{"id":"edee82a8-754f-4b18-9973-d1a86d30c83f","arxiv_id":"2504.18348","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A curriculum-based loss scheduler for steganography trains first on embedding quality, then decoding, then steganalysis resistance, but its security gains are inconsistent across datasets.","lead":"This paper proposes a two-stage curriculum-learning loss scheduler (TSCL) for deep-learning image steganography, aiming to balance embedding, recovery, and steganalysis losses during training. The authors report mixed improvements across three datasets, with security sometimes worse than the fixed-weight baseline, so the headline claim that TSCL improves security is not consistently supported.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that TSCL improves security is contradicted by the paper's own Table 1, where steganalysis scores worsen in 5 of 9 configurations, including all ImageNet rows.","rationale":"The reader rejected the paper, and my analysis supports that rejection. My concern is stronger than the reader's stated weakest assumption: the paper does not merely fail to prove the curriculum ordering is optimal; its own Table 1 contradicts the advertised security improvement. Because the steganalysis score is explicitly defined so lower means more secure, the ImageNet rows and two VOC rows show TSCL produces more detectable stego images. The paper's text acknowledges security decrease on ImageNet yet the abstract and conclusion assert security improvement. This undermines the strongest claim regardless of whether the curriculum ordering is optimal. I did not focus on missing DWA/GradNorm baselines, because the security contradiction is sufficient for rejection. The reader's weakest_assumption about the D_Encode > D_Decode > D_steganalysis ordering is related but not the most load-bearing: even if the ordering were optimal, the data do not support the security claim. I agree with rejecting; no changes to the verdict are needed. A re-run of the security metric with the original detector would settle whether the table is reproducible, but as reported the claim fails.","tokens_in":13090,"tokens_out":3458,"duration_ms":31892,"concrete_test":"Obtain or re-run the authors' trained models and the fixed steganalysis detector used for Table 1. On the ImageNet 1bpp test split, compute mean steganalysis scores and detection accuracy/AUC for baseline and TSCL stego images across at least 3 random seeds. If TSCL's mean score or detector AUC is higher than baseline (reproducing 0.414 vs 0.362), the security-improvement claim is falsified; if the score is lower, the table may be erroneous and the claim could stand. The same check on VOC2012 3bpp (0.357→0.449) would confirm the contradiction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's abstract and contribution list claim TSCL improves steganography quality, decoding accuracy, and security simultaneously. The security claim is directly contradicted by the experimental table. In Table 1, steganalysis score is [0,1] and lower is better; TSCL raises the score (worsens security) in VOC2012 1bpp (0.434→0.457), VOC2012 3bpp (0.357→0.449), ImageNet 1bpp (0.362→0.414), ImageNet 2bpp (0.325→0.391), and ImageNet 3bpp (0.413→0.423). Only ALASKA2 1bpp and 3bpp improve; ALASKA2 2bpp worsens. Table 4 also shows VOC2012 security worsens (0.357→0.449). The text even concedes 'security ... slightly decreased on the ImageNet dataset' (§4.3.1), yet the abstract and §5 claim security improvement. Since one of the three headline metrics is not improved by the method in most reported configurations, the central claim as stated is unsupported and internally inconsistent. This is not a matter of missing comparison or hyperparameter detail; the paper's own evidence refutes the claimed all-three improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TSCL, a two-phase loss-weight scheduler for deep-learning image steganography models that combine an embedding network, a decoding network, and a steganalysis network. The first phase, \"curriculum control,\" sequentially up-weights the embedding loss, then the decoding loss, then the steganalysis loss; the second phase, \"loss dynamics control,\" adjusts weights according to the ratio of each loss between successive iterations, multiplied by a priori coefficients that keep the embedding loss dominant. The authors report experiments on ALASKA2, VOC2012, and ImageNet at capacities 1-3 bpp and claim simultaneous improvements in steganographic image quality, decoding accuracy, and security.","tokens_in":13450,"tokens_out":5094,"duration_ms":49446,"significance":"If the central claim were supported, TSCL would be a simple, drop-in loss-balancing scheme for a widely used class of steganography models, and the paper's use of three datasets and several curriculum-schedule variants would provide a reasonable experimental skeleton. However, the paper's own tables contradict the headline claim: security (measured by the steganalysis score, lower is better) worsens in 6 of 9 configurations in Table 1, including all ImageNet rows, and the text itself concedes that security decreased on ImageNet. In addition, key schedule parameters are unspecified and the a priori coefficients in Eq. (12) are hand-set to keep embedding loss dominant, which pre-configures one of the claimed benefits. The contribution is therefore not established as stated.","major_comments":[{"comment":"The abstract and conclusion claim that TSCL \"improves the quality of steganography, decoding accuracy and security,\" but Table 1 shows that the steganalysis score (lower is better) worsens under TSCL in 6 of 9 configurations: ALASKA2 2 bpp (0.402 to 0.442), VOC2012 1 bpp (0.434 to 0.457), VOC2012 3 bpp (0.357 to 0.449), and all three ImageNet rows (0.362 to 0.414, 0.325 to 0.391, and 0.413 to 0.423). Table 4 likewise shows VOC2012 security worsening from 0.357 to 0.449 under the full TSCL scheme. Since one of the three headline metrics is not improved by the method in most reported configurations, the central claim as stated is internally inconsistent and unsupported by the paper's own evidence.","section":"§4.3.1, Table 1; Abstract; §5"},{"comment":"The a priori coefficients D_Encode=1, D_Decode=0.8, D_steganalysis=0.4 are hand-set to keep the embedding loss dominant, and the second-stage weight update in Eq. (13) multiplies each loss-decline ratio by these coefficients. Consequently, the observation that imperceptibility does not degrade is partly guaranteed by construction, while the steganalysis loss is always the least weighted, which makes a security improvement structurally less likely. The paper provides no sensitivity analysis with respect to this ordering, and the ImageNet results suggest the ordering may be dataset-dependent. A concrete ablation varying the relative values and order of D_Encode, D_Decode, and D_steganalysis is needed before the claimed benefits can be attributed to the proposed method.","section":"§3.3, Eq. (12); §4.2"},{"comment":"The curriculum scheduler is not reproducible from the manuscript because the schedule hyperparameters a0, a1, a2, C1, C2, C3, the iteration-step sizes, and the exact α values used in the experiments are never reported. Equations (8)-(10) define families of schedule functions, but no values are given for any dataset or capacity, and the text states only that the maximum number of iterations is 120 without connecting this to the epoch-based curriculum boundaries. Since the curriculum timing is the core mechanism of TSCL, the reported comparisons in Tables 2-4 cannot be reproduced or independently evaluated without these parameters.","section":"§3.2, §4.2"}],"minor_comments":[{"comment":"The dataset description contains a typo: \"the datasets used are three large public datasets, ImageNet, ALASKA2 and ImageNet,\" where VOC2012 is presumably meant in one of the two mentions.","section":"§4.1"},{"comment":"The text cites \"Li et al [17]\" for a new low-frequency wavelet loss and FSIM loss, but reference [17] is Chekatamala et al.; the citation-to-reference mapping is inconsistent.","section":"§2.1, References"},{"comment":"Figure captions and labels alternate between \"TSLS\" and \"TSCL\" (e.g., \"Cover TSLS Baseline\" and \"TSLSOnly_loss\"), which is confusing and should be corrected.","section":"Figures 2-4"},{"comment":"All experimental comparisons report a single run with no error bars, confidence intervals, or significance tests; given the small differences in some metrics (e.g., accuracy 0.99 vs. 1.00), the robustness of the reported improvements is unclear.","section":"§4.3, Tables 1-4"},{"comment":"The exponential schedule function is not properly normalized: for epoch = C2 the argument (e^{(epoch-C1)/(C2-C1)} - 1) evaluates to e-1 rather than 1, so the claimed transition to a2 at epoch = C2 is not smooth; please clarify the intended formula.","section":"§3.2, Eq. (10)"}],"recommendation":"reject","confidential_remarks":"The decisive issue is not a missing comparison or a debatable modeling choice: the paper's own Table 1 and Table 4 contradict the security claim in the abstract and conclusion. Because one of the three headline contributions is refuted by the reported experiments, the paper cannot be recommended for publication in its current form. A future revision would need to either restrict the claims to the metrics that actually improve or substantially rework the method and experimental protocol, including a full specification of the schedule hyperparameters and sensitivity analyses for the a priori coefficients."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is narrow: the authors take a task-level curriculum ordering (embedding first, then decoding, then steganalysis) and bolt on a DWA-style loss-ratio weighting term with hand-set priority coefficients. That particular combination applied to deep image steganography is not in the prior literature, and the paper describes it clearly enough that someone could reimplement it. I want to give credit for that, and for running experiments on three public datasets rather than just one toy set. The tables also report the raw numbers honestly, which is more than many papers in this area do.\n\nThe problem is that the central claim is contradicted by those very numbers. The abstract and conclusion say TSCL improves 'quality, decoding accuracy and security.' Table 1 shows steganalysis scores worsening on ImageNet at all three capacities (0.414 vs 0.362 at 1 bpp, for example), and VOC2012 gets worse at 1 bpp and 3 bpp. Table 4 shows the same for VOC2012. The text even concedes in §4.3.1 that security 'slightly decreased on the ImageNet dataset.' So one of the three headline metrics is not improved in most reported configurations. That is not a missing baseline or a hyperparameter detail; it is the paper's own evidence killing the abstract.\n\nThe other soft spots are real but secondary. The a priori coefficients D=1, D=0.8, D=0.4 are hand-set to keep embedding loss dominant, which pre-loads the imperceptibility result. No comparison against DWA, GradNorm, or other loss-balancing baselines is provided, so even the favorable results can't be attributed to the proposed mechanism. And the curriculum epoch boundaries (C1, C2, C3) and schedule amplitudes (a0, a1, a2) are never given, which makes the experiments unreproducible.\n\nWho is this for? Someone working on loss design for deep steganography might find the two-stage idea worth testing, but only if the claims are fixed and the missing baselines are added. As it stands, the paper is not internally consistent enough to be published. My recommendation: reject, but tell the authors to revise the abstract and conclusion to match their own data, add DWA/GradNorm comparisons, and report the schedule hyperparameters. A corrected version could be a useful niche contribution.","headline":"The method is a sensible, clearly described mash-up of known loss-balancing tricks, but the paper's own tables refute its headline claim that security improves, so it needs major revision before it can be taken seriously.","tokens_in":13968,"tokens_out":1200,"would_cite":false,"duration_ms":14460,"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":"The paper argues that its two-stage curriculum-learning loss scheduler TSCL improves steganography quality, decoding accuracy, and steganalysis resistance on ALASKA2, VOC2012, and ImageNet.","keywords":["image steganography","curriculum learning","loss balancing","multi-task learning","adversarial training","steganalysis","deep learning","loss scheduling"],"falsifier":"Train the same model on ALASKA2 at 3 bpp with the curriculum order reversed—steganalysis first, then decoding, then embedding—keeping every other TSCL setting identical; if the reversed order matches or beats the proposed one on PSNR, decoding accuracy, and steganalysis score, the claimed ordering is not the source of the reported gains.","tokens_in":12866,"feed_emoji":"🖼️","tokens_out":14110,"duration_ms":118226,"temperature":0.7,"pith_summary":"Deep-learning image steganography trains an encoder, a decoder, and a steganalysis detector together, and previous work usually sums the embedding, recovery, and steganalysis losses with fixed weights. This paper argues that fixed weights ignore both the model's changing learning state and the natural priority of the three tasks: the hidden image must first be invisible, second be recoverable, and third resist detection. It proposes TSCL, a two-stage loss scheduler: first, a curriculum raises the weight of one loss at a time in the order embedding, decoding, steganalysis; second, each weight is adjusted by the ratio of that loss between consecutive iterations, with prior coefficients keeping embedding dominant. The paper reports that on ALASKA2, VOC2012, and ImageNet, TSCL improves steganographic image quality and decoding accuracy, and on several settings also improves resistance to steganalysis, relative to fixed-weight training.","feed_headline":"Curriculum loss scheduler lifts stego quality and decoding","feed_subtitle":"TSCL orders the three losses by priority, then reweights each from its loss drop, on three public datasets.","key_machinery":"The load-bearing object is the TSCL scheduler, which wraps the existing total loss $L_{\\text{total}}=L_{\\text{Encode}}+L_{\\text{Decode}}+L_{\\text{steganalysis}}$ with a two-phase weighting scheme. Phase one, the prior curriculum, uses a scheduling function or fixed step rule to raise the weight of the embedding loss first, then the decoding loss, then the steganalysis loss, so the model's focus moves through the three tasks in that order. Phase two, loss dynamics control, measures each task's learning speed at iteration $t$ by the loss-decline ratio $W_k(t)=L_k(t-1)/L_k(t-2)$ and sets the final weight to $\\lambda_k(t)=D_k W_k(t)$, with prior coefficients $D_{\\text{Encode}}=1$, $D_{\\text{Decode}}=0.8$, and $D_{\\text{steganalysis}}=0.4$. This two-stage mechanism converts a fixed-weight steganography trainer into one whose loss emphasis follows both a designed curriculum and the measured difficulty of each task.","core_discovery":"The paper's central claim is that in an encoder-decoder-steganalysis steganography framework, the schedule of loss weights is itself a control variable, and a curriculum order that matches task priority—embedding first, recovery second, resistance to detection last—produces better final models than static weighting. TSCL implements this in two stages. In the curriculum-control stage, weights are raised one task at a time either by a continuous function (sine, linear, or exponential) or by discrete step changes, so the model masters invisibility before moving to decoding accuracy and then to steganalysis resistance. In the loss-dynamics stage, the scheduler computes the per-task loss-decline ratio $W_k(t)=L_k(t-1)/L_k(t-2)$ and sets the effective weight to $\\lambda_k(t)=D_k W_k(t)$, where the prior coefficients satisfy $D_{\\text{Encode}} > D_{\\text{Decode}} > D_{\\text{steganalysis}}$. The paper reports gains in PSNR, SSIM, and decoding accuracy on ALASKA2, VOC2012, and ImageNet, and states the overall result as simultaneous improvement in steganography quality, decoding accuracy, and security.","pith_inferences":["Because TSCL only reweights losses that already exist in an encoder-decoder-steganalysis pipeline, it should transfer to invertible-network steganography models without changing the network, a direction the paper does not test.","The ImageNet security decrease shown in the tables suggests the fixed order $D_{\\text{Encode}} > D_{\\text{Decode}} > D_{\\text{steganalysis}}$ may encode dataset-specific task difficulty; an adaptive variant that learns the order from loss-decline rates is a plausible next step.","The loss-decline ratio $W_k(t)$ is noisy from iteration to iteration, especially late in training; smoothing it over a short window could make the second stage more stable while preserving the mechanism."],"forward_implications":["If the paper's claim is right, any encoder-decoder-steganalysis steganography trainer can adopt TSCL by replacing fixed loss weights with the two-stage scheduler, leaving the network unchanged.","The embedding-first curriculum means early training effort on hiding quality is preserved when later phases shift focus to decoding and security, because previous loss weights are not lowered.","Since the second stage uses only scalar loss values and no gradients, TSCL adds negligible memory and compute overhead, making it a practical drop-in for large-scale training.","The largest decoding-accuracy gains in the reported tables appear at 3 bpp capacity, suggesting TSCL helps most when the decoding task is hardest."],"supporting_citations":[{"why":"Supplies the steganalysis network architecture used in the experiments and the SSIM/MS-SSIM loss family that defines the embedding loss.","marker":"[16]"},{"why":"Defines the three-way adversarial encoder-decoder-steganalysis framework that the paper's loss terms and training setup are built on.","marker":"[11]"},{"why":"Introduces the adversarial game of encoding, decoding, and steganalysis that motivates treating the three losses as separable tasks.","marker":"[10]"},{"why":"Provides the uncertainty-based loss-weighting baseline from multi-task learning that TSCL's dynamic weighting conceptually builds on.","marker":"[18]"},{"why":"Proposes Dynamic Weight Average, the loss-change-ratio method that the second stage of TSCL directly adapts.","marker":"[21]"},{"why":"Supplies the gradient-magnitude loss balancing idea that the paper positions as related dynamic weighting work.","marker":"[20]"}],"fun_headline_variants":["Curriculum loss ordering boosts stego quality and decoding","Loss-drop ratios schedule stego training tasks","Two-stage scheduler orders stego losses by priority","Curriculum scheduler balances stego losses for security"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fixed task ordering in TSCL is right—embedding loss should stay dominant, then decoding, then steganalysis—and that this priority holds across datasets, capacities, and architectures; if the ideal ordering differs, the scheduler's benefits could vanish or reverse.","fun_headline_variants_meta":{"raw":{"variants":["Curriculum loss ordering boosts stego quality and decoding","Loss-drop ratios schedule stego training tasks","Two-stage scheduler orders stego losses by priority","Curriculum scheduler balances stego losses for security"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00049,"raw_usage":{"total_tokens":2459,"prompt_tokens":1043,"completion_tokens":1416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":1359}},"tokens_in":659,"tokens_out":1416,"duration_ms":11126,"temperature":1.0,"reasoning_tokens":1359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:17:55.521575+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same model on ALASKA2 at 3 bpp with the curriculum order reversed—steganalysis first, then decoding, then embedding—keeping every other TSCL setting identical; if the reversed order matches or beats the proposed one on PSNR, decoding accuracy, and steganalysis score, the claimed ordering is not the source of the reported gains.","supporting_citations":[{"cited_title":"Invisible steganography via generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"Supplies the steganalysis network architecture used in the experiments and the SSIM/MS-SSIM loss family that defines the embedding loss."},{"cited_title":"Generating steganographic images via adversarial training","cited_arxiv_id":null,"evidence_quote":"Introduces the adversarial game of encoding, decoding, and steganalysis that motivates treating the three losses as separable tasks."},{"cited_title":"Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-magnitude loss balancing idea that the paper positions as related dynamic weighting work."}],"review_version":1}