{"id":"1056b467-11e6-4bed-bdbb-ffe8bfd247dd","arxiv_id":"1908.01947","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DCDT transforms spatial steganographic costs into JPEG costs via an exponent-weighted inverse-DCT footprint, and the authors report gains over UERD, J-UNIWARD and GUED against GFR/SCA-GFR.","lead":"This paper introduces DCDT, a way to turn a spatial steganography cost function into a JPEG-domain cost function by weighting each DCT coefficient change by the spatial pixel changes it causes. The authors report that the resulting scheme resists modern JPEG steganalysis better than several prior schemes while staying computationally cheap.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed GFR/SCA-GFR superiority depends on p tuned per steganalyzer on the same benchmark; with the fixed SCA-GFR p used in the main tables, DCDT-HiLL is not consistently superior to GUED or J-UNIWARD against GFR at Q75.","rationale":"The reader's weakest_assumption focuses on the linearity of spatial distortion costs. That is a real conceptual gap, but the more immediately load-bearing threat to the empirical headline comes from Section 3.2 and Section 3.4: p is tuned per steganalyzer on BOSSBase, the main table uses only the SCA-GFR-optimal p for all adversaries, and no error bars are reported, so the 'superior against GFR' claim is partly a product of parameter selection. I agree with the reader's CONDITIONAL verdict: the paper should be accepted conditional on held-out p selection, variance reporting, and ideally a sensitivity analysis of the linearity assumption. These are addressable without changing the core construction, but they must be addressed before the broad superiority claim is taken at face value.","tokens_in":16857,"tokens_out":11254,"duration_ms":121660,"concrete_test":"Fix p once per quality factor using only a held-out validation set (e.g., BOWS2 or a 20% BOSSBase split), with no per-steganalyzer selection on the test benchmark, and rerun Table 5 and Table 6 for GFR and SCA-GFR over the ten random train/test splits, reporting mean PE plus standard deviation. If DCDT-HiLL with this fixed p is not significantly better than GUED or J-UNIWARD against GFR at Q75, the abstract's 'superior' claim should be restricted to the per-steganalyzer-tuned p setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 tunes the exponent p in Eq. (5) on a 5000-image BOSSBase split to maximize PE separately for each steganalyzer, yielding p=0.5/0.9 for SCA-GFR and p=0.7/1.1 for GFR. Section 3.4 then reports the headline DCDT-HiLL using only the SCA-GFR-optimal p, while only the separately labeled DCDT-HiLL-pro uses the GFR-optimal p. The abstract's broad claim of superiority against GFR is therefore evaluated at a setting that Section 3.4 itself calls suboptimal for GFR. In Table 5 (GFR, Q75), DCDT-HiLL with p=0.5 has PE 40.85/29.33/18.62 at 0.1/0.2/0.3 bpnzAC versus GUED's 41.57/29.93/19.13, so the method is not superior at those payloads; only DCDT-HiLL-pro (p=0.7, tuned on the same BOSSBase) overtakes GUED. No standard deviations or confidence intervals are reported for the ten random test splits, and several reported gaps are below 1%, so the ordering may be sampling noise. The central claim as stated is conditional on per-adversary p selection on the evaluation benchmark, not a robust property of the DCDT transformation itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new JPEG steganographic distortion cost function obtained by transforming a spatial-domain embedding cost into the DCT domain. The core formula, Eq. (5), computes the cost of a unit change of a DCT coefficient as the p-th power of pixel-level spatial costs weighted by the inverse-DCT footprint of that coefficient change. The authors instantiate the framework with the HiLL spatial cost, tune the exponent p per steganalyzer and quality factor on BOSSBase, and report experiments against CC-JRM, GFR, and SCA-GFR on BOSSBase and BOWS2, comparing with UERD, J-UNIWARD, GUED, and BET-HiLL. They also propose a mutually dependent embedding extension and measure computational complexity.","tokens_in":17215,"tokens_out":7288,"duration_ms":74345,"significance":"If the claims were fully supported, the paper would offer a simple, interpretable, and computationally efficient way to transfer well-developed spatial steganographic costs to JPEG steganography, with particular benefit against phase-aware JPEG steganalyzers. The derivation of Eq. (5) is straightforward, the computational advantage over BET-HiLL is practically relevant, and the BOWS2 transfer test gives some independent support for the SCA-GFR setting. However, the central security claim is weakened by the per-steganalyzer tuning of p on the same benchmark used for evaluation, the absence of variance information, and a stated superiority over GUED that is not borne out in the main table for GFR at Q75. The core idea is defensible, but the current presentation overstates the evidence.","major_comments":[{"comment":"The abstract and the conclusion claim that DCDT-HiLL is superior to GUED in resisting GFR, but Table 5 does not support this for Q75. Under the main setting p=0.5 (chosen for SCA-GFR), DCDT-HiLL has lower PE than GUED at 0.1, 0.2, and 0.3 bpnzAC (40.85 vs 41.57, 29.33 vs 29.93, 18.62 vs 19.13). Even the GFR-tuned DCDT-HiLL-pro with p=0.7 is lower than GUED at 0.1 bpnzAC (41.30 vs 41.57). The claim of superiority against GFR should therefore be weakened to \"comparable or mixed\" for Q75.","section":"Section 3.4, Table 5"},{"comment":"The exponent p in Eq. (5) is selected by maximizing PE on a 5,000-image subset of BOSSBase separately for each steganalyzer, and the main security tables are then reported on BOSSBase without stating whether the tuning subset is disjoint from the final training/testing images. Because p differs by steganalyzer (0.7 for GFR vs 0.5 for SCA-GFR at Q75), the reported gains against GFR and SCA-GFR are partly a product of per-adversary benchmark tuning. The BOWS2 transfer experiment in Table 9 only evaluates SCA-GFR with the SCA-GFR p, so it does not independently validate the GFR-specific claims. Please report results with a fixed p chosen without reference to the target steganalyzer, or with a fully disjoint validation set, and explicitly state the relationship between the tuning split and the evaluation split.","section":"Sections 3.2 and 3.4"},{"comment":"The linear regression rule p = 0.02*(QF-75)+0.48 is fitted to optimal p values at QFs 75, 80, 85, 90, and 95, and Table 11 then evaluates DCDT-HiLL at QFs 80, 85, and 90 using p values from this regression. Since the evaluation QFs are the same ones used to fit the regression, this is not an independent test of the extrapolation rule. To support the claim that the method works for a variety of quality factors, validate on QFs not used in constructing Eq. (9), such as Q82 or Q88, or use leave-one-QF-out cross-validation.","section":"Section 3.6, Eq. (9), Tables 10 and 11"},{"comment":"The entire DCDT transformation rests on the assumption that spatial distortion cost is linearly proportional to the amplitude of the pixel modification. This is not self-evident because |s_{a,b}(i,j)| can exceed 1 and varies with the DCT mode and QF. The paper provides no empirical validation of this linearity or comparison with alternative amplitude models. Please add a robustness test, for example by replacing |s| with |s|^q or a saturating nonlinearity, or provide explicit justification for the linear model at the tested QFs and payloads.","section":"Section 2.3, Eq. (4)"},{"comment":"Only mean classification error probabilities over ten random test splits are reported, without standard deviations, confidence intervals, or significance tests. Several differences that the paper treats as security gains are below one percentage point, e.g., in Table 5 at Q75 with GFR at 0.4 bpnzAC, DCDT-HiLL has PE 10.97 vs GUED 11.14, and at 0.5 bpnzAC, 6.27 vs 6.10. These differences are likely within sampling noise. Please report error bars or statistical tests for the main comparisons that support the headline claims.","section":"Sections 3.1 and 3.4"}],"minor_comments":[{"comment":"The sentence \"we set the threshold T and penalty factor v in Eq. (8) as 10\" is ambiguous: it is unclear whether both T and v are set to 10 or only one of them. Please clarify.","section":"Section 2.4"},{"comment":"The method name is inconsistently capitalized as both 'DCDT-HiLL' and 'DCDT-Hill' (e.g., Sections 3.4 and 3.7, Table 7). Please unify the notation.","section":"Throughout"},{"comment":"The text says that p 'flexibly adjusts the embedding distributions,' but the precise effect of larger p on block and mode selection is not discussed until Section 3.2. Adding one sentence after Eq. (5) explaining that larger p penalizes high-cost pixels more strongly would improve readability.","section":"Section 2.3, after Eq. (5)"},{"comment":"There are typographical errors, including 'espacilly' in Section 3.7 and 'stgeanographic' in Section 4. A careful proofread is recommended.","section":"Sections 3.7 and 4"},{"comment":"The statement that using the SCA-GFR parameter setting for other steganalyzers 'won't change much' is supported only by the Q75 DCDT-HiLL-pro rows in Table 5. Consider providing a corresponding comparison for Q95 GFR or CC-JRM, or soften the claim.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The derivation of DCDT is technically sound and the computational efficiency result is likely useful to the community, but the evaluation protocol and the abstract's claims need substantial reworking before publication. The key issue is that p is tuned per steganalyzer on the same benchmark used for the headline comparison, and the stated superiority over GUED for GFR is not actually present in Table 5 for the main configuration. I see no indication of misconduct, but the authors should be asked to clarify the train/test/tuning split and to either provide independent validation or weaken the claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate extension of the JPEG steganography cost-design line, and the complexity result is real. But the abstract overstates the security win: the GFR/SCA-GFR superiority only holds for p values tuned per steganalyzer on BOSSBase, and the main tables use the SCA-GFR-optimal p for all comparisons, which makes DCDT-HiLL not superior to GUED against GFR at Q75. Read the claims with that caveat.\n\nWhat's actually new: Eq. (5) is a clean way to lift a spatial cost into JPEG domain via the IDCT footprint. The DCDT framing is more general than J-UNIWARD/GUED/BET-HiLL, and the exponent p is a useful tuning knob. The paper is honest about GFR degradation at Q75 and presents DCDT-HiLL-pro to show the effect of choosing the right p. The speed advantage over BET-HiLL (roughly an order of magnitude, Table 8) is concrete and reproducible. The related work is cited fairly; the debt to J-UNIWARD, GUED, and BET-HiLL is acknowledged.\n\nWhere it's soft: the main tables use p=0.5 for Q75 regardless of steganalyzer, which is the SCA-GFR-optimal value. For GFR, Table 5 shows DCDT-HiLL at p=0.5 is below GUED at every payload; only DCDT-HiLL-pro (p=0.7) beats GUED at most payloads. So the abstract's blanket 'superior to GUED for GFR' is not what the table with the stated setting shows. p selection is also done on BOSSBase, and the same database is used for headline comparisons; BOWS2 transfer helps but only checks SCA-GFR against J-UNIWARD, not the GFR claim. No error bars or per-split variance are reported, and several gaps are under 1%, so the ranking could be noise. The linearity assumption (spatial cost proportional to modification amplitude) is stated but never validated; it's a modeling choice that could be wrong, and there is no sensitivity analysis.\n\nVerdict: worth engaging with. The central idea is sound, the math in Eq. (5) is simple and correct, and the complexity result is valuable. I would send it to peer review with a request to fix the reporting: use per-steganalyzer p consistently in main tables or clearly mark which p is used, report variance, and add sensitivity analysis on the linearity assumption. As is, the abstract is stronger than the evidence.\n\nRecommendation: send to peer review. The work is solid enough to deserve referee time, and the flaws are correctable rather than fatal.","headline":"Useful extension of JPEG cost design with a real speed advantage, but the headline security claim depends on benchmark-tuned p and the abstract overstates GFR superiority.","tokens_in":17764,"tokens_out":5707,"would_cite":true,"duration_ms":99286,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"JPEG steganography becomes harder to detect when the distortion cost is computed on decompressed pixels and then transformed to DCT coefficients through inverse-DCT footprints, rather than designed directly on DCT coefficients.","keywords":["JPEG steganography","distortion cost function","domain transformation","phase-aware steganalysis","GFR","SCA-GFR","HiLL","exponential model"],"falsifier":"Take a single 8-by-8 block with known texture, flip one DCT coefficient, and measure the actual increase in GFR or SCA-GFR feature responses in both the DCT and decompressed spatial domains; if the ranking of candidate coefficients by DCDT cost does not match the ranking by measured feature increase, the linear-proportionality and same-priority assumptions are contradicted.","tokens_in":16651,"feed_emoji":"🖼️","tokens_out":9170,"duration_ms":80071,"temperature":0.7,"pith_summary":"JPEG steganography hides messages by slightly changing quantized DCT coefficients, and its security is judged by steganalyzers that now look at both DCT and spatial statistics. This paper proposes DCDT, a Distortion Cost Domain Transformation, that computes each DCT coefficient's embedding cost by mapping a spatial-domain distortion cost back into the JPEG domain: the cost of a unit change at a DCT mode is the sum, over the 8-by-8 pixel block, of the pixel's spatial cost raised to an exponent $p$ and multiplied by the absolute amplitude of that mode's inverse-DCT footprint. The claim is that, equipped with the HiLL spatial cost function, this transformed cost outperforms DCT-native schemes (UERD, J-UNIWARD, GUED) against the phase-aware steganalyzers GFR and SCA-GFR, and matches BET-HiLL while being about ten times faster. If true, the design philosophy for JPEG steganography shifts: instead of inventing new DCT-domain cost functions, one can reuse mature spatial-domain cost functions and convert them.","feed_headline":"Pixel-to-DCT cost transform hides JPEG messages better","feed_subtitle":"Reusing spatial embedding costs via inverse-DCT footprints beats DCT-native schemes against GFR and SCA-GFR.","key_machinery":"The central object is the DCDT transformation function in Eq. (5): $\\rho_{a,b}^{m,n} = \\sum_{i,j} (d^{m,n}(i,j))^p |s_{a,b}(i,j)|$. It carries the argument by converting a spatial-domain distortion cost, defined per pixel per unit change, into a DCT-domain cost per coefficient per unit change through the inverse-DCT footprint of each mode. The exponent $p$ is the tunable mechanism: it controls how strongly pixel costs are amplified before summation, which in turn shifts embedding away from low-frequency modes whose spatial footprints are large, and the authors select $p$ by grid search, summarizing it with an empirical linear rule $p = 0.02 (\\mathrm{QF} - 75) + 0.48$ for the quality-factor interval 75 to 95. The basis of the whole construction is the stated equivalence that a DCT block and its decompressed pixel block have the same embedding priority, plus the assumption that spatial distortion is linearly proportional to pixel modification amplitude.","core_discovery":"The central discovery is that the embedding priority of an 8-by-8 DCT block is the same as that of its decompressed pixel block, and that this equivalence can be made operational through a generalized distortion cost domain transformation. The load-bearing formula is Eq. (5), $\\rho_{a,b}^{m,n} = \\sum_{i,j} (d^{m,n}(i,j))^p |s_{a,b}(i,j)|$, in which $d^{m,n}(i,j)$ is the spatial distortion cost of the $(i,j)$-th pixel for a unit change, $p$ is an exponent that redistributes embedding across blocks, and $s_{a,b}(i,j)$ is the inverse-DCT spatial footprint produced by a unit modification at DCT mode $(a,b)$. With this formula, JPEG embedding is posed as minimizing distortion in the decompressed spatial domain, which the authors argue is equivalent to maximizing statistical undetectability against phase-aware steganalyzers. In their experiments, DCDT-HiLL (HiLL as the spatial cost) improves classification error over UERD, J-UNIWARD, and GUED against GFR and SCA-GFR when $p$ is tuned, and matches or exceeds BET-HiLL depending on quality factor, all with roughly one order of magnitude less computation; an optional mutually dependent embedding update gives a further gain.","pith_inferences":["The authors do not claim, but a natural corollary is that DCDT should generalize to any transform codec with a linear synthesis map, such as HEVC or AV1, since the derivation only needs each coefficient's spatial footprint.","The authors do not claim, but a learned cost function trained to predict DCDT costs could retain the security gain at even lower runtime, which would matter for video or large-batch image applications.","The authors do not claim, but the dependence of the optimal exponent on quality factor suggests that a content-adaptive $p$, chosen per block from local texture, might outperform the globally fixed $p$ used in the paper.","The authors do not claim, but an adaptive steganalyzer retrained on DCDT-specific selection channels could narrow the reported gap, so future work should test DCDT against continual retraining rather than fixed features."],"forward_implications":["Other spatial-domain cost functions (S-UNIWARD and MiPOD are tested in the paper) can be lifted to JPEG by DCDT, with HiLL giving the best tested security, so advances in spatial steganography transfer directly to JPEG.","A single scalar exponent $p$ gives the steganographer a lever to trade detectability between DCT-focused and phase-aware steganalyzers, and the paper's empirical rule makes $p$ predictable from the JPEG quality factor.","At equal embedding payload, DCDT-HiLL reports higher detection-error probability than UERD, J-UNIWARD, and GUED against GFR and SCA-GFR, meaning those steganalyzers would need larger payloads or stronger features to detect it.","The computation-time measurements, about 0.05 seconds per 512-by-512 image versus about 0.8 seconds for BET-HiLL, imply that the security gain is achievable in practice, not only in principle.","The mutually dependent embedding extension, although exponential in the number of modified coefficients per block, shows that accounting for joint pixel changes within a block yields additional security, pointing toward joint-cost models."],"supporting_citations":[{"why":"Supplies the HiLL spatial cost function that DCDT transforms into the JPEG domain.","marker":"[11]"},{"why":"UERD baseline; DCDT claims higher security at similar computational cost.","marker":"[14]"},{"why":"Defines J-UNIWARD baseline and the universal-distortion approach that DCDT extends.","marker":"[15]"},{"why":"GUED baseline, introduced decompressed pixel block changes into JPEG costs.","marker":"[16]"},{"why":"BET baseline; BET-HiLL is the rival whose embedding-entropy domain transformation motivates DCDT.","marker":"[17]"},{"why":"Provides STC minimal-distortion embedding, the framework that consumes the DCDT costs.","marker":"[19]"},{"why":"Defines GFR, the phase-aware steganalyzer whose detection DCDT aims to resist.","marker":"[20]"},{"why":"Defines SCA-GFR, the selection-channel-aware steganalyzer used for the main security comparisons.","marker":"[21]"}],"fun_headline_variants":["Spatial-to-DCT cost transform outwits JPEG phase-aware steganalysis","JPEG stego: cost from pixel domain beats DCT-native schemes","DCDT: a faster JPEG stego cost that rivals BET-HiLL","Reusing pixel costs via inverse-DCT improves JPEG stego","New JPEG cost function reduces detection by GFR and SCA-GFR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes that a pixel's spatial distortion cost grows linearly with the amplitude of the pixel change and that an 8-by-8 block has the same embedding priority in the DCT domain as in the decompressed pixel domain; if either fails, the transformed DCDT cost misrepresents the true embedding impact of a DCT modification.","fun_headline_variants_meta":{"raw":{"variants":["Spatial-to-DCT cost transform outwits JPEG phase-aware steganalysis","JPEG stego: cost from pixel domain beats DCT-native schemes","DCDT: a faster JPEG stego cost that rivals BET-HiLL","Reusing pixel costs via inverse-DCT improves JPEG stego","New JPEG cost function reduces detection by GFR and SCA-GFR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000964,"raw_usage":{"total_tokens":4190,"prompt_tokens":1117,"completion_tokens":3073,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":733,"completion_tokens_details":{"reasoning_tokens":2974}},"tokens_in":733,"tokens_out":3073,"duration_ms":24643,"temperature":1.0,"reasoning_tokens":2974,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:58:16.187383+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a single 8-by-8 block with known texture, flip one DCT coefficient, and measure the actual increase in GFR or SCA-GFR feature responses in both the DCT and decompressed spatial domains; if the ranking of candidate coefficients by DCDT cost does not match the ranking by measured feature increase, the linear-proportionality and same-priority assumptions are contradicted.","supporting_citations":[{"cited_title":"Holub, J","cited_arxiv_id":null,"evidence_quote":"Defines J-UNIWARD baseline and the universal-distortion approach that DCDT extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BET baseline; BET-HiLL is the rival whose embedding-entropy domain transformation motivates DCDT."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines GFR, the phase-aware steganalyzer whose detection DCDT aims to resist."},{"cited_title":"Denemark, M","cited_arxiv_id":null,"evidence_quote":"Defines SCA-GFR, the selection-channel-aware steganalyzer used for the main security comparisons."}],"review_version":1}