acceptodds
Under review as a conference paper at ICLR 2027

PACER: Provably Secure Steganography under Asymmetric Receiver Access with Certified Recovery

Abstract

Many high-capacity provably secure steganography (PSS) schemes for large language models require the receiver to reproduce the sender's token distribution using the same model and prompt. In practice, privacy or resource constraints may prevent the receiver from accessing the sender's prompt or model, creating asymmetric receiver access. Existing asymmetric baselines use token-only recovery to avoid this dependence, but often have low embedding rates or limited recovery reliability and do not benefit from receiver-side resources that may remain available. We present PACER, a provably secure linguistic steganography framework for asymmetric receiver access with certified recovery. PACER separates the sender's token distribution used during generation from the receiver-side information used for recovery. When a receiver-side distribution is available, PACER combines maximal coupling with token-class coupling to extract recovery evidence while preserving the sender's token distribution. Otherwise, token-class coupling alone enables token-only recovery. An e-process accumulates this evidence and certifies payload blocks under a prescribed error bound. We prove PACER's computational security and show that the probability of incorrect recovery in any certification process is at most this bound. Experiments show that PACER achieves reliable recovery across different receiver access modes and that receiver-side distributions increase the embedding rate over token-only recovery. With receiver-side information, PACER achieves more than 42 times the embedding rate of the reliable asymmetric baseline. In token-only mode, PACER achieves more than 16 times that rate, with faster encoding and decoding.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, provably secure steganography, asymmetric receiver access, certified recovery

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Rb2KOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PACER studies linguistic steganography under asymmetric receiver access, where the receiver may lack the sender’s prompt, model, or both. It separates distribution-preserving generation from the evidence used for recovery. Maximal coupling exploits an available receiver-side distribution, while token-class coupling extracts additional evidence from residual probability mass and supports token-only decoding. An e-process accumulates evidence for candidate payload blocks and certifies a block upon a unique threshold crossing. The paper establishes distribution preservation and a per-certification error bound under idealized randomness and exact sampling and arithmetic, and reports substantial embedding-rate improvements across three model families and four datasets.

Critical strengths
  1. A clear problem and a well motivated construction.The receiver’s inability to reproduce the sender’s distribution is a meaningful limitation of high-capacity steganography. PACER addresses this limitation by exploiting information that remains available rather than requiring exact distribution reconstruction. The combination of maximal coupling, residual token-class coupling, and sequential certification is a creative design: it uses overlapping probability mass for proposal evidence and the remaining mass for class evidence while preserving the generation distribution. The same framework also supports receivers without a model.

  2. Substantial gains under matched receiver resources. The token-only comparison is particularly convincing because the improvement does not depend on additional receiver-side model access. On Llama3/C4, PACER-TO achieves 213.7 bits per 1,000 tokens compared with 13.2 for PRSSA, with zero observed BER for both. Table 3 further supports the proposed mechanism: residual token-class coupling generally provides larger gains when sender and receiver distributions overlap less.

Critical weaknesses
  1. The full recovery comparison needs a clearer evaluation protocol. Section 4 excludes unrecovered suffixes from the BER denominator. This measures the correctness of recovered bits, but does not by itself establish delivery of a prespecified message. In Table 5, the watermarking baselines use fixed payloads of 4, 16, and 64 bits, whereas the corresponding prespecified payload length for PACER is not equally clear. Consequently, it is difficult to determine whether PACER’s reported 100% full recovery uses the same success criterion. Please specify its payload length and the denominator of this metric, and report full recovery, noncompletion, and incorrect output under matched payload lengths and token budgets. This concern affects the interpretation of Table 5 rather than directly invalidating the main embedding-rate gains.

  2. Retokenization results need a clearer connection to the protocol.Table 4 reports a modest average rate reduction and no observed bit errors after retokenization, which is encouraging. However, retokenization can change token positions and certification times, while the sender advances between payload blocks by reproducing receiver certification decisions. The paper should explain how block alignment is maintained or recovered, and what the reported sender synchronization operation actually does. The proof of Corollary 3.1 should also explicitly connect its independence argument to this end-to-end decoding procedure. This is a request to clarify a potentially important gap in exposition and justification, not a claim that the corollary is necessarily incorrect.

Questions for authors
  1. What is PACER’s prespecified payload length in Table 5? Does 100% full recovery require recovery of the entire prespecified message? Can you provide a matched comparison using a 64-bit payload and a 300-token budget?

  2. In a fixed-payload experiment, how often does recovery remain incomplete because of insufficient evidence, and how are those samples included in the reported metrics?

  3. Please describe sender synchronization under retokenization and explain how the protocol handles differences in certification times and payload-block boundaries between the endpoints.

2 comments

Thank you for describing PACER’s construction as “a creative design”, recognizing its “Substantial gains under matched receiver resources”, and finding the token-only comparison “particularly convincing”. Below, we add an evaluation of full recovery and explain synchronization and its connection to Corollary 3.1.

Rebuttal to Weakness 1 & Questions 1–2: Full Recovery and Incomplete Messages

Our original experiments mainly evaluated PACER’s embedding rate and bit error rate. The sender continuously embedded a secret bitstream during generation. We measured how many bits could be recovered within a given text length and whether those bits were correct, without specifying a fixed message length that had to be delivered in full. PACER’s 100% recovery in the original Table 5 corresponds to this setting.

You suggested distinguishing the correctness of recovered bits from full recovery of a prespecified message, and using matched payload lengths and token budgets. Following this suggestion, we completed a comparison using a fixed payload of 64 bits and a budget of 300 tokens on Llama-3.1-8B/C4.

The comparison includes all four PACER modes, PRSSA, Disreo, USLM, and the two watermarking methods MPAC and BiMark.

Full recovery requires all 64 bits to be output correctly. Any unrecovered suffix counts as a failure to recover the complete message. Each sample is classified as full correct recovery, incomplete recovery without errors, or incorrect output. The results are shown below.

MethodFull correct recovery ↑Incomplete without errors ↓Incorrect output ↓PPL ↓
PACER-PD97.4%2.6%0%6.80
PACER-MD95.7%4.3%0%6.76
PACER-DD90.8%9.2%0%6.74
PACER-TO71.7%28.3%0%6.80
PRSSA0%100.0%0%6.71
Disreo0%15.3%84.7%6.77
USLM0%60.3%39.7%6.85
MPAC (64-bit)18.1%0%81.9%14.65
BiMark (64-bit)1.3%0%98.7%7.11

These results directly answer your question about incomplete messages. For PACER-PD, MD, DD, and TO, respectively, 2.6%, 4.3%, 9.2%, and 28.3% of samples do not complete all certifications within the budget. All of these samples count as failures to recover the full message. No incorrect outputs were observed in any of the four modes.

PACER retains a clear advantage in the evaluation with a fixed message length. PACER-PD achieves 97.4% full correct recovery. PACER-TO, which you highlighted in your review, achieves 71.7% without any receiver-side model. In comparison, PRSSA, Disreo, and USLM never recover the full message correctly within this budget. MPAC and BiMark achieve full recovery rates of 18.1% and 1.3%, respectively. MPAC also has a PPL of 14.65, indicating a larger loss in text quality. These results further show PACER’s advantages in full recovery and text quality.

The new experiment further supports PACER’s performance in recovering messages of a fixed length and complements the original results on embedding rate that you recognized.

Rebuttal to Weakness 2 & Question 3: Sender Synchronization and Corollary 3.1

Thank you for your suggestions on synchronization and Corollary 3.1. We would like to clarify that sender synchronization in PACER confirms the payload prefix ultimately certified by the receiver. Certification positions and payload-block boundaries after retokenization may differ from those during generation. The sender replays the receiver’s decoder on the final text to determine which blocks have been recovered and which remain to be transmitted. Below, we illustrate this process and explain its connection to Corollary 3.1.

During generation, the sender runs the receiver’s certification algorithm alongside embedding and uses its decisions to determine when to embed the next block. If receiver-side retokenization produces the same token sequence, both endpoints make the same certification decisions. If the sequence changes, the sender reruns the receiver’s decoder on the final text to confirm the actual recovery progress. Sender Synchronization in Table 4 reports the cost of this additional check.

Replay itself is a simple operation. For many steganographic methods that rely on exact token sequences for decoding, retokenization can produce incorrect bits or cause later decoding states to become misaligned. Even if the sender detects these errors through replay, the receiver may not be able to determine on its own which information it can reliably accept.

Replay is effective because PACER certifies and recovers the payload block by block. The receiver outputs the current block and advances to the next one only after a unique certification. Blocks with insufficient evidence remain unrecovered. This lets the receiver determine its recovery progress while controlling the probability of incorrect output. The sender then confirms this progress through replay.

Consider PACER-TO sample 24. The sender locally certifies three 16-bit payload blocks at positions 94, 192, and 224 in the generated token sequence, for a total of 48 bits. It starts embedding the next block after each certification.

When the receiver retokenizes the final text, the first difference occurs in the text corresponding to position 172 in the generated token sequence. The original sequence splits saramar into two tokens, sar and amar. Retokenization instead gives three tokens, s, aram, and ar. The text is unchanged, but the changed token sequence affects the subsequent accumulation of certification evidence.

The first block has already been certified at position 94, before this change. The receiver outputs these 16 bits and enters the second block. Even after the sender has started embedding the third block during generation, the receiver continues certification using its state for the second block. In this sample, the second block does not accumulate enough evidence before the text ends, so the receiver produces no further output. Replaying the receiver’s decoder on the same final text also recovers only the first 16-bit block, confirming that the second and later blocks still need to be transmitted.

Incorrect recovery occurs when an incorrect candidate happens to accumulate enough evidence, is the only candidate to reach the threshold, and is output. Corollary 3.1 uses the independence between the random values for an incorrect candidate and the generated text to extend certification error control to the deterministically retokenized sequence. The receiver computes the wrong-candidate baseline and updates the evidence from the retokenized history, so the analysis applies to the decoding process that the receiver actually runs. Under the theoretical assumptions in the paper, the probability of outputting an incorrect block in each certification process remains at most ε\varepsilon. Its default value is 10−610^{-6}, or one in a million. This guarantee allows the receiver’s certification positions to differ from those during generation.

Neither the retokenization experiment in the original Table 4 nor the new fixed 64-bit experiment produced any observed incorrect PACER outputs. In addition, independent replay checks on 4,000 texts across the four modes found complete agreement between sender replay and the receiver’s final certification results.

Changes to the paper

We have updated the original Table 5, now Table 6 in the revised paper, with the new fixed 64-bit results and clarified how incomplete samples are counted. We have also clarified synchronization under retokenization and its connection to the certification error guarantee in Section 3.2 and Appendix C.2.

Thank you again for your constructive suggestions.

Reviewer_hpN1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies provably secure steganography under asymmetric receiver access. In many practical settings, privacy or computational constraints prevent sharing the sender’s model and prompt. However,existing asymmetric receiver-access (ARA) methods rely on token-only recovery and are limited in both embedding rate and reliability.

To this end, this work presents PACER, a provably‑secure linguistic steganography framework for asymmetric receiver access. It uses two coupling mechanisms for distribution‑preserving embedding and n e‑process for certified payload recovery, greatly boosting embedding rate without requiring receivers to replicate the sender’s model or prompt.

Experiments are conducted across four datasets and three model families under diverse receiver‑access modes, evaluating embedding rate, recovery reliability, and steganalysis performance against existing asymmetric baselines.

Critical strengths
  1. This work studies a very interesting and valuable problem, i.e., asymmetric‑receiver gaps in existing provably secure steganography.

  2. This work has formally defined four clear practical receiver modes in the asymmetric receiver‑access steganography problem.

  3. A novel method that combines maximal coupling and token‑class coupling to preserve the sender’s token distribution under varied resources.

  4. Experiments have covered four datasets, three model families, and four receiver modes. The following results have shown higher embedding rates than baselines, while maintaining near‑random steganalysis detectability.

  5. Complete proofs are provided.

Critical weaknesses
  1. PACER's computational overhead is not well in the four PACER variants.

  2. PACER treats four receiver modes as discrete cases; no adaptive mechanism for the gradual degradation of receiver‑side distribution quality.

  3. Text quality is simply measured by the perplexity. It is better to conduct a human annotation study.

Questions for authors
  1. Could you provide simple downstream‑task results to show the usability of stego‑text?
5 comments

Thank you for recognizing the “very interesting and valuable problem” and the four practical receiver access modes, as well as the method’s novelty, complete proofs, and experimental performance. Below, we address computational overhead and add experiments on degradation of the receiver distribution and downstream tasks.

Rebuttal to Weakness 1: Computational Overhead

PACER’s computational overhead can be assessed through generation speed, encoding speed, and decoding speed. Table 1 in the original paper already reports these metrics across three model families. The Llama3/C4 results are shown below.

MethodGeneration speed (tokens/s)Encoding speed (bit/s)Decoding speed (bit/s)
Random Sampling31.11——
PACER-PD19.5111.2917.30
PACER-MD23.9910.9821.61
PACER-DD23.578.4416.91
PACER-TO27.445.8689.02
PRSSA25.760.3438.84

The modes offer clear trade-offs between resources and efficiency. PD uses a receiver-side model of the same size as the sender’s model and therefore has a higher generation cost. MD and DD use a smaller receiver-side model, while TO requires no receiver-side model. In this setting, MD, DD, and TO increase the generation time for texts of the same length by about 29.7%, 32.0%, and 13.4%, respectively, compared with Random Sampling. TO generates 27.44 tokens/s, close to Random Sampling at 31.11 tokens/s.

More importantly, PACER’s higher embedding rate brings clear advantages in encoding and decoding speed. The four modes achieve encoding speeds of 5.86–11.29 bit/s, or 17.2–33.2 times that of PRSSA, the reliable asymmetric baseline. PACER-TO requires no receiver-side model and decodes at 89.02 bit/s, about 2.3 times PRSSA’s speed.

Thus, PACER achieves higher encoding efficiency while adding some generation cost. The four modes also offer different resource choices. PD, MD, and DD use receiver-side resources to increase the embedding rate, while TO provides fast recovery without a receiver-side model.

Rebuttal to Weakness 2: Adaptation to Gradually Degraded Receiver Distributions

Thank you for your suggestion on changes in the quality of the receiver distribution. We would like to clarify that PACER’s four modes describe the model and prompt resources available to the receiver. Even within one mode, the overlap between the receiver distribution QQ and sender distribution PP can change continuously during generation.

PACER’s maximal coupling and residual token-class coupling already adapt naturally to these changes. When PP and QQ are close, maximal coupling accepts more proposals. As overlap decreases, more samples enter residual token-class coupling to provide recovery evidence. Both use the distributions at each step, without predefined quality levels or manual adjustment of coupling weights.

Original Table 3 varies receiver model sizes and prompts, with mean overlap of 0.619–0.828. Residual token-class coupling improves embedding rate by 6.39%–43.85%, generally with larger gains at lower overlap.

To test adaptation to continuous changes in QQ, we added PACER-MD/DD experiments on Llama-3.1-8B/C4 with Llama-3.2-1B as the receiver model. MD retains the original prompt, while DD omits it.

Keeping PP unchanged, we mix the original receiver distribution with the uniform distribution over the vocabulary, using λ\lambda to control degradation.

Qt(λ)(v)=λtQtbase(v)+(1−λt)1∣V∣Q_t^{(\lambda)}(v)=\lambda_t Q_t^{\mathrm{base}}(v)+(1-\lambda_t)\frac{1}{|V|}.

Here, tt is the generation step, vv a token in vocabulary VV, and ∣V∣|V| its size. Qtbase(v)Q_t^{\mathrm{base}}(v) and Qt(λ)(v)Q_t^{(\lambda)}(v) are the receiver’s token probabilities before and after mixing. The weight λt∈[0,1]\lambda_t\in[0,1] sets the original distribution’s share, while 1/∣V∣1/|V| assigns equal probability to all tokens.

At λ=1\lambda=1, QQ is the original receiver distribution. At λ=0\lambda=0, it is uniform, with no model prediction preferences. Decreasing λ\lambda reduces predictive information, allowing us to test degradation without changing the receiver model.

We use two experimental settings.

  1. Static degradation. We hold λ\lambda at 1, 0.75, 0.5, 0.25, or 0 throughout each text to compare embedding rate and recovery reliability across distribution quality levels.
  2. Gradual degradation. Within each 300-token text, λ\lambda decreases linearly from 1 to 0, simulating a gradual loss of predictive information. It follows the token position and continues across certification boundaries.

Both endpoints use the same mixture and λ\lambda schedule. We test 1,000 texts per configuration, totaling 12,000, with continuous embedding and the original 300-token budget and certification parameters. Only QQ is degraded. Encoding and certification algorithms remain unchanged.

The results are shown below. Embedding rates are in bits per 1,000 tokens.

Distribution settingMD mean overlapMD embedding rateDD mean overlapDD embedding rate
λ=1\lambda = 10.7557457.50.6886357.9
λ=0.75\lambda = 0.750.6186426.00.5600375.1
λ=0.50\lambda = 0.500.4375373.90.4024331.6
λ=0.25\lambda = 0.250.2298295.80.2143278.6
λ=0.00\lambda = 0.000.0004207.50.0004213.5
Linear degradation within each text0.4170308.90.3641277.0

The results show PACER’s continuous response as the receiver distribution degrades.

As overlap decreases in the static experiments, MD’s rate falls from 457.5 to 207.5 bits per 1,000 tokens. DD starts at 357.9 and still achieves 213.5 bits per 1,000 tokens at very low overlap. Token-class coupling supports recovery even at near-zero overlap, when maximal coupling contributes little.

When the distribution gradually degrades within the same text, MD and DD achieve 308.9 and 277.0 bits per 1,000 tokens, respectively. This shows that the existing mechanism can recover information even when the receiver distribution keeps changing during generation, without switching receiver access modes. No errors in recovered bits were observed under any tested condition.

The four modes describe resource access, while the two coupling mechanisms adapt to changing distributions. These experiments support recovery as the receiver’s predictive information decreases.

Rebuttal to Weakness 3 & Question 1: Downstream Task Usability and Text Quality

Thank you for your suggestions on text quality and practical usability. In addition to the independent PPL and steganalysis evaluations in the original paper, we added two practical downstream tasks to directly assess the quality of text generated by PACER.

We use CNN/DailyMail summarization and WMT16 translation from English to Romanian, evaluated with ROUGE-L and SacreBLEU, respectively. Each task compares Random Sampling with all four PACER modes, using 1,000 texts per condition and 10,000 outputs in total. All methods use the same inputs and rules for generating task outputs.

The results are shown below.

MethodSummarization ROUGE-L ×100\times 100 ↑Translation SacreBLEU ↑
Random Sampling24.74522.146
PACER-PD24.76421.643
PACER-MD24.62222.138
PACER-DD24.79521.592
PACER-TO24.88021.877

All four PACER modes perform close to Random Sampling on both tasks.

For summarization, Random Sampling achieves a ROUGE-L score of 24.745, while the four PACER modes score 24.622–24.880. For translation, Random Sampling achieves a SacreBLEU score of 22.146, while PACER scores 21.592–22.138.

The new results show that text generated by PACER maintains downstream task performance close to Random Sampling. This provides further evidence from practical tasks to complement the original PPL and steganalysis results.

Human evaluation is a valuable suggestion, and we will add it in future work when conditions allow.

Changes to the paper

We have added summarization and machine translation quality results in Table 3 of the main text and experiments on degradation of the receiver distribution in Appendix H.2.

Thank you again for your constructive suggestions.

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

The paper focus on asymmetrical communication in generative text steganography where the receiver lacks access to the sender's model, prompt, or both. It proposes PACER, which separates preservation of the sender's generation distribution from payload recovery. When the receiver's distribution is available, PACER combines maximal coupling and token-class coupling to construct recovery evidence; otherwise, token-class coupling enables model-free extraction. An e-process accumulates evidence for payload candidates and outputs a block upon certification. Under the idealized assumptions, the authors establish distribution preservation, computational security, and an error bound for each certification process. A large number of experiments demonstrate substantial rate improvements over the evaluated asymmetric baselines.

Critical strengths
  1. The separation of distribution preservation from recovery evidence provides a unified and practically meaningful framework for asymmetric text steganography. PACER does not require the receiver to reconstruct the sender’s distribution exactly. Instead, it exploits available receiver-side information to improve recovery efficiency while retaining token-only recovery when no auxiliary model is available. Maximal coupling uses the overlap between the two distributions, and token-class coupling exploits the residual probability mass. The decomposition into accepted and residual contributions makes the distribution-preservation argument transparent. This design accommodates several resource constraints within one framework, and the ablations support the particular value of residual coupling when distribution overlap is smaller.
  2. The method combines adaptive decoding with explicit error control and substantial empirical rate improvements. The e-process construction permits sequential evidence accumulation and adaptive stopping while controlling the probability of an incorrect output during each certification process. This provides a more informative reliability characterization than average BER alone. The evaluation spans multiple datasets and model families: in Table 1, PACER-PD exceeds PRSSA’s recovery rate by more than 42 times, while the token-only mode exceeds it by more than 16 times, with no observed errors among recovered bits.
Critical weaknesses
  1. The full-recovery evaluation is insufficiently specified. BER excludes unrecovered suffixes, while the reported 100% full-recovery rate does not clearly specify PACER’s prespecified payload length. This limits assessment of complete message delivery within a finite text budget. The evaluation should use matched payload lengths and token budgets and count incomplete recovery as a whole-message failure.
  2. Block synchronization under retokenization is underspecified. The protocol advances blocks through shared certification decisions, but retokenization can change candidate paths and certification boundaries. Although synchronization time is reported, the coordination of subsequent embedding is not clearly described. Synchronization pseudocode and a worked boundary-change example would make the text-transmission procedure verifiable.
Questions for authors
  1. Evaluation clarification: Is PACER’s payload in the full-recovery comparison fixed before generation? Please explain how uncertified suffixes are handled and report whole-message success for all methods with a 64-bit payload and a 300-token budget, counting incomplete recovery as failure.
  2. Protocol clarification: When retokenization causes the receiver’s certification boundaries to differ from the sender’s, how is the next payload block selected for embedding? Please provide concise pseudocode or a worked example specifying block-index and evidence-state updates.
2 comments

Thank you for recognizing PACER as “a unified and practically meaningful framework” and highlighting its “adaptive decoding with explicit error control and substantial empirical rate improvements”. Below, we add a comparison of full recovery with a fixed message length and explain synchronization under retokenization.

Rebuttal to Weakness 1 & Question 1: Full Recovery Evaluation

Our original experiments mainly evaluated PACER’s embedding rate and bit error rate. The sender continuously embedded a secret bitstream during generation. We measured how many bits could be recovered within a given text length and whether those bits were correct, without specifying a fixed message length that had to be delivered in full. PACER’s 100% full recovery in the original Table 5 corresponds to this setting with continuous embedding.

You suggested using matched payload lengths and token budgets to directly compare recovery of complete messages. Following this suggestion, we completed a comparison using a fixed payload of 64 bits and a budget of 300 tokens on Llama-3.1-8B/C4. We compared all four PACER modes with PRSSA, Disreo, USLM, and the two watermarking methods MPAC and BiMark. A sample counts as successful only if all 64 bits are recovered correctly. Every incomplete sample counts as a failure to recover the complete message. The results are shown below.

MethodFull correct recovery ↑Incomplete without errors ↓Incorrect output ↓PPL ↓
PACER-PD97.4%2.6%0%6.80
PACER-MD95.7%4.3%0%6.76
PACER-DD90.8%9.2%0%6.74
PACER-TO71.7%28.3%0%6.80
PRSSA0%100.0%0%6.71
Disreo0%15.3%84.7%6.77
USLM0%60.3%39.7%6.85
MPAC18.1%0%81.9%14.65
BiMark1.3%0%98.7%7.11

PACER retains a clear advantage under a fixed payload. PACER-PD achieves 97.4% full correct recovery, while PACER-TO, which requires no receiver-side model, achieves 71.7%. In comparison, PRSSA, Disreo, and USLM never recover the full 64-bit message correctly within this budget. MPAC and BiMark achieve full recovery rates of 18.1% and 1.3%, respectively. MPAC also has a PPL of 14.65, indicating a larger loss in text quality. These results further show PACER’s advantages in full recovery and text quality.

The new experiment further supports PACER’s ability to recover complete messages under a fixed payload and a limited token budget, complementing the original results on embedding rate.

Rebuttal to Weakness 2 & Question 2: Block Synchronization under Retokenization

Thank you for your suggestions on synchronization. We would like to clarify that the certification boundaries after retokenization do not need to match those during generation. Sender synchronization here means reproducing the receiver’s decoding results on the final text to confirm which blocks have been recovered and which unrecovered block should be transmitted next.

During generation, the sender runs the receiver’s certification algorithm alongside embedding and uses its decisions to determine when to embed the next block. If receiver-side retokenization produces the same token sequence, both endpoints make the same certification decisions. If the sequence changes, the sender reruns the receiver’s decoder on the final text to confirm the actual recovery progress. Sender Synchronization in Table 4 reports the cost of this additional check.

Replay itself is a simple operation. For many steganographic methods that rely on exact token sequences for decoding, retokenization can produce incorrect bits or cause later decoding states to become misaligned. Even if the sender detects these errors through replay, the receiver may not be able to determine on its own which information it can reliably accept.

Replay is effective because PACER certifies and recovers the payload block by block. The receiver outputs the current block and advances to the next one only after a unique certification. Blocks with insufficient evidence remain unrecovered. This lets the receiver determine its recovery progress while controlling the probability of incorrect output. The sender then confirms this progress through replay.

Consider PACER-TO sample 24. The sender locally certifies three 16-bit payload blocks at positions 94, 192, and 224 in the generated token sequence, for a total of 48 bits. It starts embedding the next block after each certification.

When the receiver retokenizes the final text, the first difference occurs in the text corresponding to position 172 in the generated token sequence. The original sequence splits saramar into two tokens, sar and amar. Retokenization instead gives three tokens, s, aram, and ar. The text is unchanged, but the changed token sequence affects the subsequent accumulation of certification evidence.

The first block has already been certified at position 94, before this change. After outputting these 16 bits, the receiver advances its block index to 2, resets the evidence, and carries forward the first block’s cumulative path counts according to Section 3.2. The second block never accumulates enough evidence, so the receiver stays at that block and produces no further output before the text ends. Replaying the receiver’s decoder on the same final text also recovers only the first 16-bit block. The sender thus confirms that subsequent transmission should start from the second block.

Incorrect recovery occurs when an incorrect candidate happens to accumulate enough evidence, is the only candidate to reach the threshold, and is output. Corollary 3.1 proves that, under the theoretical assumptions in the paper, the probability of outputting an incorrect block in each certification process after retokenization remains at most the prescribed ε\varepsilon. Its default value is 10−610^{-6}, or one in a million.

Neither the retokenization experiment in the original Table 4 nor the new fixed 64-bit experiment produced any observed incorrect PACER outputs. In addition, independent replay checks on 4,000 texts across the four modes found complete agreement between sender replay and the receiver’s final certification results.

Changes to the paper

We have updated the original Table 5, now Table 6 in the revised paper, with the new fixed 64-bit results and clarified how full recovery is measured. We have also clarified synchronization under retokenization in Section 3.2 and added an example of how the receiver advances through payload blocks in Appendix H.1.

Thank you again for your suggestions.

Comments on the paper

1 comment
Area_Chair_s7E3OpenReview ↗

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.