- 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
-
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.
-
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
- 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.
- 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.
- 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.
- 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?