{"id":"a2391ac0-1446-452e-a2fa-1a4e78fe7986","arxiv_id":"2411.13819","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A boundary-preserving overflow preprocessing method combined with best-saving adaptive RS coding improves anti-steganalysis performance in upward robust JPEG steganography.","lead":"The paper proposes a JPEG steganography method that preprocesses spatial blocks to remove overflow while preserving block boundaries, and adapts error-correction strength per image to improve robustness against social-network recompression. The result claims to keep embedded messages recoverable while making them harder to detect by steganalysis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In-sample tuning of O1/O2 on the evaluation set may explain the reported security gains; a held-out validation is required.","rationale":"Agreement with the reader is partial: the reader identified the boundary-security premise from Wang et al. as the weakest assumption, but the paper includes a direct preprocessing ablation (Fig. 11) that tests the empirical effect independent of that premise. The more load-bearing threat is in-sample selection of the two new thresholds O1 and O2. The paper's own Section 4.2 chooses these thresholds from the same 2,000 images used in every subsequent figure, so the headline security gain may be a selection artifact rather than a property of the method. The check I propose is a simple disjoint-train/holdout split, which would settle the issue. Secondary concerns (the Algorithm 1 loop condition uses 'or' where 'and' is presumably intended, and adaptive RS requires Q_channel to be known at embed time) do not outweigh this primary concern.","tokens_in":13477,"tokens_out":10284,"duration_ms":102701,"concrete_test":"Split the evaluation protocol into a tuning set and a disjoint holdout: e.g., take 1,000 BOSSbase images to select O1/O2 and any RS search settings, and keep another 1,000 images untouched for final comparison. Re-run the full pipeline to regenerate Figs. 8-12 on the holdout, reporting average detection error and extraction error rates with standard errors. If the proposed method no longer consistently outperforms ROAST-ST and GMAS at the tested payloads while meeting the 0.0001 robustness budget, the central claim fails; if it does, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical, so the evidence must be out-of-sample. Section 4.1 fixes a single 2,000-image BOSSbase subset for all experiments. Section 4.2 then sweeps O1 and O2 on that same subset (Fig. 7) and fixes O1=0, O2=18. Sections 4.3-4.6 report the security, robustness, and ablation comparisons on the same 2,000 images. Consequently, the reported advantage over ROAST-ST and GMAS at payloads up to 0.5 bpnzac is generated with thresholds selected to perform well on the evaluation data. No validation split, cross-validation, or independent replication is described. If the advantage does not persist when O1/O2 are chosen on a disjoint tuning set, the central security-robustness tradeoff claim is not supported. This is the most load-bearing weak point; it is addressable, so the appropriate disposition is conditional.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a robust steganographic method for upward JPEG recompression scenarios. The main contribution is a preprocessing step that selectively removes spatial overflow in 8×8 JPEG blocks while deliberately limiting modifications to block boundaries, based on the observation that boundaries are more prone to overflow and more sensitive for steganalysis. The second contribution is an adaptive Reed-Solomon error-correction scheme that iteratively searches for a short code length that still meets a robustness threshold. The method is evaluated on a 2,000-image subset of BOSSbase v1.01 (QF 65 covers, Q_channel 85 and 95), comparing anti-steganalysis performance (CCPEV, DCTR, SRNet), robustness, and image quality against ROAST-OS, ROAST-ST, GMAS, and Adaptive-GMAS. The reported results show improved anti-steganalysis performance over baselines at payloads up to 0.5 bpnzac while keeping extraction error below 0.0001 at Q_channel 85 and 95.","tokens_in":13667,"tokens_out":5408,"duration_ms":46807,"significance":"If the reported results are reliable, the paper offers a practical improvement in the security-robustness-capacity tradeoff for robust steganography, a problem of genuine relevance for covert communication over social networks. The use of standard, publicly available steganalysis detectors (CCPEV, DCTR, SRNet) and a public dataset (BOSSbase) are strengths, as is the inclusion of ablation experiments for both the preprocessing and the adaptive error-correction components. However, the paper does not release code or data, and the experimental protocol has load-bearing weaknesses—in-sample threshold selection and missing statistical analysis—that must be addressed before the central claim can be accepted. The core idea (boundary-aware overflow removal and adaptive RS coding) is plausible and worth further study, but the current evidence is not yet out-of-sample.","major_comments":[{"comment":"The parameters O1 and O2 are selected on the same 2,000-image subset that is then used for all subsequent security, robustness, and ablation evaluations in Sections 4.3–4.6. The reported advantage over ROAST-ST and GMAS is therefore based on thresholds chosen to perform well on the evaluation set. A held-out validation set or nested cross-validation must be used to support the central empirical claim; otherwise the improvement may be an artifact of in-sample tuning.","section":"Section 4.2, Fig. 7"},{"comment":"The while condition in Algorithm 1 is 'R_error > threshold or k* >= 7', which contradicts the accompanying text ('This process continues until k* equals 7 or the error rate reaches a preset threshold') and the intended early stopping. As written, the loop will always continue down to k*=7 for every image, even when the error threshold is already met, and the output stego image corresponds to the last iteration, not to best_k*. This is a load-bearing algorithmic specification error; the pseudocode must be corrected and the actual implementation clarified.","section":"Section 3.3, Algorithm 1"},{"comment":"The steganalysis evaluation does not describe the train/test split, number of folds, or any variance estimates. Reported differences such as 'over 2% increase' (Section 4.3) and 'more than 10%' (Section 4.6) are presented without confidence intervals or significance tests. Given the small dataset (2,000 images), the empirical security claim needs error bars or a significance test to distinguish actual improvement from sampling noise.","section":"Section 4.1 and 4.3"},{"comment":"The adaptive error-correction loop requires the embedder to know the exact Q_channel and to simulate recompression at embedding time (line 7). This is a matching-robust requirement, not an upward-robust one as claimed in the contributions. The manuscript should clarify whether Q_channel is assumed to be known to the sender and how the method would behave if the actual channel quality differs from the assumed value.","section":"Section 3.3, Algorithm 1"}],"minor_comments":[{"comment":"There are several typos and grammatical issues: 'develpment' (Introduction), 'cofficients' (Introduction), 'thershold' (Table 2), 'Over an 2% increase' (Section 4.3), and inconsistent spacing such as 'partI' and 'partB'.","section":"Throughout"},{"comment":"The claim that 'spatial block boundaries are more prone to overflow' is presented as a statistical finding (Fig. 4a), but no quantitative distribution or statistical test is reported. Adding a small table or histogram with the actual overflow counts would strengthen this load-bearing observation.","section":"Section 2.4 and 3.2"},{"comment":"The sentence 'Except for Adaptive-GMAS, the stego images generated by the proposed method exhibit significantly higher image quality' is ambiguous: at payload 0.2, Adaptive-GMAS reports higher PSNR than the proposed method. Please rephrase to clarify that the comparison excludes Adaptive-GMAS where it cannot run.","section":"Section 4.3, Table 3"},{"comment":"The statement 'we conducted multiple independent repeat experiments' is not supported by any detail. Please specify what was repeated (e.g., random message seeds, classifier initializations) and report the resulting variance.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the proposed ideas (boundary-aware overflow removal and adaptive RS coding) are reasonable. However, the in-sample tuning of O1/O2 on the evaluation set and the algorithmic inconsistency in Algorithm 1 are significant issues that require a revision. I would also encourage the editor to require the authors to share code or at least a detailed experimental protocol, as the lack of reproducibility is a concern for an empirical paper in steganography. The topic fits the journal, and a careful revision could make the contribution acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhaoxia's group's new arXiv paper is a reasonable incremental step on robust steganography for social-network channels. The genuinely new pieces are the observation that 8x8 spatial block boundaries overflow more than interiors, and a preprocessing rule that skips boundary de-overflow when a block's overflow count is below a threshold (O2) and only treats interiors when their overflow count exceeds O1. Combined with an adaptive RS scheme that saves the best code seen during iteration, the method reports better anti-steganalysis than ROAST-ST at payloads up to 0.5 bpnzac while holding extraction error below 1e-4.\n\nThe experimental work is honestly described: BOSSbase subset, three steganalysis detectors, comparison against four baselines including the two most relevant. The adaptive RS tweak is modest but does what it claims, and the ablation shows it helps.\n\nThe soft spots are the usual ones for this line of work plus one that is more serious. No code or data, no error bars or significance tests. The bigger issue: O1 and O2 are tuned on the same 2,000-image subset used for the security evaluation. The sweep in Fig. 7 and the final choice (O1=0, O2=18) are made with the evaluation set, so the reported advantage over ROAST-ST is partly a product of that fit. This is not a fatal flaw -- it is fixable by adding a held-out validation split or a small tuning set -- but as published, the central security claim is weaker than it looks. The robustness comparison is less affected because O2 mainly moves the error rate floor, but the security numbers are the main selling point.\n\nAlso worth noting: the method assumes the sender knows the channel quality Q_channel, which is realistic for some OSN use cases but not all. The boundary-overflow statistical observation is grounded in a 10,000-image study and is the kind of thing that is plausible and testable by others.\n\nWho should read this: anyone working on robust steganography for JPEG recompression, especially the upward-robust branch. It is a useful data point and a reasonable technique to compare against. I would not cite it as established fact, but I would engage with it.\n\nRecommendation: a serious editor should send this to peer review. The issues are addressable and the community would benefit from the discussion. My verdict would be conditional: require the authors to demonstrate that the O1/O2 choice survives on a disjoint validation set, and ideally release code to make the numbers reproducible.","headline":"A plausible incremental improvement to robust JPEG steganography, but the security claim rests on thresholds chosen on the evaluation set.","tokens_in":14181,"tokens_out":2638,"would_cite":false,"duration_ms":24063,"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":"A preprocessing scheme that removes spatial overflow mostly inside 8x8 blocks while sparing their boundaries, combined with adaptive Reed-Solomon error correction, makes robust JPEG steganography harder to detect than ROAST-ST and GMAS at…","keywords":["robust steganography","overflow alleviation","adaptive error correction","dither modulation","JPEG recompression","boundary preservation","anti-steganalysis","social networks"],"falsifier":"Run the same pipeline twice on the same covers, once with the O2 boundary-stopping rule (the proposed method) and once with unconditional boundary de-overflowing (equivalent to ROAST-ST preprocessing), while keeping the adaptive RS loop identical, and compare the reported steganalysis detectors at a fixed payload (e.g., 0.3 bpnzac) and channel quality 85. If the boundary-preserving variant does not show a higher average detection error than the unconditional variant, the central security claim is falsified. Alternatively, directly measure whether boundary pixels are in fact more detectable than interior pixels under the large modification magnitudes typical of overflow removal; if they are not, the design premise fails.","tokens_in":13272,"feed_emoji":"🕵️","tokens_out":13992,"duration_ms":107443,"temperature":0.7,"pith_summary":"This paper claims that robust JPEG steganography can survive social-network recompression without the usual loss of undetectability. Its preprocessing step removes spatial-domain overflow selectively: overflow inside each 8x8 block is fixed aggressively, while overflow at the block's boundary is left alone unless it is mild. Because boundary modifications are the most detectable kind, and because an adaptive Reed-Solomon loop picks the shortest error-correction code that keeps extraction errors below a threshold, the method reports higher anti-steganalysis performance than ROAST-ST and GMAS at payloads up to 0.5 bits per nonzero AC coefficient while keeping the extraction error rate under 0.0001. If correct, this gives covert communication through JPEG-recompressing channels a better security–robustness–capacity tradeoff.","feed_headline":"Fewer block-edge changes make JPEG steganography harder to detect","feed_subtitle":"Sparing block edges during overflow removal plus adaptive error correction keeps covert JPEG images readable and undetectable.","key_machinery":"The central object is the boundary-preserving overflow preprocessing, which partitions each 8x8 spatial block into an interior (I, 6x6), a boundary (B, the 28 edge pixels including the four corners), and corners (C). The rule 'remove overflow in I when its count exceeds O1; then remove overflow in B only when its count is below O2' is what lets the method stabilize quantized DCT coefficients while leaving the most detectable pixels untouched. This is paired with Algorithm 1, an adaptive RS(31, k*) search that embeds, recompresses, and extracts at each iteration, records the best error rate seen, and stops when the error threshold is met or k* reaches 7. Together they turn the whole image into a robust embedding region with minimal boundary footprint and minimal coding redundancy.","core_discovery":"Analyzing overflow on a benchmark of 10,000 grayscale images compressed at quality factor 65, the authors observe that spatial truncation causes overflow disproportionately at the edges of 8x8 blocks rather than uniformly. They exploit this with a two-threshold preprocessing rule: the 6x6 interior of each block is de-overflowed when its overflow count exceeds O1 (chosen as 0), and the surrounding boundary is de-overflowed only when its own overflow count is below O2 (chosen as 18), so heavily overflowed boundaries are deliberately left modified to preserve inter-block correlations. They complement this with an adaptive RS(31, k*) error-correction search that keeps the code that achieved the lowest measured error rate so far, which avoids the pathological case—documented on smooth-texture images—where longer codes increase the error rate. The paper reports that this combination lifts the average detection error of CCPEV, DCTR, and SRNet steganalysis relative to ROAST-ST and GMAS at payloads up to 0.5 bpnzac for channel qualities 85 and 95, with extraction errors below 0.0001.","pith_inferences":["The same boundary-sparing heuristic could be transferred to other block-based codecs (e.g., HEVC, AV1) or color images, where the block geometry and quantization tables differ, though the overflow statistics would need re-measuring.","If the prior finding that boundary modifications are more detectable does not survive the large modification magnitudes that overflow removal introduces, the O2 threshold would need re-calibration by channel quality and payload.","The adaptive RS search could likely be replaced by a one-shot complexity predictor that estimates the image's error rate from block statistics, avoiding the repeated embed–recompress–extract iterations and lowering computational cost.","A steganalyzer trained specifically on boundary modifications—rather than the general CCPEV/DCTR/SRNet detectors—might reduce the reported security gap; testing against such a detector would sharpen the claim."],"forward_implications":["The method supports payloads of 0.1–0.5 bpnzac—higher than the 0.05–0.15 range typical of previous robust adaptive steganography—without collapsing security.","The adaptive RS loop with best-error tracking removes the failure mode, documented on smooth-texture images, where lengthening the error-correction code raises the extraction error rate.","Reported security gains include over 2% higher average detection error than ROAST-ST at 0.16 bpnzac under a CNN-based steganalyzer, and over 10% higher DCTR detection error than fixed-RS coding at 0.5 bpnzac.","Because overflow preprocessing stabilizes coefficients across the entire DCT domain, the method eliminates the need for robust-domain selection, simplifying the embedding pipeline."],"supporting_citations":[{"why":"Provides the ROAST-ST and ROAST-OS baselines whose overflow preprocessing the proposed method modifies to spare block boundaries.","marker":"(Zeng et al., 2024)"},{"why":"Supplies the premise that spatial block boundaries have lower empirical security, which justifies the boundary-preserving strategy.","marker":"(Wang et al., 2020)"},{"why":"Defines GMAS, the generalized dither modulation and asymmetric distortion framework that serves as the embedding base and as a primary security baseline.","marker":"(Yu et al., 2020)"},{"why":"Introduces Adaptive-GMAS, the adaptive error-correction baseline and the source of the 0.0001 robustness threshold that this paper refines.","marker":"(Duan et al., 2023)"},{"why":"Provides the Reed-Solomon error-correcting codes whose redundancy the adaptive loop tunes.","marker":"(MacWilliams and Sloane, 1977)"},{"why":"Supplies syndrome-trellis codes (STC), the embedding method used to hide the RS-encoded message with minimal additive distortion.","marker":"(Filler et al., 2011)"},{"why":"Defines DCTR steganalysis features, one of the three detectors used to measure the claimed anti-steganalysis gain.","marker":"(Holub and Fridrich, 2014)"},{"why":"Defines CCPEV features (PEV with Cartesian calibration), another detector used for the security evaluation.","marker":"(Kodovsk`y and Fridrich, 2009)"},{"why":"Provides SRNet, the deep residual steganalysis network used in the low-payload comparison.","marker":"(Boroumand et al., 2019)"},{"why":"Supplies the benchmark dataset of 10,000 grayscale images on which the overflow statistics, parameter choices, and evaluations are computed.","marker":"(Bas et al., 2011)"}],"fun_headline_variants":["Block-edge overflow fixes boost JPEG stego stealth","Smarter overflow handling makes JPEG stego more secure","Adaptive error correction cuts stego detectability","Preserving block edges improves stego robustness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The security gain rests on the premise, taken from prior work cited in the paper, that modifying spatial block boundaries is more detectable than modifying block interiors, and that this ordering still holds for the relatively large overflow-removal adjustments the preprocessing makes.","fun_headline_variants_meta":{"raw":{"variants":["Block-edge overflow fixes boost JPEG stego stealth","Smarter overflow handling makes JPEG stego more secure","Adaptive error correction cuts stego detectability","Preserving block edges improves stego robustness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1644,"prompt_tokens":995,"completion_tokens":649,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":611,"tokens_out":649,"duration_ms":6251,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:50:08.857121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pipeline twice on the same covers, once with the O2 boundary-stopping rule (the proposed method) and once with unconditional boundary de-overflowing (equivalent to ROAST-ST preprocessing), while keeping the adaptive RS loop identical, and compare the reported steganalysis detectors at a fixed payload (e.g., 0.3 bpnzac) and channel quality 85. If the boundary-preserving variant does not show a higher average detection error than the unconditional variant, the central security claim is falsified. Alternatively, directly measure whether boundary pixels are in fact more detectable than interior pixels under the large modification magnitudes typical of overflow removal; if they are not, the design premise fails.","supporting_citations":[{"cited_title":", author Chen, K","cited_arxiv_id":null,"evidence_quote":"Provides the ROAST-ST and ROAST-OS baselines whose overflow preprocessing the proposed method modifies to spare block boundaries."},{"cited_title":", author Zhang, W","cited_arxiv_id":null,"evidence_quote":"Supplies the premise that spatial block boundaries have lower empirical security, which justifies the boundary-preserving strategy."},{"cited_title":", author Li, B","cited_arxiv_id":null,"evidence_quote":"Introduces Adaptive-GMAS, the adaptive error-correction baseline and the source of the 0.0001 robustness threshold that this paper refines."},{"cited_title":", author Sloane, N.J.A","cited_arxiv_id":null,"evidence_quote":"Provides the Reed-Solomon error-correcting codes whose redundancy the adaptive loop tunes."},{"cited_title":", author Judas, J","cited_arxiv_id":null,"evidence_quote":"Supplies syndrome-trellis codes (STC), the embedding method used to hide the RS-encoded message with minimal additive distortion."},{"cited_title":", author Fridrich, J","cited_arxiv_id":null,"evidence_quote":"Defines CCPEV features (PEV with Cartesian calibration), another detector used for the security evaluation."},{"cited_title":", author Chen, M","cited_arxiv_id":null,"evidence_quote":"Provides SRNet, the deep residual steganalysis network used in the low-payload comparison."},{"cited_title":", author Filler, T","cited_arxiv_id":null,"evidence_quote":"Supplies the benchmark dataset of 10,000 grayscale images on which the overflow statistics, parameter choices, and evaluations are computed."}],"review_version":1}