acceptodds
Under review as a conference paper at ICLR 2027

-Steg: Generative Image Steganography for Integrated Covert Communication and Copyright Protection

Abstract

Generative steganography and copyright watermarking, both subsumed under information hiding, have long been studied as separate paradigms. No single model supports covert communication and copyright authentication simultaneously. Direct integration of the two tasks confronts three fundamental challenges: (1) **Steganography capacity zero-sum game**: Simultaneously hiding secret images and watermarks makes secret residuals block watermark extraction, and watermark features leak steganographic traces, jointly compromising both tasks’ security. (2) **Semantic representation deficiency**: Neither task is modeled at a semantic level, and joint two tasks in low-level numerical space causes mutual feature interference, degrading recovery fidelity. (3) **Conflicting security boundaries**: A single key for both streams makes secret data and watermarks indistinguishable under coercion: implausible secret denial (weakening steganography) or invalid ownership proof (invalidating authentication). To systematically address these issues, this paper makes the first attempt at a unified diffusion-based generative image steganography framework, **-Steg**, which integrates covert communication and copyright protection in a single pipeline and is supported by a rigorous theoretical foundation. Specifically, the proposed -Steg framework conducts a **shared Discrete Semantic Space** to elevate steganography from low-level numerical space to the high-level semantic domain. Feature mutual interference and insufficient semantic representation induced by low-level steganography are thereby eliminated. Moreover, a **Disjoint Dual-key Mask Decoupling mechanism** is designed, by which the concealing space is spatially decoupled to overcome the zero-sum capacity trade-off caused by constrained steganography resources. Furthermore, the dual-key design resolves the conflicting security boundaries: **plausible deniability for steganography** under coercion (the watermark key alone recovers a legitimate watermark while secret images remain undetectable) and verifiability of copyright watermarks. Experiments demonstrate the -Steg’s outstanding performance in steganography fidelity, security, and copyright verifiability.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Generative image steganography, copyright watermarking, deniability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_RUTnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents C2C^2-Steg, a diffusion‑based generative image steganography framework that unifies covert secret‑image communication and copyright watermarking within one end‑to‑end generative pipeline. A Theoretical Dual-Functional Unified Framework seamlessly unifies covert steganographic communication and robust copyright watermarking within a single learning paradigm, Error-bounded Semantic Steganography guarantees controllable semantic distortion and bounded extraction error for hidden messages under common perturbations, and Dual-key Decoupling separates the communication secret key and copyright verification key to support independent access control for message decoding and ownership authentication. Extensive validations are conducted across standard steganography and watermarking benchmarks under diverse realistic distortions including compression, noise addition and geometric transformations, where C2C^2-Steg achieves competitive visual fidelity, high hidden-bit extraction accuracy and reliable copyright verification performance.

Critical strengths
  1. The Theoretical Dual-Functional Unified Framework together with Error-bounded Semantic Steganography and Dual-key Decoupling realizes joint covert communication and copyright authentication within a single diffusion pipeline, supporting deniable access via separated keys while bounding semantic distortion and extraction error. This is consequential because prior generative hiding systems mostly optimize steganography or watermarking in isolation, and this integrated security formulation fills a clear gap by enabling fine-grained, secure payload control that cannot be trivially obtained from existing decoupled designs.
  2. The validation spans diverse realistic distortions, standard benchmarks and targeted ablation studies. This is consequential because the empirical evidence systematically verifies not only the baseline performance advantage but also the core security property of dual-key separation, which builds credible support for the practical feasibility of the proposed dual-functional paradigm.
Critical weaknesses
  1. The deniability claim is confusing under standard cryptographic definitions in the paper. The proposed deniability comes from separated mask regions, where the fake key only accesses the watermark payload. This differs from canonical deniable primitives, where multiple plausible plaintexts can be decoded from the same stego object and the adversary cannot identify the true one. The threat model for coercion needs clarification. Besides, the cryptographic nature of the true key and fake key remains underspecified. It is unclear whether they form a public-private pair, independent secret seeds for mask generation or derived from a master key. The key generation algorithm, key space and whether one key can be derived from the other are not clarified.
  2. The error bound in Error-bounded Semantic Steganography is hard to interpret. The definition of semantic distortion and the metrics used to quantify semantic error are not well established. It remains unclear whether the theoretical bound accounts for inherent inversion error of diffusion models, and the empirical validation relies heavily on pixel-level metrics rather than semantic evaluation. Moreover, the security implication of non-overlapping masks is insufficiently analyzed. Partitioning the initial Gaussian noise into two fixed regions may introduce statistical artifacts that can be exploited by steganalyzers. This paper does not discuss steganographic security against detection, nor the failure mode when mask overlap occurs.
Questions for authors
  1. The authors describes the mapping that converts discrete semantic tokens into standard Gaussian noise in Section 3.3. Equation (14) characterizes the token recovery error probability after majority voting over R independent copies of semantic payloads. This work assumes identically and independently distributed errors across individual copies when deriving the binomial error expression and it remains unclear how the boundary terms bs−b_s^- and bs+b_s^+ enter the computation of the single-copy base error PeP_e. The redundancy introduced by these repeated copies may degrade the steganographic indistinguishability bound from Lemma 1 and weaken the hiding security.
  2. The proposed dual-key decoupling adopts true key and fake key to generate non-overlapping masks for payload partition. The two keys do not constitute a standard public-private key pair in asymmetric cryptography while both keys are kept secret, which deviates from the typical single-secret-key setting of symmetric cryptography. The formal cryptographic classification of this dual-key mechanism and the independence assumption between the true key and fake key in the security model remain unclear.
  3. This paper claims deniability property enabled by dual-key access separation. The realization differs from canonical deniable encryption, where multiple plausible plaintexts are decodable under the same key. The mask-based deniability needs to be distinguished from the deniability definition widely adopted in cryptography and the coercion threat model adopted for this formulation lacks explicit definition.
4 comments
AuthorsOpenReview ↗

Payload-Level Key-Scope Deniability with Disjoint Dual-Key Masks, Formal Security Model, Semantic Fidelity, and i.i.d. Error Bounds

We appreciate your feedback and hope our response addresses your points.

W1 Deniability Canonical deniable encryption concerns one ciphertext with multiple explanations. Here payload-level deniability is claimed: the stego image carries secret and watermark on disjoint coordinates. With the watermark key, a verifiable watermark is obtained while the secret remains undetectable. Multiple-secret decodability is not claimed, only inaccessibility under a given key.

Key System. Keys ks,kwk_s,k_w are independent 256-bit seeds, not a key pair, not derived from a master key. Key generation is k←{0,1}256k\leftarrow\lbrace0,1\rbrace^{256}. MaskGen(k,n,Ωpool)MaskGen(k,n,\Omega_{pool}) uses SHA−256(k)SHA-256(k) as cryptographically secure pseudorandom number generator seed and outputs an ordered duplicate-free coordinate set by rejection sampling. Key space is 22562^{256}, and preimage resistance gives cost ≥2256\ge2^{256} to derive ksk_s from kwk_w. Disjoint pools Ωsec,Ωwm\Omega_{sec},\Omega_{wm} make masks statistically independent given one key.

W2 Semantic Fidelity. Ours improves all metrics over CRoSS and DiffStega on UniStega.

MethodToken Acc.CLIP ↑LPIPS ↓PSNR(dB)
CRoSS72.3%0.7420.28721.25
DiffStega79.6%0.8050.21423.29
Ours89.1%0.8830.12624.37

The odd-even split only pre-partitions pools, and positions are cryptographically secure pseudorandom number generator sampled within each pool. No significant deviation from pure Gaussian is detected.

TestOdd vs. Gauss.Even vs. Gauss.Odd vs. Even
KS test p-value0.4620.4550.458
Mean difference<0.001<0.001<0.001
Variance difference<0.2%<0.2%<0.1%

Semantic distortion is token-sequence Hamming distance, and each VQGAN entry is an independent local primitive. Inversion error is included in Eq. (9): dTV(P^,Q)≤Nhδtoken+εinv.d_{TV}(\hat{P},Q)\le N_h\delta_{token}+\varepsilon_{inv}. Overall error thus covers prior deviation and inversion error, with no omission.

Failure modes Pre-partitioned pools guarantee disjoint mask spaces. Malicious overlap causes only local token errors, without global leakage.

Q1 i.i.d. assumption. i.i.d. holds by construction: Us,r∼Uniform(s/K,(s+1)/K),zs,r=Φ−1(Us,r)U_{s,r}\sim Uniform(s/K,(s+1)/K), z_{s,r}=\Phi^{-1}(U_{s,r}). The channel is memoryless, so per-copy errors are i.i.d.

The boundary terms. Eq. (13) includes both boundary terms: Pe(s)=∫Isϕ(z)[Q(bs+−zσ)+Q(z−bs−σ)] dz,P_e(s)=\int_{I_s}\phi(z)[Q(\tfrac{b_s^+-z}{\sigma})+Q(\tfrac{z-b_s^-}{\sigma})]\,dz, with bs−=Φ−1(s/K),bs+=Φ−1((s+1)/K)b_s^-=\Phi^{-1}(s/K), b_s^+=\Phi^{-1}((s+1)/K), which are Gaussian tail probabilities.

Redundancy vs indistinguishability. Uniform prior gives δtoken=0\delta_{token}=0 and zero TV independent of RR. Non-uniform prior gives conservative NhδtokenN_h\delta_{token}, but Pinsker gives dTV(P⊗Nh,Q⊗Nh)≤12Nh KL(Pz∥Qz)=O(Nh δtoken),d_{TV}(P^{\otimes N_h},Q^{\otimes N_h})\le\sqrt{\tfrac{1}{2}N_h\, KL(P_z\|Q_z)}=O(\sqrt{N_h}\,\delta_{token}), so distinguishability grows as Nh\sqrt{N_h}, not NhN_h. Gains saturate at R=5. Detection in Table 5 is near random (SRNet 50.62%, XuNet 50.96%, YeNet 51.07%).

Q2 Cryptographic classification. The scheme is a symmetric multi-key scheme with disjoint key domains, described by (K,Ω,Partition,MaskGen)(K,\Omega,Partition,MaskGen), with K={0,1}256K=\lbrace0,1\rbrace^{256} and Ωsec∪Ωwm=Ω\Omega_{sec}\cup\Omega_{wm}=\Omega. Deniability arises from disjoint domains, not multi-plaintext decodability.

Independence. Keys satisfy I(ks;kw)=0I(k_s;k_w)=0, and conditional independence holds: Pr⁡[Mwm=M∣ks]=(∣Ωwm∣n)−1,\Pr[\mathbf{M}_{wm}=M\mid k_s]=\binom{|\Omega_{wm}|}{n}^{-1}, so H(Mwm∣ks)=H(Mwm)H(\mathbf{M}_{wm}\mid k_s)=H(\mathbf{M}_{wm}).

The security model. In ExpA,Πdeny(λ)Exp^{deny}_{\mathcal{A},\Pi}(\lambda), xstx_{st} is given, at most one key query is allowed, and a guess b′b' is output. Payload-level deniability holds if Pr⁡[ExpA,Πdeny(λ)=1]≤12+negl(λ).\Pr[Exp^{deny}_{\mathcal{A},\Pi}(\lambda)=1]\le\tfrac{1}{2}+negl(\lambda). This parallels IND-CPA, with ciphertext replaced by stego image and plaintext by payload existence.

Q3 Key-Scope Deniability vs Canonical. Canonical deniable encryption requires random coins with Enc(m0;r0)=Enc(m1;r1)=c,Enc(m_0;r_0)=Enc(m_1;r_1)=c, so one ciphertext opens to multiple messages. Here multi-plaintext decodability is not claimed. Under single-key disclosure, the secret is inaccessible and the watermark is genuine. In Exp^{ks\text{-}deny}_{\mathcal{A},\Pi}(\lambda)$, (ks,kw)(k_s,k_w) are sampled, the scheme and partition are published, b←{0,1}b \leftarrow \lbrace0,1\rbrace is sampled, both payloads are embedded if b=0b=0 and only the watermark if b=1b=1, xstx_{st} is given to A\mathcal{A}, at most one key query with disclosure {kw}\{k_w\} is allowed, and b′b' is output with advantage Adv=∣Pr⁡[b′=b]−12∣Adv=|\Pr[b'=b]-\tfrac12|. Key-scope deniability holds if Adv≤negl(λ)Adv\le negl(\lambda), and it fails if both keys are coerced. Deniability is access-control-based, not equivocation-based, which is weaker than canonical deniable encryption but matches copyright coercion.

Reviewer_RUTnOpenReview ↗

Official Comment by Reviewer

Thanks for authors' response. I think the rigorous responses have not alleviated my concern about the watermark keys, especially for the failure modes in this paper. Therefore, I decide to maintain my rating score.

Dear Reviewer RUTn,

Thank you again for your insightful feedback on our submission and your careful re-evaluation. To better clarify the concerns regarding watermark keys and corresponding failure modes, could you kindly elaborate on the specific points that remain unclear or unsatisfactory?

We have tried our best to respond to your comments in the rebuttal document. Should you have any further questions or additional concerns, we are more than happy to provide further clarification. If additional experiments are needed, please note that they will require extra time to complete. We are willing to carry out such experiments if you believe they can help resolve your concerns.

We hope to clear up any potential misunderstandings and fully address your remaining doubts.

Thank you once again for your time and thoughtful consideration. We look forward to your response.

Sincerely,

Authors of Paper #13014

AuthorsOpenReview ↗

Response to the Comment on Watermark Key Failure Modes

The continued feedback from the reviewer is sincerely acknowledged and hope our response addresses your points.

The remaining concern regarding watermark key related failure modes is fully respected. To make the security boundaries transparent, all relevant failure modes analyzed in this work, their impact scopes, and the concrete revision plan are systematically summarized below.

First, it is clarified that all watermark key related failure scenarios have been quantitatively evaluated. None of them is found to lead to catastrophic or abrupt security breakdown. The security degradation is always bounded and gradual.

  1. Imperfect diffusion inversion
    Security degradation is smooth and bounded with no abrupt failure. Even with 2.6×2.6\times relative inversion error, deniability remains valid, with KS test p value >0.05>0.05 and steganalysis accuracy <55%<55\%. The error upper bound is formally characterized by the TV distance in Theorem 1 rather than being treated as an unquantified risk.

  2. Cryptographic implementation flaws
    Mask generation is built on standardized SHA256 and cryptographically secure pseudorandom number generator primitives instead of custom cryptographic designs. Mask uniformity has been validated across three independent standard library implementations, and no detectable statistical bias is observed. Routine implementation imperfections will not break security. Only full disclosure of the key can compromise confidentiality, which is a fundamental property of all symmetric key systems rather than a design flaw of this work.

  3. Accidental mask overlap
    Under the standard algorithmic procedure, the two masks are generated from strictly disjoint candidate pools, so overlap cannot occur in normal deployment. For the extreme case of artificially forced overlap, the failure is strictly local. Only tokens at overlapping positions are corrupted, and performance degradation scales linearly with the overlap ratio. No global secret leakage or complete deniability breakdown occurs.

  4. Partition detection under public scheme
    When the dual key design is known to the adversary and spatial nonstationarity across partitions is tested, all statistical tests, including KS, mean, and variance tests, fail to distinguish the two subspaces. Binary detection accuracy is close to random guess. This attack vector does not constitute an effective threat to deniability.

  5. Watermark key leakage
    The two keys are independently sampled and cannot be derived from each other under the SHA256 preimage resistance assumption. Leakage of the watermark key reveals no information about the secret key and causes no compromise of steganographic secrecy. This is the core security guarantee of the dual key decoupling design.

It is acknowledged that the original manuscript did not organize these analyses into a dedicated failure mode section, which may have created the perception that some edge cases were untested. This is a presentation gap rather than a methodological deficiency.

The comment is highly valued. It is hoped that this systematic clarification and the supplementary content in the final revision will further alleviate the reviewer’s concern and improve the rigor of the work.

Reviewer_YbkYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The problem formulation relies on strong assumptions, and the proposed method appears closely related to existing distribution-preserving embedding strategies. The current analysis does not sufficiently demonstrate a substantial improvement over prior work.

Critical strengths

The paper studies an interesting direction of integrating steganography and watermarking in generative models.

Critical weaknesses
  1. The paper assumes that secret information and copyright watermarks should share the same latent space, leading to a capacity trade-off. However, it is unclear whether this conflict is inherent to steganography and watermarking or mainly caused by the proposed joint embedding design. In practice, these two tasks are usually controlled by different parties. Why is a shared hidden space necessary?

2.The core method maps semantic tokens to Gaussian noise through quantile mapping, which is closely related to existing distribution-preserving latent embedding methods. It is unclear whether extending the message representation from bits to semantic tokens provides any fundamental theoretical or security advantage.

3.The security guarantee relies on Gaussian distribution preservation under a uniform token distribution assumption. However, real semantic tokens are highly non-uniform and structured. Does the proposed mapping preserve the complete high-dimensional Gaussian prior under realistic token distributions, including dependencies among different dimensions?

  1. The security conflict addressed by the dual-key design appears to originate from the shared-key assumption. If secret communication and watermarking already use independent keys and embedding spaces, does this conflict still exist?
Questions for authors

See weaknesses

2 comments
AuthorsOpenReview ↗

Dual-Key Semantic Tokenization with Disjoint Mask Decoupling for Error-Bounded and Deniable Steganography–Watermarking

We appreciate your feedback and hope our response addresses your points.

W1 Two Notions of Conflict. Two notions of conflict should be separated. The first is an intrinsic total capacity budget constraint. Any single carrier image has finite capacity, and any dual-functional method shares this budget.

The second is zero-sum interference, namely mutual degradation from direct superposition of secret and watermark residuals on the same latent or pixel representation. This conflict is not intrinsic but an artifact of naive joint embedding. The latter is identified in Section 1, and its elimination is the design goal of C²-Steg. Through discrete semantic tokenization and disjoint dual-key masks, the two payloads are assigned to mutually exclusive coordinate pools, so feasible regions are independent and no cross-penalty term exists.

The necessity of a shared hidden space follows from deniability. Under deniability, only a single stego image should be visible upon inspection of all deliverable artifacts. If secret and watermark were placed in separate carriers, a second carrier would require explanation after disclosure of the watermark key, and the secret could not be plausibly denied. Co-location is the logical precondition for the claim that one watermarked image denies the existence of a secret. Control by different parties motivates dual keys. The secret key is held by the covert communicator and the watermark key by the copyright holder. Masks generated by the two keys reveal their payloads independently, with neither party accessing the other's information.

W2 Semantic Tokenization. Firstly, bit-level distribution-preserving embedding is a special case at K=2. The method generalizes bit-level embedding to a K-ary codebook with no additional theoretical loss.

Secondly, the advantage lies in error boundedness. Bit errors may propagate arbitrarily during decoding, and no bound on semantic distortion exists. Token-level mapping reduces maximum-likelihood detection under a uniform prior to interval decision (Eqs. 11-13), yields a closed-form single-copy error probability, and constrains semantic distortion after VQGAN decoding through the nearest-neighbor codebook distance, providing a closed-form revealing-error bound.

Thirdly, the security advantage is reflected in the capacity-security trade-off. The bound in Theorem 1, dTV(P^,Q)≤Nhδtoken+ϵinv,d_{TV}(\hat{P},Q)\le N_h\delta_{token}+\epsilon_{inv}, has the same form as its bit-level counterpart, yet the per-coordinate payload increases from 1 bit to log⁡2K=14\log_2 K=14 bits, a 14-fold increase under the same NhN_h and bound. Exact preservation P=Q under a uniform prior holds independently of K. Token-level representation also permits a natural-image watermark, the precondition for substantiating deniability with a high-fidelity watermark.

W3 Non-Uniform Priors and Dependence. A uniform prior is only sufficient for exact preservation P=Q not necessary for boundedness. Under a non-uniform prior, the theorem provides a bound with δtoken=∑s∣πs−1/K∣\delta_{token}=\sum_s|\pi_s-1/K|, and the uniform prior is recovered at δtoken=0\delta_{token}=0. The bound is relaxed by a measurable deviation term. For spatial dependence, an extended bound follows from Pinsker's inequality,

=\sqrt{\tfrac{1}{2}(\sum_{i=1}^{L}KL(P_{Z_i}\|Q)+ TC(S))},$$ where $TC(S)$ is total correlation, estimable on real data. Dependence enters as an additive term and does not break preservation. Measured deviations stay below 0.04, far below the steganalyzer threshold, as summarized below. |Dataset|$\delta_{token}$|Single-coord. TV|KS p|Cov. Homog. p| |---|---:|---:|---:|---:| |UniStega|0.032|0.032|0.412|0.387| |DIV2K|0.028|0.028|0.435|0.402| |Stega260|0.035|0.035|0.396|0.364| W4 **Naive Splitting versus Dual-Key Decoupling.** The security boundary conflict is not introduced by the shared-key assumption but is an endogenous contradiction of bearing both tasks on one carrier. Even with independent keys and naive spatial splitting, three conflicts persist, namely zero-sum capacity trade-off, feature mutual interference, and deniability failure. Controlled experiments compare the naive baseline and the proposed scheme, as summarized below. These results verify that simple key and space splitting cannot eliminate the core conflicts. |Metric|Naive Split|Ours|Improvement| |---|---|---|---| |Permission isolation|Yes|Yes|-| |Average recovery PSNR|21.41dB|24.33dB|+2.92dB| |SRNet accuracy|60.32%|50.62%|-9.70%| |Plausible deniability|No|Yes|Qualitative| |Distribution consistency|No|Yes|Qualitative| The dual-key design is not a superficial split of one key into two but an architectural-level solution coupled with discrete semantic space, disjoint mask decoupling, and distribution preservation. Permission isolation, distribution consistency, low feature interference, and deniability under coercion are simultaneously achieved.

Dear Reviewer YbkY,

Thank you again for your careful review and constructive suggestions. We sincerely appreciate the time and effort you have devoted to evaluating our work.

We have provided point-by-point responses, additional experiments, and clarifications in our rebuttal. As the discussion period draws to a close, we would be grateful if you could let us know whether these responses have addressed your concerns and which issues, if any, remain unresolved. We welcome further discussion and would be happy to provide timely clarifications before the deadline.

Thank you for your careful consideration and the effort you have invested in helping us improve our work.

Best regards,

Authors of Paper #13014

Reviewer_vqbVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes C2C^2-Steg, a diffusion‑based unified generative image steganography framework leveraging a shared discrete semantic space and disjoint dual‑key mask decoupling to simultaneously realize covert secret‑image communication and verifiable copyright watermark protection with plausible deniability under coercion, alongside supporting theoretical bounds and comprehensive experimental validations against competing steganography approaches.

Critical strengths

This paper resolves three core inherent conflicts (capacity zero‑sum game, low‑level feature interference, contradictory security requirements) that hinder joint covert communication and copyright watermarking by introducing discrete semantic space representation and disjoint dual‑key mask decoupling, enabling plausible deniability absent from prior generative steganography methods.

Critical weaknesses
  1. The practical embedding capacity is constrained by the disjoint dual‑mask spatial partitioning design, which splits the available embedding space for secret images and copyright watermarks and imposes hard trade‑offs between the payload size of hidden secret content and watermark information.
  2. Although the method achieves plausible deniability when disclosing only the watermark key, the security guarantees heavily rely on the cryptographic mask‑generation mechanism and exact diffusion inversion. Any inversion inaccuracy or cryptographic implementation flaw could break deniability, steganographic secrecy and copyright verification without thorough real‑world adversarial testing.
Questions for authors
  1. Could the authors quantitatively analyze the embedding‑capacity trade‑off induced by disjoint mask partitioning and provide actionable guidance for tuning payload sizes for secret images and copyright watermarks under different practical usage scenarios?
  2. Can the authors add more real‑world adversarial evaluations, such as cryptanalytic attacks against your SHA‑256‑based mask generation and imperfect diffusion inversion perturbations, to verify the robustness of the claimed plausible deniability and secret security guarantees?
  3. Would the authors clarify how the discrete semantic space design compares conceptually and empirically against existing semantic‑aware hiding works, and better highlight where their innovations diverge from prior incremental steganography‑watermark combination attempts?
2 comments
AuthorsOpenReview ↗

Per-Pool Budget, Error-Bounded Security, and Innovation Boundary of a Unified Dual-Key Steganography Framework

We appreciate the feedback and hope our response addresses the points.

W1&Q1: Per-Pool Budget. Disjoint mask partitioning imposes a per-pool budget constraint rather than a coupling constraint. The two payloads are allocated independently within their pools. With N=196608, ∣Ωsec∣=∣Ωwm∣=98304|\Omega_{sec}|=|\Omega_{wm}|=98304. Each stream needs LR=5120LR=5120 coordinates at default L=1024, R=5, (5.21%) of the pool. The pool ceiling is Cpool=98304×145≈275251bits.C_{pool}=\frac{98304\times14}{5}\approx275251 bits. A margin of about 19.2×19.2\times remains over the 14336-bit payload. The constraint binds only when LR→98304LR\to98304 (L≥19660L\ge19660 at R=5R=5), far beyond the present setting. The default bottleneck is thus L and K, not mask partition. Partitioning costs a constant factor of two in coordinate budget, about one bit per coordinate, and is tunable through ratio-γ\gamma partitionwith ∣Ωsec∣=γN|\Omega_{sec}|=\gamma N.

Recommended configurations are listed below, where utilization is computed as LR/∣Ωpool∣LR/|\Omega_{pool}|.

ScenarioLsecL_{sec}RsecR_{sec}LwmL_{wm}RwmR_{wm}Sec./Wm. Coord.Utilization
Balanced (default)10245102455120 / 51205.21%
Covert-comm. dominant40965256520480 / 128020.8%
Copyright dominant2565102451280 / 51205.21%
Robustness priority10249102499216 / 92169.38%
Pool ceiling (R=5R=5)19660519660598304 / 98304100%

W2&Q2: Error-Bounded Security. Security rests on bounded-error bounds and standardized primitives, not on exact diffusion inversion or custom cryptography. Theorem 1 incorporates inversion error into the TV bound dTV(P^,Q)≤Nhδtoken+εinv.d_{TV}(\hat{P},Q)\le N_h\delta_{token}+\varepsilon_{inv}.

Imperfect inversion is simulated by fewer steps and quantization noise on UniStega at 256×256256\times256 (VQGAN f8, K=16384, R=5).

Imperfect Inversion SettingRel. ErrorSRNet Acc. (%)KS pDeniability
Full steps1.0$\times$50.620.458Yes
70% steps1.4$\times$51.370.421Yes
50% steps2.05$\times$51.840.386Yes
50% steps + quantization noise2.6$\times$53.190.347Yes
25% steps + strong noise4.1$\times$56.830.212Yes

Mask generation uses SHA-256 and a cryptographically secure pseudorandom number generator, and tests on 100,000 keys show no exploitable bias.

Cryptanalytic TestMetricExpectedMeasured
Chi-square uniformityp-value0.50.472
Serial autocorrelationcorr. coeff.07.2×10−57.2\times10^{-5}
Birthday collisioncollision rate0.00520.0051
Shannon entropy per positionbitslog⁡2(98304)≈16.6\log_2(98304)\approx16.616.58

Under the strong adversary with full dual-key knowledge, four detectors remain close to the 50% random baseline.

DetectorPure (%)Stego (%)
SRNet50.3150.62
XuNet50.4750.96
YeNet50.5551.07
StegExpose51.2051.89

Q3: Innovation Boundary.

DimensionPrior WorksOurs
Semantic roleSemantic guidance for position selectionNative embedding in discrete semantic domain
Embedding levelPixel / low-level feature spaceHigh-level semantic token space
Integration paradigmPost-hoc combination of two separate modulesEndogenous unified framework
Capacity relationZero-sum trade-off with mutual interferenceSpatially decoupled with negligible mutual interference
Security modelSingle key, no deniabilityDual-key isolation with plausible deniability
Theoretical guaranteeNo unified theoretical foundationProvable distribution preservation & error bound
Differences from existing semantic-aware hiding works.

Conceptually. Prior methods use semantic to guide steganography and the secret has no semantic attribute. The proposed method encodes secret images and watermarks into discrete semantic tokens and conceals them in the high-level semantic domain, upgrading steganography from pixel values to visual semantic.

Empirically. Semantic-guided low-level concealing cannot avoid feature mutual interference. The proposed semantic-space embedding suppresses entanglement, improving recovery PSNR over semantic-guided baselines.

Differences from steganography-watermark combination attempts.

Different architectural essence. Prior schemes are post-hoc splicing of separate modules in pixel or feature space, inheriting zero-sum capacity trade-off, feature mutual interference, and single-key security boundary conflict. The proposed framework solves these challenges architecturally through spatial decoupling and dual keys.

Different security capabilities. Existing schemes lack steganographic deniability, and permission isolation is unattainable under a single key. Plausible deniability is realized for the first time in the proposed generative steganography framework by the dual-key mechanism.

Different theoretical support. Prior schemes are engineering splicing without unified guarantees. A complete proof system covering distribution preservation, error bound, and collision probability is provided in the proposed method.

Dear Reviewer vqbV,

Thank you again for your careful review and constructive suggestions. We sincerely appreciate the time and effort you have devoted to evaluating our work.

We have provided point-by-point responses, additional experiments, and clarifications in our rebuttal. As the discussion period draws to a close, we would be grateful if you could let us know whether these responses have addressed your concerns and which issues, if any, remain unresolved. We welcome further discussion and would be happy to provide timely clarifications before the deadline.

Thank you for your careful consideration and the effort you have invested in helping us improve our work.

Best regards,

Authors of Paper #13014

Reviewer_WuvrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes C²-Steg, the unified diffusion-based generative image steganography framework that simultaneously supports covert communication and copyright protection in a single pipeline. The key ideas are: (1) a Shared Discrete Semantic Space that tokenizes both the secret image and the watermark via a pretrained VQGAN and maps discrete tokens to Gaussian noise through randomized quantile mapping, thereby elevating steganography from pixel-level to semantic-level manipulation; and (2) a Disjoint Dual-Key Mask Decoupling mechanism that spatially partitions diffusion noise coordinates into non-overlapping secret and watermark regions, driven by separate true and fake keys. The dual-key design eliminates the zero-sum capacity trade-off and provides plausible deniability under coercion. Theoretical results include a TV-distance distribution-preservation bound and an error-probability analysis under majority-vote decoding. Experiments on UniStega, DIV2K, and Stega260 show improved stego quality, recovery fidelity, watermark verifiability, and anti-steganalysis performance over CROSS and DiffStega.

Critical strengths
  1. Novel and timely problem formulation. Unifying generative steganography and copyright watermarking within one diffusion pipeline, and explicitly addressing the plausible-deniability requirement, is a timely and well-motivated problem. The three-challenge decomposition (zero-sum capacity, semantic deficiency, conflicting security boundaries) is clearly articulated.
  2. The dual-key disjoint mask design is conceptually strong. Spatially partitioning noise coordinates into secret and watermark regions is an elegant way to break the mutual interference. The plausible-deniability argument (under coercion, revealing only the fake key recovers a high-fidelity watermark while the secret remains statistically unrecoverable) is a meaningful security contribution that existing generative steganography methods lack.
  3. Reasonable theoretical grounding. The randomized quantile mapping analysis (Proposition 1, Theorem 1) provides a TV-distance bound on how far the concealed noise distribution deviates from a true Gaussian, and the repetition-coding/majority-vote error-probability derivation gives a principled justification.
Critical weaknesses
  1. "Semantic" is overclaimed. The paper repeatedly frames the VQGAN codebook as a "high-level semantic domain," but VQGAN tokens are perceptual codebook entries learned from reconstruction, not semantic abstractions in the concept sense. The claim that this "eliminates feature mutual interference" is not directly verified.
  2. Watermark robustness is entirely absent. A copyright watermark must survive JPEG compression, cropping, scaling, noise, filtering, and potentially adversarial removal. The paper reports only PSNR/SSIM on the pristine stego image and anti-steganalysis detection rates, but never evaluates watermark recovery under any distortion. This is a fundamental gap for a method that claims "copyright protection." If the watermark vanishes after a single JPEG compression, the copyright-protection claim is unsupported.
  3. Very limited baseline comparison. Table 1 compares only CROSS, DiffStega, and DiffStego. Table 5 (anti-steganalysis) compares against cover-based steganography methods (UDH, HiNet, ISN) and two generative methods, but the copyright watermarking side has no baseline at all.
  4. Plausible deniability is not tested against a suspecting adversary. The deniability argument assumes the adversary only knows the fake key and cannot distinguish. But if an adversary suspects a dual-key scheme (and the paper itself publishes the design), can they run statistical tests on the stego image to detect that the odd-indexed and even-indexed noise coordinates have different distributions? An adversary who knows the scheme could test for this spatial non-stationarity. This security analysis is missing.
  5. Capacity is not reported. How many bits per image can be hidden? With L=1024 tokens and K=16384 codebook entries, the theoretical secret capacity is L⋅log⁡2(K)≈1024×14=14336L \cdot \log_2(K) \approx 1024 \times 14 = 14336 bits, but with R=5 repetition coding this drops to ~2867 bits. What is the actual secret image size (256×256 RGB? compressed?) and how does it compare to CROSS/DiffStega? No capacity comparison is provided.
  6. Computational cost is underreported. Table 1 reports inference time (10.57s on RTX 4090, 33.94s on RTX 3090) and memory (11.43 GB), but no training time. More importantly, the R=5 repetition factor means 5× more samples per token and majority voting. What is the actual decoding latency?
Questions for authors
  1. Watermark robustness: Can you report watermark recovery accuracy (Bit Error Rate or SSIM) after JPEG compression (quality 10–90), center cropping, Gaussian noise, and scaling? Without this, the "copyright protection" claim is difficult to evaluate.
  2. Deniability under adversarial suspicion: Given that the odd/even partition is public and the dual-key design is now published, can an adversary statistically test whether odd-indexed and even-indexed noise coordinates are i.i.d. Gaussian? Does the TV bound in Theorem 1 still hold when an adversary tests for spatial non-stationarity across the two masks?
  3. Capacity: What is the actual bit-rate of the hidden secret (bits per pixel)? How does it compare to CROSS and DiffStega under the same stego quality?
  4. Why VQGAN specifically? The paper uses VQGAN (Sun et al. 2024) with a 16384-entry codebook. What happens with a different VQGAN (e.g., VQGAN-f16, or a larger codebook)? Is the choice driven by reconstruction quality or by codebook size affecting the quantile-mapping interval width?
2 comments
AuthorsOpenReview ↗

Security, Robustness, Capacity, and Design Justification of the Unified Dual-Key Steganography–Watermarking Framework

Thank you for the feedback. We hope the clarifications could address your points.

W1: Semantics and Interference Scope. The term semantic refers to codebook-level visual semantics of VQGAN rather than concept-level abstraction, as defined in Sec. 3.2. The term discrete codebook-level semantic token space is uniformly adopted.

The interference-elimination claim is restricted to direct residual superposition and is guaranteed by the disjoint masks Msec∩Mwm=∅\mathbf{M}_{sec} \cap \mathbf{M}_{wm} = \emptyset with supporting evidence provided by Fig. 7, Fig. 8, and Tables 1-2.

W2&Q1: Watermark Robustness. Robustness is evaluated in Appendix E through recovery PSNR under JPEG compression, Gaussian noise, Gaussian blur, and Poisson noise. The observed gains are attributed to R=5R=5 repetition coding, majority voting, and error-tolerant token decoding. The same mechanism protects both the secret image and the watermark.

W3: Baseline Selection and Watermark Comparison. Baselines are task-matched. Generative steganography SOTA (CRoSS, DiffStega) is used in Table 1, while both cover-based and generative methods are included in Table 5. No direct copyright baseline was available, since prior methods lack the unified setting and rely on bit-string watermarks. Post-embedding methods are incompatible with generative steganography, and dedicated generative watermarking is single-task.

Four baselines are compared under 256×256256\times256 resolution and an identical backbone.

MethodWatermark PSNR ↑Watermark SSIM ↑Steganography CompatibleDeniability Supported
Tree-Ring19.410.63NoNo
Gaussian Shading19.720.64NoNo
MaXsive20.100.66NoNo
MaskWM20.350.67NoNo
Ours24.280.82YesYes

Higher watermark fidelity is achieved by the proposed method, and all baselines are single-task, whereas steganography and watermarking with deniability are supported by the proposed method.

W4&Q2: Spatial Non-Stationarity Tests. Under a uniform token prior, odd/even candidate-pool coordinates are i.i.d. standard normal, and the TV bound in Theorem 1 holds per subspace with value zero. The odd/even split only pre-partitions pools, and actual positions are sampled within each pool by a cryptographically secure pseudorandom number generator. Hence the pools cannot be distinguished from noise by any statistical test. For non-uniform priors, within-pool TV is bounded by LR⋅δtokenLR\cdot\delta_{token}, which is of the same order as the global bound, and detection is further diluted by hidden positions and background noise.

Statistical TestPure Generated ImagesStego Images
KS test p-value (odd vs. even)0.4720.458
Mean difference test p-value0.4610.449
Variance difference test p-value0.4830.467
Binary detection accuracy51.2%52.1%

All p-values exceed 0.05, and binary accuracy is close to random (0.5), which confirms deniability under spatial non-stationarity tests.

W5&Q3: Capacity and Bit-Rate. Payload is not reduced by repetition coding. Additional coordinates are used for error correction. The secret image uses L=1024L=1024 tokens from K=16384K=16384: Csec=Llog⁡2K=1024×14=14336bits.C_{sec}=L\log_2K=1024\times14=14336 bits. With R=5R=5, coordinates increase to LR=5120LR=5120, average information becomes 14/5=2.814/5=2.8 bits, and total payload is unchanged: bpp=14336256×256≈0.219.bpp=\frac{14336}{256\times256}\approx0.219. The watermark uses another LRLR coordinates and the same 14336 bits, giving 28672 bits total. CRoSS/DiffStega hide full images and the proposed method hides a 256×256256\times256 RGB secret at comparable quality and additionally recovers a watermark. All three methods are image-level generative, and higher secret fidelity at comparable capacity is achieved by the proposed method, which doubles the total information capability.

Image-in-image steganography capacity is commonly measured as the number of secret images concealed per stego image

W6: (1) No additional end-to-end training is required, since the framework is training-free. Public pretrained weights are used for VQGAN and BELM. (2) R=5R=5 decoding overhead is limited. Coordinate-wise majority over L=1024L=1024 groups of R=5R=5 requires only 5120 CDF/floor operations, all O(LR)O(LR). Dominant latency is diffusion inversion G−1G^{-1}, not voting. The increment over R=1R=1 is millisecond-order.

Q4: VQGAN Selection. VQGAN is chosen by a Pareto trade-off among reconstruction, capacity, and robustness. Codebook size KK affects capacity and quantile interval width C=Llog⁡2K.C=L\log_2K. Per-token error (Eq. 13) depends on bs±b_s ^ \pm. Under standard normal, z-domain width is ≈1/(Kϕ(z))\approx1/(K\phi(z)). Larger KK raises per-token information but narrows intervals, increases single-copy error, and weakens RR-fold repetition gain. Downsampling ff sets L:f=8⇒L=1024,f=16⇒L=256L: f=8\Rightarrow L=1024, f=16\Rightarrow L=256. Table 3 shows f=16f=16 is comparable to CRoSS, while f=8f=8 outperforms CRoSS/DiffStega, giving a favorable capacity–reconstruction balance.

Dear Reviewer Wuvr,

Thank you again for your insightful feedback on our submission. We have carefully addressed all of your concerns in the response document and believe we have fully clarified the points you raised.

Should you have any further questions or additional concerns, we are more than happy to assist you in any way we can. If any further experiments are required, please note that these would take a bit of additional time to conduct, but we will gladly proceed if you think they would add value.

We kindly hope that our explanations have helped to resolve any misunderstandings. We would be grateful for a fair and thorough evaluation of our revised work.

Thank you once again for your time and thoughtful consideration. We look forward to your response.

Sincerely,

Authors of Paper #13014

Comments on the paper

1 comment
Area_Chair_ijf2OpenReview ↗

Please Participate in the Discussion Phase

Dear Reviewers,

The authors have posted their rebuttal and responses addressing your review comments and questions.

Please carefully read their responses (as well as the updated manuscript, if applicable) and actively engage in the reviewer-author discussion phase. If you have any remaining concerns or follow-up questions, please post them so the authors can address them before the discussion period closes.

Your timely participation and updated assessments are essential to helping us reach a fair and well-informed recommendation for this submission.

Best regards,

Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.