acceptodds
Withdrawn from ICLR 2027

SURPRISAL-AWARE PREPROCESSING FOR LLM CLASSIFICATION OVER ENCRYPTED TEXT

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Fine-tuning large language models on sensitive data is challenging in security-critical applications, where plaintext exposure may be unacceptable and encryption is therefore essential. However, encrypted representations can substantially reduce model utility by disrupting the lexical, syntactic, and distributional regularities learned during pretraining. We address this problem with a surprisal-aware preprocessing framework that transforms selected plaintext spans before protection to improve the learnability of the resulting protected representations without modifying the downstream classifier or protection mechanism. The framework preserves task-relevant and structural content, generates context-compatible substitutions using a frozen language model, and selects replacements through pairwise ordinal aggregation over sentence-level semantic similarity, token-level contextual similarity, contextual surprisal, and part-of-speech compatibility. Across eight cross-domain classification datasets, preprocessing improves accuracy under all evaluated protection configurations. Under the primary S-FF3-1 setting, macro-average accuracy increases from 0.8230 to 0.8794, with positive point-estimate gains on all eight datasets; standard FF3-1 and AES-CBC show similarly consistent improvements. Controlled rewriting baselines and criterion ablations further support the contribution of the proposed selection strategy. Because encrypted representations may still leak information through observable sequence length, we additionally introduce a randomized padding extension that weakens the relationship between input length and protected-sequence length while preserving classifier utility. Overall, the results show that representation-aware preprocessing can improve the practicality of encrypted-data model training, while randomized padding provides complementary protection against length-based information leakage.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text Preprocessing, Privacy-Preserving NLP, Encrypted Text, Format-Preserving Encryption, Large Language Models, Text Classification, Surprisal-Aware Preprocessing, Secure Machine Learning

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_B3c8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper looks at fine-tuning an LLM on text that has been encrypted before it leaves the data owner. The authors argue that encrypted text is hard for a pretrained model to learn from, and that this can be partly repaired by rewriting the plaintext before encryption. Eligible word spans are replaced with candidates proposed by a frozen Qwen3-4B-Instruct. The replacement is picked by a pairwise ordinal vote (Copeland/Borda style, with tolerance-based ties) over four criteria: sentence-embedding similarity, token-level contextual similarity, length-normalised surprisal of the candidate, and POS match. Entities, numbers, negations, high self-information words and similar spans are excluded from editing, and an edit budget caps the number of replacements.

Critical strengths
  1. The setup isolates the variable of interest. Key, tweak policy, data split, initialisation and training recipe are held fixed across paired runs, and only the pre-encryption rewrite changes, for three different ciphers. That is the right design for this question. The preprocessing is also cheap and independent of both the cipher and the classifier, so if the effect turns out to be real it is a usable add-on.

  2. The limitations are stated openly: no security claim for S-FF3-1, no formal length-hiding guarantee for padding, no significance claim, single seed.

Critical weaknesses
  1. The security framing has no threat model, and the cryptographic basis is shaky.

The paper motivates itself with security-critical data but never says what the party doing the fine-tuning is assumed to see. FF3-1 has published cryptanalysis (Beyne, CRYPTO 2021), and NIST's second draft of SP 800-38G Rev. 1 (Feb 2025) drops FF3 and FF3-1 altogether. The revised drafts also require a minimum domain size of one million, which single short words do not meet, and the paper does not explain how short spans are handled. S-FF3-1 itself is described only in the supplement.

More fundamentally, the stated goal is to make ciphertext easier for a pretrained model to learn, and no experiment asks whether this also makes it easier for an adversary to learn something (frequency analysis over repeated spans, plaintext or attribute inference from the ciphertext or from the fine-tuned model). Saying the preprocessing "adds no cryptographic guarantee" sidesteps the issue. The learnability gain is exactly the quantity that needs an attack-side measurement.

  1. The padding component is not evaluated on the security side, and its reported cost is probably optimistic.

Table 3 gives accuracy and coverage only. Nothing shows how well a length predictor does with and without padding, which is the whole purpose of the component. There is no threat model. The attacker is the lightweight predictor the fillers were optimised against (independent attackers are listed as future work), and there are no cheap baselines such as padding to a fixed length or to bucket boundaries. The gate leaves 14% to 35% of records uncovered, so those either go out unpadded or disappear from the evaluation. If the gate's utility criterion looks at the label or at correctness, it cannot be run at deployment and the 1.2 to 7.85 point cost is understated.

There is also an inconsistency. The S-FF3-1 + Pre column in Table 3 is supposedly computed on the gate-covered subset, yet it matches the full-set value in Table 1 to four digits on five of eight datasets (Finance, MedMCQA, Alpaca, Law, EXAMS), even though coverage is only 65-86%. Finally, for MedMCQA, Alpaca, Law and EXAMS the padded accuracy falls below the no-preprocessing S-FF3-1 baseline in Table 1. The subsets differ, so this is not a clean comparison, but it suggests padding can cancel the headline gain.

  1. The accuracies under encryption are hard to believe, and nothing in the paper rules out a leak through unencrypted channels.

AES-CBC with preprocessing reaches 0.8984 on MedMCQA and 0.9152 on Finance, and even without preprocessing AES-CBC is at 0.8451 on MedMCQA. A 4B model with rank-8 LoRA should not be able to answer knowledge-heavy medical questions at that level from AES ciphertext of the question. I would expect something close to the label prior unless another channel carries the signal. On Alpaca, S-FF3-1 + Pre (0.9302) beats the plaintext baseline (0.9042), which is also odd if encryption removes information. The plaintext numbers themselves (0.9156 on MedMCQA) look well above what I know of models this size on the original benchmark.

Several datasets are not natively multiple choice (Alpaca, FinQA), and the main text does not say how options and distractors were built. If they were generated, the tasks may be solvable from surface cues. It is also unclear what is actually encrypted. Fig. 1 says "encrypt rewritten spans", while Sec. 3.3 excludes entities, numbers, answer-bearing terms and high-information words from editing. Are those spans encrypted or left in plaintext? Are the answer options part of the protected field? Without this I cannot tell what the headline numbers measure. At minimum I would need blind baselines (same-length random strings in place of the protected field, label-only, options-only), a train/test duplicate check, and a field-level description of what is encrypted for each dataset.

  1. The surprisal-based mechanism is asserted rather than tested, and simpler explanations are not excluded.

All four selection criteria are computed on plaintext candidates, but the classifier only ever sees ciphertext. For FF3-1 or AES, the output for a word depends on its length and alphabet, the key and the tweak. It does not depend on how predictable the word is in context, so I do not see why lowering plaintext surprisal would make the ciphertext easier to learn. No ciphertext-side analysis is given (subtoken counts and fragmentation, ciphertext perplexity, anything of that kind).

Meanwhile the rewrite also helps plaintext training (+4.84 points on Alpaca), and the plaintext gains on the other seven datasets are almost identical (0.60 to 0.74 points) despite very different data. That looks more like a generic effect of rewriting (normalisation, shorter or more frequent words, regularisation) than representation compatibility. The controls I would want are random replacements matched on length and character class at the same edit rate, rewriting applied only at training time or only at test time, and some measurement showing which property of the protected sequence actually changes.

Questions for authors

Q1. For each dataset, which fields are encrypted (question, options, instruction)? Are protected-term spans encrypted or left as plaintext? Under AES-CBC, what is the IV policy and the encryption granularity (span, sentence, whole field)?

Q2. How were Alpaca and FinQA converted to four-option multiple choice, and what are the test-set sizes per dataset?

Q3. Does the padding gate use ground-truth labels or correctness? What happens at inference to records that fail the gate? Why do the S-FF3-1 + Pre numbers in Table 3 equal the Table 1 full-set values on five datasets?

Reviewer_mYZ1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a surprisal-aware preprocessing framework to improve LLM classification over encrypted text. The framework transforms selected plaintext spans before protection while preserving task-relevant and structural content, without modifying the downstream classifier or protection mechanism. The authors evaluate the framework across eight cross-domain classification datasets under S-FF3-1, standard FF3-1, and AES-CBC protection configurations. Under the primary S-FF3-1 setting, preprocessing improves macro-average accuracy from 0.8230 to 0.8794, with positive gains across all eight datasets. Controlled rewriting baselines and criterion ablations further support the proposed selection strategy. The paper additionally introduces a randomized padding extension to reduce information leakage through observable sequence length.

Critical strengths

1. Novel perspective on improving the utility of encrypted text.

The paper explores a relatively novel direction by introducing a preprocessing step before text encryption to improve the learnability of encrypted representations. Unlike existing privacy-preserving text transformation methods, the proposed framework focuses on improving downstream task utility without modifying the underlying protection mechanism or classifier.

2. Extensive evaluation across multiple datasets and encryption configurations.

The paper evaluates the proposed method on eight datasets spanning different domains and three protection configurations (S-FF3-1, FF3-1, and AES-CBC). The consistent accuracy improvements across these settings provide empirical evidence for the effectiveness and potential generalizability of the preprocessing framework.

3. Consideration of additional privacy leakage through sequence length.

Beyond improving classification utility, the paper also considers potential information leakage through observable sequence length and introduces a randomized padding extension. This demonstrates an effort to address privacy concerns beyond content encryption, while acknowledging the trade-off between padding and downstream classification performance.

Critical weaknesses
  1. Insufficient implementation details regarding the protection pipeline and task construction. The paper lacks sufficient details to interpret and reproduce its core results. Although the paper provides a high-level description of S-FF3-1 tweak derivation, important implementation details remain unclear. Moreover, the construction of the multiple-choice classification tasks is insufficiently explained. Several datasets, such as FinQA and Alpaca, do not originally follow a multiple-choice format, yet the paper does not clarify how answer candidates and distractors are generated. These omissions make it difficult to determine whether the high classification accuracy reflects meaningful information learned from encrypted representations or residual structural information and other unencrypted cues.

  2. Lack of comparisons with established privacy-preserving text transformation methods. Although the paper discusses prior text sanitization approaches such as SanText and CusText, it does not provide empirical comparisons with these methods. The experiments primarily focus on conventional encryption configurations and alternative rewriting strategies. While these approaches have different privacy guarantees and threat models, comparisons with established text sanitization methods, or evaluations of whether the proposed preprocessing can improve their downstream utility, would help clarify the contribution and applicability of the proposed framework. Without such evidence, it remains unclear how the method relates to existing privacy-preserving NLP approaches(such as SanText/CusText/TextObfuscator etc.) beyond the specific encryption configurations studied.

  3. Unclear threat model and privacy protection objectives. The paper does not clearly define its threat model. Although preprocessing and encryption are performed in a trusted environment, it remains unclear who the adversary is, what information they can access (e.g., protected inputs, metadata, training labels, or model parameters), and what privacy properties the system aims to guarantee. In particular, the paper considers both content confidentiality and length-based information leakage, but the relationship between these two threat scenarios is not adequately specified. A more explicit threat model is necessary to understand the security assumptions and evaluate the practical privacy guarantees of the proposed framework. Prior studies have demonstrated that protected text representations can be vulnerable to reconstruction and inference attacks under various threat models. Although these attacks are not directly transferable to properly encrypted ciphertext, the paper should clarify what information remains observable under its specific protection configurations and evaluate whether this information enables sensitive-token recovery, attribute inference, or partial text reconstruction.

Questions for authors
  1. Threat model: Could the authors clarify the assumed threat model, including the adversary's capabilities, accessible information (e.g., encrypted inputs, unencrypted metadata, labels, and model parameters), and intended privacy protection goals? In particular, what information is the framework designed to protect against inference or reconstruction?

  2. Protection and task construction: Please specify the protection configurations, unencrypted fields, and multiple-choice construction procedures for datasets such as FinQA and Alpaca. What evidence distinguishes learning from encrypted content from reliance on unencrypted or structural cues?

  3. Privacy attacks: Have the authors evaluated potential information leakage beyond sequence length, such as sensitive-token recovery, partial plaintext reconstruction, or attribute inference? Given the readable fragments observed in the FF3-1-based representations (Supplementary Figure 4), how robust are these protected representations against such attacks under the assumed threat model?

  4. Baseline comparisons: How does the proposed framework compare empirically with established text sanitization methods such as SanText and CusText, accounting for their different privacy guarantees and threat models?

  5. Length-inference evaluation: What quantitative reduction in length-inference attack performance does randomized padding achieve against independently trained attackers? How are records without any accepted padding draw handled, and what is the resulting privacy–utility trade-off over the full test set?

Reviewer_Zs89OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper studies text classification when the training/inference inputs must be protected (e.g., via encryption) and proposes a surprisal-aware preprocessing framework that selectively rewrites plaintext spans before protection to make the resulting protected representations more compatible with pretrained LLMs. The core idea is a conservative, constraint-checked rewriting pipeline that generates context-appropriate candidates using a frozen LM and selects replacements via a pairwise ordinal aggregation of four criteria: sentence-level semantic similarity, token-level contextual similarity, contextual surprisal, and POS compatibility. Results show that for eight multiple choice datasets and three types of protection improve performance over the model with a random padding to reduce length side channel leakage at a reasonable cost.

Critical strengths
  • The paper frames classification of encrypted text as a representation-compatibility problem, and proposes a principled constrained rewriting approach to improve learnability without changing the downstream classifier or protection mechanism.
  • The pairwise ordinal selection over four heterogeneous criteria is a neat alternative to weighted score fusion, easy to justify, and empirically validated via ablations.
  • The randomized padding extension is a pragmatic attempt to address residual length leakage and the idea of building a filler-template bank tuned against a predictor while preserving classifier behavior is interesting.
Critical weaknesses
  • The security gains from padding have not been tested against a set of highly targeted adversaries or different predictors. The analysis does not report any measure of leakage reduction that makes security gains difficult to assess.
  • Rewriting relies on a frozen LM and a few thresholds . The paper does not describe how these hyperparameters are selected across datasets or their sensitivity, leading to questions about tuning and generality.
  • The protected-term filter is heuristic may allow unintended changes, or may be excessively conservative in certain areas; there is no human or automatic label-preserving audit reported.
  • No runtime/latency/throughput measurements for preprocessing and padding are provided. The cost of LM-based candidate generation and ordinal selection could be material in production.
  • Results include only one training seed, and no variance or paired significance testing is reported. Given the size of gains, this is likely robust, but the lack of variability analysis weakens claims of general effectiveness.
  • Some deployment details (e.g., candidate pool size k, retokenization edge cases, error handling) are deferred to the supplement, making it hard to fully assess reproducibility from the main paper.
Questions for authors
  • How were εm, δ, ρ, Bmax, and the candidate pool size k chosen? Are they fixed across datasets or tuned per-domain? Please provide a sensitivity analysis.
  • What sentence/context encoders did you use for S_i(c) and T_i(c), and how robust are results to swapping these?
  • Could you compare against at least one model-adaptation baseline (e.g., SentinelLMs) on a subset to contextualize the utility vs. deployment complexity trade-off?

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

Related papers

Discussion (0)

Sign in to comment.