Teams building non-English language models increasingly buy or borrow "native" instruction and evaluation data. Much of what is sold as native is in fact machine-translated English. That matters. Models tuned on translated data look fine on translated benchmarks but lose ground on text written by native speakers. On XQuAD, Qwen-14B reaches 36.5% exact match when trained on native data and 28.5% when trained on translated data (Is It Good Data for Multilingual Instruction Tuning or Just Bad Multilingual Evaluation for Large Language Models?). Translated text carries a statistical fingerprint that classifiers can detect. This fingerprint is called translationese: flatter vocabulary, extra explicit pronouns and connectives, and word order borrowed from the source language. Detecting it, however, normally means reading the text, and data owners will not hand over the text before a deal is made.
We propose a way to measure how much translationese a corpus contains while the corpus stays encrypted. The same ciphertexts can later be handed to the buyer by an untrusted cloud server. The tool that makes this possible is proxy re-encryption: a scheme in which a server holding a special "re-encryption key" can convert data locked for Alice into data locked for Bob, without ever seeing the contents. We pair a lattice-based proxy re-encryption scheme that also supports simple arithmetic on encrypted data with a translationese detector deliberately restricted to linear scoring. The server can then compute each document's translationese score on ciphertext and forward only the scores to a prospective buyer. After purchase, it forwards the full content. The buyer learns how native the data is before seeing a single sentence. Post-purchase checks make it costly for the seller to lie.
Core Claim
A translationese detector built as a linear function of per-document feature counts loses little accuracy compared with the nonlinear classifiers used in current translationese studies. Because it is linear, an honest-but-curious server can evaluate it on encrypted data using only cheap plaintext-by-ciphertext multiply-adds. Combined with separate encryption keys for "audit scores" and "content," this gives a pre-purchase audit that reveals nothing about the text beyond one score per document. The scores it produces predict the downstream native-benchmark gap caused by training on that data.
Two things would be new. The first is using translationese statistics as a certified property of private data, not as a post-hoc observation about public corpora. The second is showing that the audit number is decision-relevant: it forecasts how much a model will gain on native benchmarks if the corpus is bought and used.
Where This Sits
The file contains no work at the exact intersection of these two areas, so the closest neighbors come from each side separately.
On the privacy side, EncChain protects retrieval-augmented language model pipelines by decrypting data inside Intel TDX hardware enclaves, which are processor-isolated virtual machines. It pays only a 0.3–1.5% latency overhead. It trusts the hardware vendor, and it never asks what kind of text it is protecting. Glyph supplies a method we reuse. It moves costly work on encrypted data into ciphertext-times-plaintext operations by keeping public components (for Glyph, frozen early layers) unencrypted. Here the detector weights play that public role. Glyph works only on vision models and never re-keys data between parties.
The cryptographic core comes from HRA-Secure Homomorphic Lattice-Based Proxy Re-Encryption with Tight Security. That scheme builds proxy re-encryption on BGV homomorphic encryption and adds calibrated noise so that a curious proxy watching many re-encryptions cannot extract the owner's key. It reports roughly 2 ms per single-hop re-encryption on a desktop CPU. It was tested only on simulated 5G network slices. We are its first language-data application, and we need its homomorphic arithmetic, not just its delegation.
Identity-Based Proxy Re-encryption motivates our requirement that delegation be one-way: the buyer's key must not let anyone reverse the path. The single-hop restriction analyzed in Research on Non-interactive Construction based on Fuzzy Conditional Proxy Re-encryption motivates forbidding the buyer's copy from being re-delegated. Conditional re-encryption in that line of work, meaning keys that work only on ciphertexts carrying a given tag, is the principled way to separate "score" from "content." The lattice scheme we use lacks tags, so we emulate them with two key pairs.
On the language side, the detector design comes from An Investigation of Translationese in the Generations of Multilingual Large Language Models. It trains support vector machine classifiers on native versus translated Europarl text, using lexical diversity, function-word rates, part-of-speech n-grams, and lengths. Source Language Markers in EUROPARL Translations shows that plain word n-gram frequencies identify a translation's source language with up to 96.7% accuracy. That result is the main reason we expect linear models over count features to be strong. The motivating harm comes from Is It Good Data… and ArabicMMLU. In ArabicMMLU, GPT-4 drops from about 80% on translated MMLU to 72.5% on native Arabic exams.
One constraint needs stating plainly. Can you map it to English? shows that multilingual models route non-English inputs through English-aligned middle layers. This means human-authored text post-edited by a language model may carry translationese too. Our detector therefore measures "translated-like" text, not provenance. We report it that way.
Research Program
The protocol involves four parties. The owner O holds a corpus. The prospective buyer B wants to know its quality. A public auditor publishes a detector. An untrusted cloud stores everything.
The owner computes a feature vector $x_i$ for each document $i$ on their own machine. It contains hashed counts of function words, part-of-speech and character n-grams, type-token ratio, and length statistics. Nonlinear features such as ratios are acceptable because the owner computes them before encryption. The owner encrypts $x_i$ under an audit key $pk_{O}^{aud}$ and the document text under a separate content key $pk_{O}^{con}$. Both ciphertexts go to the cloud.
The auditor publishes weights $w$ and bias $b$. The cloud homomorphically computes the encrypted score
$$\mathrm{Enc}(s_i) = \mathrm{Enc}(x_i)\cdot w + b,$$
where the dot product is between an encrypted vector and a public plaintext vector. BGV packs many feature slots into one ciphertext, so this costs one multiply-and-rotate pass per document.
The owner then gives the cloud a re-encryption key $rk^{aud}_{O\to B}$. The cloud converts the score ciphertexts so that B can decrypt them. B then sees the full distribution of per-document scores, and from it the estimated share of translated-like documents overall and within any metadata slice the owner exposes. If B buys, O issues $rk^{con}_{O\to B}$ and the cloud re-encrypts the content.
Honesty is enforced in two ways. First, before purchase, B may request content delegation for a small random subset chosen after the scores are fixed. B recomputes $x_i$ on those documents and checks them against the encrypted features, which are then also delegated. Second, after purchase, B can recompute every feature, so any mismatch becomes evidence for a contractual penalty. Neither check is cryptographic proof. Both make cheating detectable with probability that grows with the size of the spot check, and we quantify that probability.
The machine-learning work has three parts. First, we test whether linear detectors over hashed count features match the nonlinear classifier of An Investigation of Translationese…. We test this across human translation, Google-style MT (GTNC), and language-model generations, in high-resource and low-resource languages. Second, we test robustness against a seller who post-edits machine translation with a language model to evade the detector. Third, we test whether the audited translationese rate predicts downstream harm.
Baselines
The naive reference point is no audit: the buyer accepts the owner's "native" label. In our mixture experiments this corresponds to training on whatever blend of native and translated data is delivered. The expected cost is the native-versus-translated gap measured by Is It Good Data…, namely 36.5% versus 28.5% XQuAD exact match for Qwen-14B.
The strongest existing detector is the support vector machine of An Investigation of Translationese in the Generations of Multilingual Large Language Models, run on plaintext. That paper reports native human German flagged as translationese 0% of the time, Gemini's German generations 60.2%, and Llama's 37.6%. We retrain it with its own feature set on the same Europarl-UdS splits we use, giving it the same hyperparameter search budget as our linear detector. It serves as the accuracy ceiling. Our encrypted linear detector "wins" if its detection accuracy and flag rates are within two points of this ceiling at every language and generation source. A slightly lower number is acceptable because the baseline requires plaintext access. If the gap exceeds that, the linear restriction is too costly and the proposal fails on its main premise.
The strongest privacy competitor is running that same nonlinear classifier inside a hardware enclave, following EncChain. It reports 0.3–1.5% end-to-end overhead for its pipeline. We run it on the same corpus sizes and hardware class and report wall-clock time and dollar cost alongside ours. We do not expect to beat it on speed. The comparison makes the price of removing hardware trust explicit.
The plaintext-sample baseline has the owner reveal a random fraction of documents, and the buyer runs the full detector on them. We match its revealed fraction to our spot-check fraction, so both methods leak the same amount of content.
The key ablation removes the homomorphic evaluation: the owner computes and reports scores themselves, with the same spot check. This isolates whether computing on the committed ciphertexts, rather than trusting self-reported scores, actually raises the chance of catching a cheating seller.
For latency of homomorphic dot products on text features, we found no reported result on Lacuna. The only anchor is the 2 ms single-hop re-encryption figure in HRA-Secure Homomorphic Lattice-Based Proxy Re-Encryption.
Why not just…?
Why not just reveal a random plaintext sample and run the best detector on it? This is the most serious alternative. For estimating a single contamination rate it may work nearly as well: a few thousand sampled documents give a tight confidence interval. Our design is justified only in two cases. One is when the owner cannot reveal any content before sale, as with licensed, medical, or benchmark data where a leaked item is a burned item. The other is when contamination is concentrated in small slices that a sample misses. We test both directly. We plant machine-translated documents at 1–10% rates, either spread across the corpus or confined to one subdomain. We then compare slice-level detection power at equal revealed content. If sampling matches full encrypted coverage in every setting, we will report that the cryptographic audit is only worth it under a zero-reveal constraint.
Why not just use a hardware enclave? Enclaves are fast, and EncChain's overhead numbers show they would beat us on latency. The trade is about trust. An enclave requires trusting the chip vendor and the absence of side channels. It also decrypts everything in one place, so it provides no key-level separation between score access and content access. Proxy re-encryption provides that separation, and one-way delegation keeps the buyer from re-delegating. We present these as different trust assumptions with measured costs, not as a claim that enclaves are unsafe.
Why not just download, decrypt, and re-encrypt on the owner's side? This requires the owner to stay online and serve every buyer. More importantly, it gives no assurance that the scores the buyer sees were computed from the same documents that are later delivered. The self-report ablation above measures exactly this binding benefit.
Why not just trust scaling or alignment training to remove translationese from newer data? The file's evidence says it does not disappear. Gemini's German is still flagged 60.2% of the time. High-resource languages show more translationese than Pashto or Greek, because more translated text exists online. The native-data advantage correlates with model strength at about 0.81 (Is It Good Data…).
Proposed Experiments
Detector experiments use Europarl-UdS native versus translated text in German and Spanish, with Pashto from OLDI Seed and Wikipedia as a low-resource check. GTNC supplies machine-translation signatures. Language-model generations follow the protocol of An Investigation of Translationese…. For each language we compare three detectors: the nonlinear support vector machine, a linear model on the same features, and a linear model on hashed n-gram counts at several hash sizes. We report accuracy and flag rates. For robustness, we have a language model post-edit machine-translated documents to sound native, and we measure how far flag rates fall.
Downstream prediction uses the native and translated instruction data from Is It Good Data…. We build training mixtures with known translated fractions, audit each mixture through the full encrypted pipeline, and fine-tune small open models on each. We evaluate on native benchmarks (CMMLU, C-Eval, TyDi QA Russian, and ArabicMMLU where an Arabic mixture is feasible) and on translated ones (MT-MMLU, XQuAD, MGSM).
The decisive checks are three. First, the correlation between the audited translationese rate and the native-benchmark drop must be strong. In the same experiments, the translated benchmarks should show the flat, "negligible differences" pattern the source paper found on MT-MMLU. If the audit does not predict native-benchmark loss, it is not a useful purchase signal. Second, encrypted scores must match plaintext scores after decryption, up to the fixed-point quantization error from encoding real-valued weights as integers. We also run the curious-proxy leakage analysis supported by the lattice scheme's noise flooding. Third, we plant false features for a fraction of documents and measure how often the spot check plus post-purchase recomputation catches them, compared with the self-report ablation. This yields a detection-probability curve against spot-check size, which a buyer can use to set contract terms.
We will report costs per million documents: encryption, homomorphic scoring, re-encryption, and decryption, set against the enclave and plaintext-sample baselines.
References
- Is It Good Data for Multilingual Instruction Tuning or Just Bad Multilingual Evaluation for Large Language Models?
- EncChain: Enhancing Large Language Model Applications with Advanced Privacy Preservation Techniques
- Glyph: Fast and Accurately Training Deep Neural Networks on Encrypted Data
- HRA-Secure Homomorphic Lattice-Based Proxy Re-Encryption with Tight Security
- Identity-Based Proxy Re-encryption
- Research on Non-interactive Construction based on Fuzzy Conditional Proxy Re-encryption
- An Investigation of Translationese in the Generations of Multilingual Large Language Models
- Source Language Markers in EUROPARL Translations
- ArabicMMLU: Assessing Massive Multitask Language Understanding in Arabic
- Can you map it to {E}nglish? The Role of Cross-Lingual Alignment in the Multilingual Performance of {LLM}s
- GTNC: A Many-To-One Dataset of Google Translations from NewsCrawl