acceptodds
Under review as a conference paper at ICLR 2027

LLMs as Linguistic Chameleons: Decoupling Semantics and Structure for Privacy-Preserving Communication

Abstract

As Large Language Model (LLM) APIs become increasingly integrated into privacy-sensitive workflows, ensuring inference-time privacy without compromising task utility remains a major challenge. Existing approaches preserve most of the original semantic content to maintain downstream performance, but this also leaves exploitable cues for reconstructing the original text. This work investigates semantic decoupling, which replaces original semantics with alternative content while preserving the structure needed for LLM reasoning. Based on this idea, we propose CROSS-MAP, a bidirectional framework that maps private inputs into a different semantic domain before inference and recovers the corresponding outputs afterward. Local models are trained with multi-objective optimization to maximize semantic divergence in the mapping stage while minimizing semantic inconsistency in the recovery stage. Experiments show that CROSS-MAP reduces reconstruction success across multiple attack settings while outperforming existing baselines in utility. Code is available at https://anonymous.4open.science/r/CROSS-MAP-7B2C/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models (LLMs), Privacy-Preserving Inference, Inference-Time Privacy, Text Sanitization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8KAGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper discusses user privacy in cloud-based LLM inference and proposes an information-theoretic approach. The method keeps semantic information locally and uses a mapper to replace the original prompt while preserving its structure. After the cloud LLM generates a response, a local recoverer reconstructs the final output. Experiments show a better privacy–utility trade-off than the baselines and robustness against various attacks.

Critical strengths

S1: Unlike previous semantics-preserving approaches, the authors separate text into structure and semantics and explain the effectiveness of their approach from an information-theoretic perspective.

S2: The method uses a learned cross-domain semantic mapping instead of direct content replacement to protect user privacy while preserving utility on certain tasks.

S3: The authors discuss multiple attack scenarios and show that the method provides some level of protection against different attacks.

Critical weaknesses

W1: The paper provides extensive information-theoretic proofs, but their connection to the actual mapper training and recoverer design is unclear.

W2: The method seems to require the LLM to reason correctly with preserved structure but replaced semantics, and the recoverer to reconstruct the final output correctly. This is a strong assumption. The method may fail on semantics-dependent tasks such as coding and math, but the authors do not evaluate such tasks.

W3: Table 2 uses three metrics defined in the paper, along with Downstream QA Utility and Recovered Text Utility, to demonstrate a better privacy–utility trade-off. Although the authors evaluate attack success rates in Section 6.2, evaluating privacy and downstream utility using attack success rates would provide more convincing evidence of the claimed trade-off.

W4: Table 2 uses Qwen2.5-14B as the base model for both the mapper and recoverer, which may require too much local computing power for typical API users. Moreover, Figure 4 in Appendix F.2 shows that performance strongly depends on the model sizes of the mapper and recoverer. This suggests that the method may be largely limited by the capabilities of its base models.

Questions for authors

Q1: Does the method work on semantics-dependent tasks such as coding and math?

Q2: If users can only run small models locally, would the method be limited by the base model capabilities and fail to achieve the performance reported in Table 2?

Q3: Since the method adds two local inference steps, have the authors measured the additional inference latency?

Reviewer_7ihfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

he paper proposes CROSS-MAP, which uses a local mapper to transform private inputs into a different semantic domain. The theme and the dictionary mapping original terms to their replacements remain local, while only the rewritten text is sent to an external LLM. A local recoverer then reconstructs the answer in the original domain. Training uses synthetic examples, supervised fine-tuning (SFT), and multi-objective direct preference optimization (DPO) of the mapper, balancing dictionary coverage, semantic distance, fluency, and recovery consistency. The paper also uses the information bottleneck framework to characterize when semantic decoupling is beneficial, and evaluates attribute inference, black-box span restoration, and white-box DPO attacks with dictionary leakage. Across four QA benchmarks, CROSS-MAP achieves both greater semantic distance between the original and mapped text and higher ACC than existing methods like HAS. The paper provides extensive non-QA generalization experiments and ablations of training, dictionary properties, and scale, but the main results come from single runs.

Critical strengths
  1. Cross-domain mapping provides a meaningful extension of existing input-transformation and output-recovery approaches. CROSS-MAP combines domain-level semantic rewriting, an explicit mapping dictionary, and local recovery to explore whether substantial changes to input semantics can preserve downstream utility. Although the general protection pipeline and trainable recovery modules have precedents, this particular design offers a useful direction for reducing the semantic cues exposed to external LLMs.

  2. The multi-objective training design explicitly addresses competing requirements of the defense. The scoring function combines dictionary coverage, dictionary- and text-level semantic distance, fluency, and recovery consistency to construct preferences for DPO. This is an interesting mechanism for balancing semantic transformation against recoverability. The SFT-only and DPO comparisons in Table 15 provide evidence that training can adjust this trade-off, increasing semantic distance and format validity while incurring some loss of task utility.

Critical weaknesses
  1. Important experimental and training details are insufficiently documented, limiting reproducibility and interpretation. Although the appendices contain substantial additional experiments, the fixed settings, changed factors, and relationships among configurations are not consistently specified. In particular, the synthetic-data generation, filtering, and splitting procedures are unclear, as are the reward weights and key preference-pair construction settings. A specific section of experimental settings and a reproducible description of the SFT–DPO data pipeline are needed.

  2. The evaluation does not sufficiently establish the practical privacy–utility benefits of the proposed system. The main experiments use Qwen2.5-14B for both local modules but an LLAMA-8B model for external inference, leaving the deployment rationale unclear for users already capable of running a larger model locally.

  3. Inadequate privacy measurement. Table 2 measures privacy primarily through coverage and semantic-distance proxies, even though semantic distance is itself optimized during training. Section 6.2 provides useful attack evaluations, but their differing datasets, baseline coverage, and metrics make them difficult to connect directly to the main table(Table 2). Table 15 also compares SFT-only and DPO checkpoints without directly evaluating their relative resistance to attacks. Matched utility–attack comparisons across training stages, together with reward-component ablations and deployment-cost measurements, are needed to establish the practical contribution of DPO and the overall defense.

Questions for authors
  1. Training and data construction: Could you specify more data generation details? For example, how the synthetic SFT data were generated? For DPO, what were the candidate sampling settings, preference-pair selection threshold, number of resulting pairs, and five reward weights?

  2. Experimental configurations: Could you clarify which settings are held fixed and which change between Table 2 and the appendix experiments, particularly Tables 12 and 15? Please identify the exact model checkpoints and explain how the reported training checkpoints were selected.

  3. Deployment rationale: What motivates using Qwen2.5-14B for both local modules while outsourcing inference to LLAMA-8B in the main experiments?

  4. Privacy–utility comparison: Could you report attack success alongside task utility for CROSS-MAP and the principal baselines on matched evaluation data, with consistent attacker capabilities and budgets? This would help determine whether the semantic-distance improvements in Table 2 correspond to lower actual leakage.

Reviewer_pQupOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper provides a mechanism for privacy during API based inference from LLM's by semantically replacing tokens in the text in a structured manner.

Critical strengths

1.) It is a good attempt at using information reducing ideas, to reduce privacy leakage. Similar ideas have been previously explored in the information bottleneck method, and other works in rate-distortion tradeoffs in communication and in other works in DP such as https://arxiv.org/abs/1608.03677 .

2.) A reasonable set of attacks were considered in evaluating the mechanism, including de-anonymization attacks, black box span restoration attacks and white box attacks.

Critical weaknesses

1.) Although the overall guiding principle is reasonable, it is not clear how these privacy guarantees compare to differential privacy or to other setups.
2.) Different mechanisms such as https://openreview.net/forum?id=WLK37mn0El were introduced for privacy during LLM inference. how does your work compare with such methods? 3.) methods for reducing the information bottleneck were attacked under some assumptions in works like https://arxiv.org/pdf/2012.02670 before.

Questions for authors

information bottleneck method for deep learning in general has been critically reviewed in https://artemyk.github.io/assets/pdf/papers/Kolchinsky%20et%20al_2019_Caveats%20for%20information%20bottleneck%20in%20deterministic%20scenarios.pdf do these criticisms apply to your setting?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.