valid measure of intrinsic image difficulty.","section":"§3.1 and §4.3.5, Table 6"},{"comment":"The difficulty score is computed with teacher models that share the same architecture and the same SSIM/PSNR-family losses that are later used to evaluate the student, and the thresholds α1, α2, μ1, μ2 are hand-set (with values 0.9, 0.8, 20, 12) with no sensitivity analysis. The paper provides no evidence that the resulting difficulty ordering transfers to a newly initialized student or to a different architecture. Without such validation, the reported improvements could be an artifact of subset composition, threshold choice, or initialization rather than of an easy-to-hard curriculum.","section":"§3.1-§3.2 and §4.2"},{"comment":"No error bars, repeated runs, or significance tests are reported anywhere in the experimental section. Several of the improvements are extremely small (for example, SSIM 0.9934 versus 0.9932 for VOC2012 D=2 in Table 1), and without an estimate of run-to-run variance it is impossible to assess whether STCL improves over random-order training beyond chance. The authors should provide multiple seeds and a statistical comparison for the main tables.","section":"§4.3, Tables 1-6"},{"comment":"There are citation and description inconsistencies in the generalization experiments. \"SteganoGAN [17]\" does not point to reference [17] (which is RoSteALS), and \"FC-DenseNet [32]\" points to a speaker-verification paper rather than a DenseNet-based steganography model. In addition, §3.1 calls the teacher model a \"three-way adversarial model,\" but the base model described in §4.2 is an encoder-decoder trained with SSIM/MSSSIM/RMSE and binary-cross-entropy losses, with no adversarial component; the discrepancy should be resolved.","section":"§4.3.7"}],"minor_comments":[{"comment":"There are two items labeled \"6)\" in Section 4.3: \"Comparison of training convergence nodes\" and \"Security Testing\". The numbering should be corrected.","section":"§4.3"},{"comment":"Equation (1) introduces CN, but the method uses only three teacher models T1, T2, T3, while Figure 1 shows an arbitrary number N of teachers; the relationship between these should be clarified.","section":"§3.1"},{"comment":"The sentence \"the input RGB image in Duan et al.'s study [32] is modified to binary information\" appears to cite the wrong reference, since [32] is a speaker-verification paper and not a study on RGB image steganography.","section":"§4.3.7"},{"comment":"The caption contains the typo \"ASKAKS2\" for ALASKA2.","section":"Figure 2"},{"comment":"The text states that max_iter is 120, but later says the first training stage takes 15-30 epochs to reach the knee point. Please clarify whether 120 refers to epochs or to iterations, and how \"epoch\" and \"iteration\" are used in the protocol.","section":"§4.2"},{"comment":"Grammatical issues such as \"an knee point\" and the long, repetitive introduction should be corrected; these do not affect the technical content but reduce readability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains several internal contradictions between the text and its own tables, particularly around decoding accuracy (Table 1) and the difficulty-order premise (Table 6). These are not mere presentation issues: they bear directly on whether the proposed difficulty score and curriculum schedule work as claimed. I recommend asking the authors to provide raw per-run results, a clarified and reproducible definition of Eq. (4), validation of the difficulty ranking across architectures and initializations, and corrected references and overclaims before reconsidering the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible training recipe for deep image steganography, and the authors do real multi-dataset, multi-architecture experiments. The novelty is narrow—combining a teacher-model consistency difficulty score with a knee-point scheduler—and that combination is genuinely new in this subfield. The quality-metric gains (SSIM, PSNR) are consistent enough across ALASKA2, VOC2012, and ImageNet that the recipe probably does something. But the load-bearing element, the difficulty score, is never validated as a property of the images, and one of their own tables suggests it isn't.\n\nWhat's good: the framing is clear, the related work is reasonable, and the citation pattern is honest—Bengio, Ye, Lee, and Bernat are the right anchors. They test across 1–3 bpp, include steganalysis results (Table 8, directionally supportive), and show generalization to SteganoGAN and FC-DenseNet. The knee-point stopping rule is a sensible efficiency idea, and Table 7 shows it doesn't hurt final performance.\n\nWhere it gets shaky. Equations 2–4 define difficulty by consistency of SSIM/PSNR across three teacher models trained for C1<C2<C3 epochs, but those teacher training lengths are never reported. Eq. 4 is under-specified: it doesn't state whether the Easy/Hard condition must hold for all teachers or any teacher, so the assignment rule isn't reproducible as written. More importantly, nothing shows the ordering transfers to a newly initialized student or to different architectures. The thresholds (alpha1=0.9, alpha2=0.8, mu1=20, mu2=12) are hand-set, with no sensitivity analysis.\n\nThe bigger problem is Table 6, which gives internal evidence against the premise. On ImageNet, the 'Hard' subset has baseline PSNR 35.389 versus 'Easy' 34.126; on VOC2012, Hard has 34.337 versus Easy 32.108. Under PSNR, the 'hard' images are actually easier for a random baseline. SSIM tells the opposite story, so the ordering is metric-dependent—which means the reported gains could reflect subset composition or hand-tuned thresholds rather than a curriculum effect. Then there are the decoding accuracy drops: ImageNet 2 bpp goes from 0.99 to 0.73, VOC2012 3 bpp from 0.92 to 0.83. The abstract claims the strategy maintains decoding accuracy, but the tables show declines at higher capacities. No error bars, no repeated runs, no significance tests, so we can't tell whether any of these differences are noise.\n\nNet: the recipe is plausible and worth knowing about, but the central mechanism is unvalidated. This deserves peer review, not desk rejection, because the concerns are empirically addressable. The referee should demand: (a) a fully specified difficulty rule, (b) reported teacher training lengths, (c) validation that the difficulty ordering transfers across architectures and initializations, and (d) error bars on the headline numbers. If the authors can supply those, it becomes a solid training-recipe paper. For now, treat the quality gains as suggestive, not established.","headline":"A plausible curriculum-learning recipe for image steganography, but the teacher-consistency difficulty score that carries the method is never validated as intrinsic or transferable, and Table 6 actually undercuts the easy-to-hard premise.","tokens_in":16408,"tokens_out":2533,"would_cite":false,"duration_ms":23881,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Easy-to-hard training curricula lift stego image quality and security","keywords":["steganography","curriculum learning","teacher model difficulty scoring","knee point scheduling","image quality metrics","steganalysis","deep learning training strategy"],"falsifier":"Train a student on the same three-stage schedule but in reverse difficulty order (hard first, easy last) on ALASKA2; if the reverse-curriculum student matches or beats STCL on PSNR, SSIM, and XuNet score, the claim that easy-to-hard ordering causes the gains fails. Alternatively, build the difficulty ranking with one architecture and evaluate the curriculum with a structurally different student; a large drop in the reported advantage would show the ordering does not transfer.","tokens_in":15388,"feed_emoji":"🖼️","tokens_out":3319,"duration_ms":28469,"temperature":0.7,"pith_summary":"This paper proposes a curriculum learning training strategy for deep-learning image steganography models. The strategy first scores every training image by how consistently three teacher models of different training maturity preserve stego-image quality in terms of SSIM and PSNR, then splits the data into easy, medium, and hard subsets. Training begins on the easy subset, stops at a knee point where performance growth levels off, and progressively adds the medium and hard subsets until convergence on the full dataset. The authors report that students trained this way beat randomly trained baselines on ALASKA2, VOC2012, and ImageNet, with higher SSIM and PSNR, comparable or better decoding accuracy, and lower steganalysis scores from a XuNet detector.","feed_headline":"Easy-to-hard training sharpens stego image quality","feed_subtitle":"Ordering images by teacher-model difficulty boosts PSNR and SSIM and cuts steganalysis scores on three datasets.","key_machinery":"The central object is the teacher-model difficulty score from Equations 2 through 4: three teacher models with the same architecture as the student are trained for increasing numbers of epochs, and each training image's difficulty is set by the consistency of SSIM and PSNR of the stego images the teachers produce. An image is classified as easy if all teacher scores lie above thresholds $\\alpha_1$ and $\\mu_1$, hard if any lies below $\\alpha_2$ and $\\mu_2$, and medium otherwise. This score carries the argument because it defines the easy-to-hard curriculum, while the knee-point rule sets when to switch stages.","core_discovery":"The central claim is that ordering steganography training data from easy to hard, according to a teacher-model consistency score, produces stego images of higher visual quality and greater resistance to steganalysis than training on the same data in random order, without sacrificing decoding accuracy. The difficulty score treats an image as easy when all three teachers produce similar high SSIM and PSNR stego versions, and hard when any teacher produces a low-quality stego image. The knee-point schedule stops each early stage just as performance saturates, which the authors argue prevents overfitting on small subsets and accelerates convergence. The paper further claims the strategy transfers across three datasets and two alternative steganography network structures.","pith_inferences":["The difficulty ranking produced by SSIM and PSNR consistency is likely to correlate with texture richness, so the same ranking could be reused for cover selection in traditional steganography, not only for curriculum scheduling.","A testable prediction is that the ordering transfers across architecture families; one could check whether a difficulty ordering built from teachers of one architecture helps a student of a different architecture as much as the same-architecture case.","The knee-point heuristic could be made adaptive by monitoring validation loss rather than fixing hand-set thresholds, but that is a modification the paper does not test."],"forward_implications":["If STCL generalizes as claimed, training steganography models on easy-first curricula should improve stego image quality on flat-color and low-texture regions that hard subsets contain.","The knee-point schedule should shorten early training, with stages one and two reaching their stopping point after roughly 15 to 30 epochs on the tested datasets.","The strategy should maintain decoding accuracy at or near the random-training baseline while improving image quality, including at higher embedding capacities of 2 to 3 bits per pixel.","The reported gains should appear across different steganography network architectures, not only the paper's own encoder-decoder design."],"supporting_citations":[{"why":"Supplies the curriculum learning paradigm and the easy-to-hard ordering rationale that STCL adapts.","marker":"[8]"},{"why":"Defines the adversarial steganography framework with encoder, decoder, and steganalysis critics that the teacher and student models instantiate.","marker":"[6]"},{"why":"Provides the SteganoGAN architecture that serves as the base model and one of the generalization test structures.","marker":"[3]"},{"why":"Gives the XuNet steganalysis network used to measure the security improvement via steganalysis scores.","marker":"[40]"},{"why":"Supports the claim that difficult and noisy samples improve model generalization, motivating the final stage of training on hard images.","marker":"[39]"}],"fun_headline_variants":["Easy-to-hard data ordering boosts stego image quality","Teacher-model scores order stego training data for better output","Curriculum learning sharpens stego images via teacher difficulty","Easy-first training for stego nets improves image quality","Knee-point scheduling cuts overfitting in stego training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The teacher-model consistency score is assumed to be a valid, transferable measure of intrinsic image difficulty for a newly initialized student model; if the ordering is an artifact of the teachers' own training trajectories, the curriculum has no principled basis.","fun_headline_variants_meta":{"raw":{"variants":["Easy-to-hard data ordering boosts stego image quality","Teacher-model scores order stego training data for better output","Curriculum learning sharpens stego images via teacher difficulty","Easy-first training for stego nets improves image quality","Knee-point scheduling cuts overfitting in stego training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2679,"prompt_tokens":944,"completion_tokens":1735,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":1657}},"tokens_in":560,"tokens_out":1735,"duration_ms":10494,"temperature":1.0,"reasoning_tokens":1657,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:34:55.801954+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a student on the same three-stage schedule but in reverse difficulty order (hard first, easy last) on ALASKA2; if the reverse-curriculum student matches or beats STCL on PSNR, SSIM, and XuNet score, the claim that easy-to-hard ordering causes the gains fails. Alternatively, build the difficulty ranking with one architecture and evaluate the curriculum with a structurally different student; a large drop in the reported advantage would show the ordering does not transfer.","supporting_citations":[{"cited_title":"Structural design of convolutional neural networks for steganalysis","cited_arxiv_id":null,"evidence_quote":"Gives the XuNet steganalysis network used to measure the security improvement via steganalysis scores."},{"cited_title":"Curriculumnet: Weakly supervised learning from large-scale web images","cited_arxiv_id":null,"evidence_quote":"Supports the claim that difficult and noisy samples improve model generalization, motivating the final stage of training on hard images."}],"review_version":1}