acceptodds
Under review as a conference paper at ICLR 2027

PII-RADAR: RApid Detection And Replacement for PII Scrubbing

Abstract

Large language models have been shown to leak their training data, which becomes a privacy risk when this data contains personal information. In practice, vulnerable data is commonly removed using scrubbers, *i.e.*, tools that detect personally identifiable information (PII) in a document and remove it before training. However, scrubbers are imperfect and miss some PII in the training data. In this work, we first show that, while scrubbing leaves less information exposed, the PII it misses becomes significantly more vulnerable. The leakage we observe increases with the detection rate, so scrubbers with stronger detectors increase the risk for missed PII. We identify PII removal as the primary cause of this leakage, because it leaves only a small amount of real PII in the dataset, which the model then memorizes. While synthetic rewriting of the detected PII should, in theory, resolve this issue, current rewriting tools generate highly repetitive or malformed text, increasing the risk similarly to removal. To address this issue, we propose PII-RADAR, a novel scrubber that replaces each detected PII with a realistic one of the same type, drawn from a pool of validated synthetic identities and formatted to match the original document. Contrary to other rewriting scrubbers, PII-RADAR's replacements are diverse, well-formed, and non-repetitive, lowering the privacy risk of any PII it misses **up to 3×** compared to baselines. Moreover, PII-RADAR outperforms all evaluated scrubbers on five of seven benchmarks, establishing a new state of the art in PII detection, while being up to 10× faster than other scrubbers that replace rather than redact.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
privacy, scrubbers, Personally identifiable information (PII)

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HPsEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates the privacy risk of personally identifiable information (PII) that remains in training data after imperfect scrubbing. It shows that masking detected PII can increase the vulnerability of missed PII to membership-inference and extraction attacks, and that existing replacement methods may provide limited protection when their synthetic values are repetitive or poorly formatted.

To address this issue, the paper introduces PII-RADAR, an LLM-based scrubber that detects PII and replaces it with synthetic counterparts drawn from validated profiles. A copy gate distinguishes non-PII tokens, which are preserved, from PII tokens, which are replaced. The method uses profile-conditioned CTC decoding to produce replacements of varying lengths that match the original PII’s type and format. Evaluation across seven detection benchmarks and controlled leakage experiments demonstrates high detection recall, improved replacement quality, reduced leakage of missed PII compared with the evaluated baselines, and faster processing than other replacement-based scrubbers.

Critical strengths
  1. A useful evaluation perspective supported by meaningful controls - The paper evaluates privacy risk for missed PII instead of relying exclusively on detector accuracy. Section 3 compares the same missed PII under scrubbed and unscrubbed training and includes random-masking controls. Figure 6 additionally holds PII-RADAR’s detections fixed while changing the replacement strategy. These comparisons provide convincing evidence that the treatment of detected PII can affect the risk faced by the PII left behind.
  2. An efficient system with substantial empirical improvements - The combination of a copy gate, profile conditioning, and CTC replacement provides an efficient alternative to regenerating an entire document. The reported results cover detection, replacement validity and diversity, downstream leakage, and runtime. PII-RADAR achieves 0.936 pooled span recall, approximately 92.3% structurally valid replacements, and roughly 3–10× higher throughput than the evaluated rewriting systems.
Critical weaknesses
  1. The model’s contribution is not isolated from the synthetic profile resource. The masking ablation shows that replacement helps, but does not compare against a simpler replacement procedure using the same detected spans and profile bank. Such a comparison would clarify whether the gains require the proposed architecture or primarily come from better replacement data.
  2. The method description is fragmented. Important details about the backbone, copy gate, and training appear outside Section 4. The method section should explain how the copy gate works and classifies tokens as PII or non-PII, including its training labels and decision threshold. It should also clarify the combined training objectives and span selection during training and inference to support reproducibility.
Questions for authors
  1. Provenance and calculation of statistics: Are the statistics in lines 71–77 and 87–96 taken from prior work or measured in your experiments—for example, the statement that “Anonymizer-4B edits 26.3% of detected credit card numbers”? Although Section 5.3 and Appendix E provide evaluation details, please explicitly link these introductory claims to the relevant experiments. Please specify the datasets used, what each percentage is calculated over, and any exclusions applied when collecting replacements.
  2. Story generation: Appendix D.1 identifies three models used to generate the stories. How many stories each contribute, what prompts and sampling procedure are used, and how are their outputs distributed across the training and evaluation splits? Does the observed leakage trend hold separately for each generating model?
  3. Missed-PII evaluation: How are missed PII instances identified and counted in Section 3.1, particularly when a scrubber detects or replaces only part of a PII span? How are such cases handled in membership inference and extraction, and how is the same evaluation set maintained between treatment and control?
  4. Missed-PII evaluation: How are missed PII instances identified and counted in Section 3.1, particularly when a scrubber detects or replaces only part of a PII span? How are such cases handled in membership inference and extraction, and how is the same evaluation set maintained between treatment and control.
  5. Runtime evaluation: Why are the runtime measurements supporting the approximately 3–10× speedup claim presented only in Appendix C.4? Could you summarize the timing setup in the main text and directly reference Table 6, clarifying batching, input-length handling, and whether preprocessing, profile retrieval, and decoding are included?
7 comments
AuthorsOpenReview ↗

PII-RADAR Protects Like an Oracle; Method Section Reworked; Results Clarified

We would like to thank the Reviewer for a positive assessment of our work, for recognizing that our evidence is convincing, and for highlighting that PII-RADAR is an efficient system with substantial empirical improvements. Below we address each of the comments one by one:

W1: The model’s contribution is not isolated from the synthetic profile resource. The masking ablation shows that replacement helps, but does not compare against a simpler replacement procedure using the same detected spans and profile bank. Such a comparison would clarify whether the gains require the proposed architecture or primarily come from better replacement data.

We implement the simpler replacement procedure with an oracle that knows the correct type and format for each detected PII. With this oracle, we insert values from the profile bank into spans, and get MIA Leakage Gain similar to the one default PII-RADAR’s:

MethodMIA Leakage Gain
Default PII-RADAR1.53 ± 0.02
Detect with PII-RADAR, replace from profile bank1.55 ± 0.02

We added these results to Appx. C.8 (Tab.14).

High-quality replacements thus protect the missed PII, and PII-RADAR's replacements protect them as well as inserting valid PII directly.

In real scrubber deployments there is no oracle, so the scrubber must infer types and formats before it can select and format a replacement value. As shown in table below, existing scrubbers that classify PII types assign a wrong or overly broad type to 25-39% of detected PII (below, we report type accuracy, %, columns refer to detection benchmarks):

ScrubberNemotron300kREDACTSPY legalSPY medicalPrivasis vanillaPrivasis hardAll benchmarks
Presidio76.865.777.475.873.673.474.075.1
Piiranha76.061.068.362.158.664.768.070.2
OPF80.538.678.075.974.772.572.167.6
OpenPipe71.238.488.967.566.772.780.061.3

We added these results to Sec. 5.3 (Tab.4).

W2: The method description is fragmented. Important details about the backbone, copy gate, and training appear outside Section 4. The method section should explain how the copy gate works and classifies tokens as PII or non-PII, including its training labels and decision threshold. It should also clarify the combined training objectives and span selection during training and inference to support reproducibility.

We used the additional 1 page for the rebuttal in the main paper to add the important details. We also restructured Sec. 4 to clarify PII-RADAR. Sec. 4.1 now describes the architecture, LLM backbone, and the copy gate with its training labels, decision threshold, and inference-time PII classification. Sec. 4.2 now details the replacement mechanism, including synthetic-profile conditioning, CTC decoding, and frame expansion. Sec. 4.3 describes the training data, ground truth for the copy gate and replacement, joint training objective, and full inference procedure, clarifying that training uses ground-truth PII spans, while inference uses spans predicted by the copy gate. We also updated Fig. 3.

Q1: Provenance and calculation of statistics: Are the statistics in lines 71–77 and 87–96 taken from prior work or measured in your experiments—for example, the statement that “Anonymizer-4B edits 26.3% of detected credit card numbers”?

We measured all of these numbers ourselves, using PII produced by rewriting scrubbers for the benchmark documents, following Appx. E.2. Anonymizer-4B replaced 3374 of 12850 credit card numbers with “XXXX XXXX XXXX XXXX” (26.3 %). We clarified these results in the introduction, linking them to the experiments in Appx. E, and added the experimental setup and format rules there. Tab. 19 (Appx. E.4) reports the par-value shares.

AuthorsOpenReview ↗

Evaluation Details for Claims Clarified; Generation Details of Long Stories (Short) Clarified; Missed PII Clarified

Q1: Although Section 5.3 and Appendix E provide evaluation details, please explicitly link these introductory claims to the relevant experiments. Please specify the datasets used, what each percentage is calculated over, and any exclusions applied when collecting replacements.

We added a comprehensive setup to Appx. E detailing that:

  • we use data from all seven PII detection benchmarks;
  • we check nine PII types that are verifiable, with format rules described in Appx. E.2;
  • we evaluate only the spans that each scrubber rewrites;
  • we exclude spans that remain unchanged, are deleted, or whose replacements cannot be identified in the output.

To identify replacement PII, we use the scrubber's list of replaced values. If the span is not in the list, we take the output text between the span's surrounding words. This fails when the rewriter also changes those words or when they repeat. It mainly affects Privasis, which provides no replacement list: we cannot identify 41% of its changed spans, vs. under 2.5% for every other scrubber. We clarify that in Appx. E.

All Sec. 5.3 percentages use as denominator the gold spans each rewriter replaced, pooled over all seven benchmarks and nine types (counts now in Tab. 17).

Below we clarify our claim and source with extended results.

ClaimShare#matching / # total replaced spansSource
Privasis Cl. 4B writes REDACTED for phones29.0%10,880 / 37,537Tab. 19
Anonymizer 4B writes XXXX XXXX XXXX XXXX for cards26.3%3,374 / 12,850Tab. 19
OpenPipe e-mails at example.com/.net/.org91.9%77,401 / 84,199Fig. 15; Tab. 19
OpenPipe valid card numbers0.0%0 / 6,436Fig. 15; Tab. 19
Valid, not a placeholder: PII-RADAR (ours)92.3%250,661 / 271,672Figs. 5, 15
Valid, not a placeholder: Anonymizer 4B38.2%102,282 / 267,625Figs. 5, 15
Valid, not a placeholder: Privasis Cl. 0.6B1.4%2,106 / 145,284Figs. 5, 15
Valid, not a placeholder: Privasis Cl. 4B0.7%1,072 / 153,071Figs. 5, 15
Ten most frequent values: PII-RADAR (ours)0.07%192 / 271,672Figs. 5, 14; Tab. 19
Ten most frequent values: Privasis Cl. 0.6B69.5%100,943 / 145,284Figs. 5; 14, Tab. 19
Ten most frequent values: Privasis Cl. 4B74.1%113,446 / 153,071Figs. 5, 14; Tab. 19

Q2: Story generation: Appendix D.1 identifies three models used to generate the stories. How many stories each contribute, what prompts and sampling procedure are used, and how are their outputs distributed across the training and evaluation splits? Does the observed leakage trend hold separately for each generating model?

We clarified in Appx. D.1 that GPT-5.6 Luna contributes 10,177, Claude Sonnet 5 contributes 10,431, and Qwen3.5-122B-A10B contributes 2,038 stories, see Tab. 15. We added the generation prompt (Fig. 12), the template example (Fig.13) and sampling details to Appx. D.1.

When training the target and reference models for InfoRMIA, we select stories at random from the set of all samples in Long Stories (Short). We follow the same sampling for the models trained for extraction experiments. We provide counts for each of the four seeds in Appx. D.1, Tab. 15.

The observed leakage trends hold for each generating model, see Figs. 22, 23 in Appx. F.4.

Q3&4: Missed-PII evaluation: How are missed PII instances identified and counted in Section 3.1, particularly when a scrubber detects or replaces only part of a PII span? How are such cases handled in membership inference and extraction, and how is the same evaluation set maintained between treatment and control?

We create stories from templates with known PII fields, so we can track every inserted value. After scrubbing, values that remain exact matches are labeled as missed PII, while partially replaced values are classified as detected. Extraction succeeds when the PII is generated from the LLM, while partial matches are counted as failures (Sec. 3.1; Appx. D.2, D3).

To compare leakage with and without scrubbing, the PII must be the same in both models and there must be no distribution shift between members and non-members. We scrub the full Long Stories (Short) and apply the same random 50/50 split of story IDs to the original and scrubbed versions. Control LLMs train on the original members and treatment LLMs on the scrubbed ones, so treatment's member and non-member PII are subsets of control's. We then run InfoRMIA on both models only for control PII that survive scrubbing, i.e., the missed PII. We clarified this in Appx. D.2.

AuthorsOpenReview ↗

Runtime Evaluation Clarified

Q5: Runtime evaluation: Why are the runtime measurements supporting the approximately 3–10× speedup claim presented only in Appendix C.4? Could you summarize the timing setup in the main text and directly reference Table 6, clarifying batching, input-length handling, and whether preprocessing, profile retrieval, and decoding are included?

Due to space constraints, we previously reported the runtime setup and results in Appx. C.4 (Tab. 9, formerly Tab. 6). We now report the results in Tab. 3 (main text thanks to the additional 1 page allowed for the rebuttal) and describe the setup in Sec. 5.1 and detailed in Appx. C.4. Each method runs in one process on an NVIDIA A40 over the same seven benchmarks, with batched documents. PII-RADAR splits documents exceeding 8192 tokens at whitespace, whereas the LLM baselines process whole documents within their context windows and OPF uses its own sliding windows. Timings include preprocessing, profile retrieval, and decoding, but not model loading.


We hope that the above answers address the Reviewer's concerns. We are also happy to address any remaining questions.

Reviewer_nv8rOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies PII scrubbing for LLM training data from the angle of downstream privacy leakage instead of detection metrics. The authors present the following observations: (1) scrubbing concentrates privacy risk on the PII it misses—the more PII a scrubber detects, the more the remaining missed PII leaks, and this effect is driven purely by the number of missed PII, not their distinctiveness; (2) existing rewriting scrubbers fail to mitigate this because their replacements are repetitive, placeholder-like, or malformed; (3) they propose PII-RADAR, a non-autoregressive sequence-editing scrubber that detects PII via per-token binary classification and replaces it in place with values drawn from validated synthetic profiles, using CTC decoding with span expansion.

Critical strengths
  1. The observation that scrubbing increases the vulnerability of missed PII, and it scales with detection rate is interesting. Showing that random masking at matched miss rates reproduces the same trend is a clean controlled experiment that convincingly isolates the mechanism (outlier/memorization effects) rather than properties of specific missed instances.
  2. The paper evaluates eight scrubbers across seven benchmarks with per-system type scoping, span- and token-level metrics, a threshold ablation, throughput/cost comparisons (including honest reporting of accelerated baselines), and multiple leakage attacks. The leakage experiments use two backbones, and the "mask-mode twin" analysis disentangles detection quality from replacement quality. Overall, the experiments are well-designed.
Critical weaknesses
  1. The proposed method is: (a) detect PII with per-token binary classification, and (b) replace detected spans with values from a validated synthetic profile pool. Both components are individually well-established. However, token-level PII detection is in the spirit of OPF, NER-based scrubbers, and replacing PII with consistent synthetic surrogates is also used in the existing work, eg. Carrell et al. 2013. The claimed novelty rests on the empirical observation that replacement quality matters and on engineering a system that achieves it. Though this is a useful contribution,  the paper would be stronger if it articulated what is fundamentally new beyond a well-executed combination.
  2. The paper does not systematically answer the following question. As different replacement strategies should yield different downstream leakage, then which aspects of the replacement actually drive the effect, and by how much?
  3. All leakage experiments are conducted on synthetic PII in small 1B models trained far beyond standard practice to induce extraction. While I understand the ethical motivation for synthetic data, this raises questions about external validity: (a) do the leakage-gain trends hold at realistic PII densities and model scales, where memorization dynamics differ? (b) synthetic PII pool is far less diverse than real-world PII — could the outlier effect be amplified by the limited pool?
Questions for authors
  1. if you take Faker-generated values, validate and format-match them with the same rules used for profile values, and use them as replacements at PII-RADAR's detection rate, how does the missed-PII leakage compare? This single experiment would substantially clarify the source of the improvement.
  2. Does cross-span consistency within a profile matter, or would independently sampled valid values per span perform identically? (E.g., replace each span from a freshly sampled profile rather than the conditioning profile.)
  3. The 10× span expansion for CTC is described as arbitrary. What failure modes appear with shorter/longer expansions, and does the CTC formulation introduce any systematic biases in replacement length (e.g., a tendency toward shorter replacements)?
  4. Are the leakage gains observed for Gemma3 quantitatively similar to OLMo2, and does the trend hold at 7B scale or with fewer training epochs?
7 comments
AuthorsOpenReview ↗

Novelty of PII-RADAR; Replacement Ablations

We would like to thank the Reviewer for their assessment that we convincingly isolate the mechanism of increased privacy leakage of missed PII, and that our experiments are well-designed. We address all the raised issues one by one below:

W1: The proposed method is: (a) detect PII with per-token binary classification, and (b) replace detected spans with values from a validated synthetic profile pool. Both components are individually well-established. However, token-level PII detection is in the spirit of OPF, NER-based scrubbers, and replacing PII with consistent synthetic surrogates is also used in the existing work, eg. Carrell et al. 2013. The claimed novelty rests on the empirical observation that replacement quality matters and on engineering a system that achieves it. Though this is a useful contribution, the paper would be stronger if it articulated what is fundamentally new beyond a well-executed combination.

High-quality replacement PII requires much more than simply a) detection followed by b) replacement from a valid profile. For example, the model needs to select the right value and format it to match the original. PII-RADAR does both, without type classifiers or format rules. This avoids common errors in existing scrubbers like OPF, which mistype 25-39% of detected PII (new Tab. 4), fail to match formats, or generate repetitive and invalid values (Tab. 19, Figs. 5, 14, 15). PII-RADAR avoids these errors, and 92.3% of its replacements are type- and format-valid.

HIPS (Carrell et al. (2013)) [1] targets direct data release, where real PII are easy to spot among masked PII, and realistic surrogates are a good solution. We instead study leakage from trained LLMs, and find that scrubbing increases leakage of missed PII as detection rates improve, and that existing rewriting methods do not prevent it. [1] implement their principle with a PII detector and hand-crafted format rules for each type. Our architecture matches types and formats without such rules.

Overall, our work introduces the following novel contributions:

  1. We show that scrubbing increases leakage of missed PII from trained LLMs as detection rates improve, and that existing rewriting methods do not prevent it.
  2. We introduce PII-RADAR, which selects PII from profiles and adapts it to the original context without explicit type classification or hand-crafted formatting rules.

We added discussion of HIPS to Sec. 2, and clarified the novelty of PII-RADAR relative to prior work in the Introduction, Sec. 2, 3.3, and 4.

W2: The paper does not systematically answer the following question. As different replacement strategies should yield different downstream leakage, then which aspects of the replacement actually drive the effect, and by how much? Q1: if you take Faker-generated values, validate and format-match them with the same rules used for profile values, and use them as replacements at PII-RADAR's detection rate, how does the missed-PII leakage compare? This single experiment would substantially clarify the source of the improvement. Q2: Does cross-span consistency within a profile matter, or would independently sampled valid values per span perform identically? (E.g., replace each span from a freshly sampled profile rather than the conditioning profile.)

We vary what we insert into the PII spans detected by PII-RADAR, assuming an oracle that knows correct PII types and formats. We find that PII-RADAR achieves protection like oracle replacements from a consistent profile (MIA Leakage Gain 1.53 vs 1.55).

We test inconsistent profiles (addressing Q2) and observe an average increase of MIA Leakage Gain (1.55→1.72) across most PII types, with the largest MIA Leakage Gain of 2.20 for emails, whose replacement values are not aligned with the corresponding names (Tab. 14; Appx. C.8).

We generated PII with Faker, which show a similar Gain of 1.77 (addressing Q1), while the Gain increases to 2.69 for emails. This happens because Faker produces valid, but repetitive example.com domains by default. Making the email domains diverse drops the Gain to 2.21, and making email usernames consistent with their names drops the Gain to 1.97 (PII-RADAR’s Gain for emails is 1.96).

Inserted valuesType and formatConsistent ProfileMIA Leakage GainEmail Gain
Our PII-RADARInferredyes1.531.96
Our pool of profilesFrom oracleyes1.551.78
Our pool of profilesFrom oracleno1.722.20
Faker profilesFrom oracleno1.772.69
Faker profilesFrom oracle+diverse email domainsno1.542.21
Faker profilesFrom oracle+diverse email domainsyes1.541.97

These findings show that consistency, value quality, and variety play an important role in reducing leakage of missed PII. We added these results to Appx. C.8.

AuthorsOpenReview ↗

Real vs Synthetic PII & Leakage Setup Clarified

W3: All leakage experiments are conducted on synthetic PII in small 1B models trained far beyond standard practice to induce extraction. While I understand the ethical motivation for synthetic data, this raises questions about external validity: (a) do the leakage-gain trends hold at realistic PII densities and model scales, where memorization dynamics differ?

Since scrubbing matters only where PII actually leaks, we need a setup with strong PII leakage, otherwise, leakage measurements are too noisy to guide the design of the new method and provide reliable comparison between the methods. Since pre-training leakage is relatively weak [3], we strengthen it by fine-tuning the LLMs on a small set of samples, all containing PII, and we employ a strong MIA that requires reference models.

We fine-tune LLMs for one epoch for MIA. On the other hand, we fine-tune for 20 epochs, aiming to induce high extraction and obtain a strong signal, which we explicitly state in Appx. D.2-D.3. With fewer epochs the Extractability Gains are simply unstable, as we now show in Tab.16.

We show that the MIA Leakage Gain trends hold across two backbones, OLMo2 7B, and for a different PII density, i.e., a mix of Long Stories (Short) and 100k real pre-training documents from DCLM [2], which brings the rate of documents with PII from 100% to 10%. (all results in Appx. F.3, F.5, Figs. 19, 24).

W3: (b) synthetic PII pool is far less diverse than real-world PII — could the outlier effect be amplified by the limited pool?

We note that the PII pool can refer to: (1) Nemotron-PII [4] used in Long Stories (Short), or (2) our 10.7M synthetic profile pool.

For (1), not all synthetic PII types are guaranteed to be less diverse, e.g., credit card numbers are digit strings with specific formatting rules, and valid values can be easily generated. Moreover, randomly retained PII leaks as much as truly missed PII (Sec. 3.2), indicating that the leakage is driven mostly by the detection rate rather than what exact PII remains.

For (2), our results show that diversity of synthetic replacement PII matters (Fig. 14, Tab. 19). Our new experiments with Faker’s PII (please see answer to W2 above) show that repetitive email domains cause higher leakage of missed emails (Appx. C.8, Tab. 14). This is why we put emphasis on making our 10.7M profile pool very diverse (see Appx. E.1, G.4; Fig. 14), with top-10 most frequent values accounting for only 0.07% of all replacements (Fig. 5).

In both cases, without a real PII database we cannot directly validate our findings on real data, which we state as a limitation in Appx. A. The code used to generate and validate profile data is provided in our submission.

AuthorsOpenReview ↗

CTC Expansion Rate Ablated; Leakage Gains in Other Settings Shown to Hold

Q3: The 10× span expansion for CTC is described as arbitrary. What failure modes appear with shorter/longer expansions, and does the CTC formulation introduce any systematic biases in replacement length (e.g., a tendency toward shorter replacements)?

The original CTC formulation introduces a systematic bias, as CTC decoding can only output sequences shorter than or equal to the input (Sec. 4.2). We overcome this with the frame expansion (Sec. 4.2). Replacement length is correct 91%-96% of the time, with errors slightly biased towards shorter replacements, usually by a single token.

Insufficient frame expansion can lead to malformed outputs when the original PII is significantly shorter than the replacement. 3× expansion is too low for 1.77% of replacements, 5× for 0.07%, while 10× covers all replacements in our tests.

We compare PII-RADARs trained with 3×, 5×, 10×, 30×, and 100× expansion rates, on 1M documents (about 39% of the data, to speed up experiments). In the table below, we report span-level detection F1 and valid replacements rate, which is the share of replacements that pass a format check for their PII kind described in Appx. E.2 We find that frame expansion does not significantly impact replacement quality, while expansion rates above 10× appear to reduce precision of PII detection. Effectively, 10× lets us cover all replacements while retaining good detection performance. Full results are in Tab. 12.

Expansion RatePII Detection F1Rate of Valid Replacements
3×0.8450.921
5×0.8350.900
10×0.8350.899
30×0.7180.880
100×0.8030.896

We further compare the length of PII-RADAR’s output with the length of the target replacement when using 10× expansion rate. 91%-96% of outputs have the correct length, about 2%-5% are too short, and 2%-4% too long. Both shorter and longer outputs typically differ from the target by a single token (median: 1; mean: 1.3-1.5 for shorter outputs and 1.13-1.65 for longer outputs), for example, “427-957-01830”“427-957-0830” or “5105100000060328”“51051000000060328” (one zero too much).

Dataset# ReplacementsCorrect length (%)Too short (%)Too long (%)
Long Stories (Short)112,00094.53.32.2
Nemotron67,14796.02.41.6
SPY legal15,73894.53.91.6
SPY medical18,31694.93.51.6
REDACT145,03595.03.31.8
pii-masking-300k112,58491.05.43.6

We added a new Appx. C.7 (Tab.13) with full details.

Q4: Are the leakage gains observed for Gemma3 quantitatively similar to OLMo2, and does the trend hold at 7B scale or with fewer training epochs?

Leakage gain is similar between Gemma3, OLMo2 1B, and OLMo2 7B (Figs. 19 and 21). The Pearson correlation of the MIA Leakage Gain is 0.77-0.85, and Fig. 19 shows that the magnitudes are also close in value. These results are for one training epoch.

With fewer than 20 epochs we do not extract enough missed PII to get stable Extractability Gain estimates, see Tab. 16. Since PII-RADAR misses the fewest PII, at early epochs its Gain rests on a small number of extracted PII. At epoch 10, one more extracted PII would move the Gain from 0.22 to 0.33 (by 50%), while at epoch 20, the results are more stable, with changes of at most 0.01. We acknowledge that this is a limitation of our study.

With OLMo2 7B, extraction becomes stable and measurable by epoch 7. There, baseline scrubbers reach extraction gain of 2.08–3.24×, while PII-RADAR stays below 1 (0.72× [0.58, 0.89]). We added the results to Tab.16.

References

[1] Carrell, D. et al. (2013). Hiding in plain sight: use of realistic surrogates to reduce exposure of protected health information in clinical text. Journal of the American Medical Informatics Association. pp. 342-348.

[2] Li, Jeffrey, et al. (2024). DataComp-LM: In Search of the Next Generation of Training Sets for Language Models. NeurIPS24

[3] Hayes, Jamie, et al. (2025). Exploring the limits of strong membership inference attacks on large language models. NeurIPS25.

[4] Steier, Amy, et al. (2025). Nemotron-PII: Synthesized Data for Privacy-Preserving AI, 2025. URL: https://huggingface.co/datasets/nvidia/Nemotron-PII.


If the above answers address the Reviewer's concerns, we would greatly appreciate it if they could consider increasing the rating. We are also happy to address any remaining questions.

Reviewer_dh7eOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper points out that the more personal identifiable information (PII) deleted from the training data, the easier it is for the model to remember and leak the missing PII. To address this, the authors propose PII-RADAR, which replaces sensitive information with high-quality, realistic synthetic PII, thereby reducing the privacy leakage risk of the downstream models. The paper identifies privacy risks with practical significance; the method design is practical; and the experimental verification is relatively complete. However, the core idea is not highlighted as distinct from existing methods.

Critical strengths
  1. The paper is no longer limited to evaluating the accuracy of PII detection, but focuses on the impact of cleaned data on the privacy leakage of downstream models. The experiment of randomly deleting PII supports the explanation that "cleaning makes missing PII easier to remember."
  2. Simplify PII detection to binary classification and combine it with a high-quality synthetic identity pool to achieve non-autoregressive replacement, achieving good effects in terms of detection recall rate, replacement quality, running speed, and downstream privacy risk.
Critical weaknesses
  1. The idea of using alternative text to protect sensitive data while trying to maintain data availability has already been a precedent, and the paper does not sufficiently highlight its unique contribution. (Hiding in Plain Sight, 2013; InferDPT, 2025)
  2. The paper lacks an evaluation of downstream model utility. It remains unclear whether its synthetic PII replacement strategy better preserves language modeling performance and downstream task capabilities compared to conventional masking.
Questions for authors

Can the method proposed in the paper maintain the normal learning ability of the model after cleaning the training data?

4 comments
AuthorsOpenReview ↗

Novelty of PII-RADAR; Downstream Performance After PII-RADAR

We thank the Reviewer for recognizing that our work identifies privacy risks with real-world significance, appreciating the practical design of our method and complete experimental validation. We address all the comments one by one below:

W1: The idea of using alternative text to protect sensitive data while trying to maintain data availability has already been a precedent, and the paper does not sufficiently highlight its unique contribution. (Hiding in Plain Sight, 2013; InferDPT, 2025)

InferDPT [1] protects sensitive data against untrusted LLM providers by replacing every prompt token. In contrast, PII-RADAR protects PII left unscrubbed from increased leakage by replacing detected PII with high-quality surrogates. Hiding in Plain Sight (HIPS) [2] and its follow-up studies [3,4] ask whether humans or machine learning models can distinguish real from the surrogate PII directly in released text whereas we focus on leakage from trained LLMs. Furthermore, HIPS uses formatting rules for each PII type, requiring accurate type classification. This matters because current scrubbers mistype 25-39% of detected PII (Tab. 4). PII-RADAR avoids these errors, and 92.3% of its replacements are type- and format-valid.

Overall, our work introduces the following novel contributions:

  1. We show that scrubbing increases leakage of missed PII from trained LLMs as detection rates improve, and that existing rewriting methods do not prevent it.
  2. We introduce PII-RADAR, which selects PII from profiles and adapts it to the original context without explicit type classification or hand-crafted formatting rules.

We added discussion of HIPS and InferDPT to Sec. 2, and clarified the novelty of PII-RADAR relative to prior work in the Introduction, Sec. 2, 3.3, and 4.

W2: The paper lacks an evaluation of downstream model utility. It remains unclear whether its synthetic PII replacement strategy better preserves language modeling performance and downstream task capabilities compared to conventional masking. Q1: Can the method proposed in the paper maintain the normal learning ability of the model after cleaning the training data?

We trained OLMo 150M models from scratch on 1.5B tokens from DCLM dataset [5], for 4 epochs, following [6,7], in the three settings: 1) No scrubbing; 2) scrubbed with PII-RADAR; 3) masked with PII-RADAR (Mask). We compare the impact of synthetic PII against masking, at a matching detection rate. We follow the evaluation protocol of [6], train three LLMs per setting, report perplexity on an unscrubbed held-out DCLM set, and an average over ten benchmarks.

InterventionPPL ↓Avg. benchmark ↑
No scrubbing29.04±0.1340.9±0.3
PII-RADAR29.82±0.2142.0±0.4
PII-RADAR (Mask)30.32±0.1541.3±0.6

Models in all three settings show similar downstream task performance. Masking remains a worse strategy, with perplexity rising by 4.41% vs 2.69% for PII-RADAR. We also compute perplexity on scrubbed held-out text, focusing only on tokens that are left unchanged after scrubbing.

ScrubberPPL (scrubbed LLMs)PPL (unscrubbed LLM)
PII-RADAR29.10±0.2229.12±0.10
PII-RADAR (Mask)29.09±0.1129.72±0.12

On scrubbed data that matches the training distribution of PII-RADAR, its perplexity matches that of the unscrubbed LLM, suggesting that normal learning ability is preserved. We added the results to our paper, see Appx. C.6 for the detailed experimental setup and Tab.10-11 for full results.

References

[1] Tong, Meng, et al. InferDPT: Privacy-Preserving Inference for Closed-box Large Language Model. IEEE Transactions on Dependable and Secure Computing 22.5 (2025): 4625-4640.

[2] Carrell, D. et al. (2013) Hiding in plain sight: use of realistic surrogates to reduce exposure of protected health information in clinical text. Journal of the American Medical Informatics Association. pp. 342-348.

[3] Carrell, D. et al. (2020). Resilience of clinical text de-identified with "hiding in plain sight" to hostile reidentification attacks by human readers. Journal of the American Medical Informatics Association. pp. 1374-1382.

[4] Carrell D. et al. The machine giveth and the machine taketh away: a parrot attack on clinical text deidentified with hiding in plain sight. J Am Med Inform Assoc. 2019 Dec 1;26(12):1536-1544. doi: 10.1093/jamia/ocz114. PMID: 31390016; PMCID: PMC6857511.

[5] Li, Jeffrey, et al. DataComp-LM: In Search of the Next Generation of Training Sets for Language Models. NeurIPS24

[6] Magnusson, Ian, et al. Datadecide: How to predict best pretraining data with small experiments. arXiv preprint arXiv:2504.11393 (2025).

[7] Muennighoff, Niklas, et al. Scaling data-constrained language models. NeurIPS23.


If the above answers address the Reviewer's concerns, we would greatly appreciate it if they could consider increasing the rating. We are also happy to address any remaining questions.

As the rebuttal period is coming to an end, we would like to thank the Reviewer once again for their constructive feedback. To address their comments, we have additionally:

  1. Highlighted our unique contribution: PII-RADAR selects PII from profiles, without explicit type classification or hand-crafted formatting rules, in contrast to prior work.
  2. Clarified our positioning with respect to InferDPT [1] and HIPS [2]: these protect sensitive data in released text, while we study leakage of missed PII from trained LLMs. We show that the leakage increases as detection rates improve.
  3. Showed that that PII-RADAR preserves downstream performance, while masking hurts perplexity.

We believe that these new additions have substantially strengthened the paper and would greatly appreciate it if the Reviewer could consider increasing their score. We are happy to answer any further questions.

Thank the authors for their efforts to address the reviewers’ concerns and improve the paper. Having considered the other reviewers’ comments and the authors’ responses, I am willing to appropriately increase the rating.

My remaining concern is whether synthetic replacements preserve semantic compatibility with the unchanged context. Appx. C.6 shows comparable overall performance, but perplexity and the ten benchmarks may not capture disruptions to specific contextual relationships. For example, replacing “8th September 1946” with “21st October 1961” in Table 18 could create inconsistencies with references to age or life events elsewhere in the document. (e.g., he was born on 8th September 1946 and retired at the age of 65 in 2011.) The profile mechanism promotes consistency among replacement values but does not necessarily ensure their compatibility with the unchanged surrounding context.

Could the authors clarify whether such contextual constraints are considered during replacement and whether semantic consistency has been directly evaluated? The clarification on this point would not affect my score, but would help me better understand the method’s practical utility and applicability in real-world settings.

Thank the authors for their efforts to address the reviewers’ concerns and improve the paper. Having considered the other reviewers’ comments and the authors’ responses, I am willing to appropriately increase the rating.

We are grateful for the increase in the rating and are glad that our answers helped to resolve the concerns. Thanks to the Reviewer’s feedback, we highlighted our novel contributions, and showed that downstream task capabilities are preserved when the data is cleaned with PII-RADAR, which further strengthens our work.

My remaining concern is whether synthetic replacements preserve semantic compatibility with the unchanged context. Appx. C.6 shows comparable overall performance, but perplexity and the ten benchmarks may not capture disruptions to specific contextual relationships. For example, replacing “8th September 1946” with “21st October 1961” in Table 18 could create inconsistencies with references to age or life events elsewhere in the document. (e.g., he was born on 8th September 1946 and retired at the age of 65 in 2011.) The profile mechanism promotes consistency among replacement values but does not necessarily ensure their compatibility with the unchanged surrounding context. Could the authors clarify whether such contextual constraints are considered during replacement and whether semantic consistency has been directly evaluated? The clarification on this point would not affect my score, but would help me better understand the method’s practical utility and applicability in real-world settings.

We currently do not implement any mechanisms to cater to these constraints, thus we did not directly evaluate the semantic consistency between the surrounding context and the inserted replacements. We ensure semantic consistency within profiles, and train PII-RADAR to insert replacements that follow the format and type of the original PII.

Our observation is that most of the PII is largely independent from the document context (for example: IP address or Tax ID). Special cases when these dependencies occur (like the mentioned example of dates) could potentially be addressed by training PII-RADAR to additionally edit non-PII text to keep the semantic consistency between replacements and the context, or by implementing additional logic when selecting profiles for replacements. We leave these extensions as our future work.

We are happy to answer any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal

We thank the Reviewers for their constructive feedback, and the ACs and PCs for their time in overseeing the reviewing process. We appreciate the Reviewers’ recognition of the real-world significance of the problem studied in our work (Reviewer dh7e), the well-designed and convincing nature of our experiments (Reviewers nv8r & HPsE), and the practicality and efficiency of PII-RADAR (Reviewers dh7e & HPsE).

Overall, our work introduces the following novel contributions:

  1. We show that scrubbing increases leakage of missed PII from trained LLMs as detection rates improve, and that existing rewriting methods do not prevent it.
  2. We introduce PII-RADAR, which selects PII from profiles and adapts it to the original context without explicit type classification or hand-crafted formatting rules.
  3. PII-RADAR significantly reduces leakage of missed PII, has the highest recall on five of seven detection benchmarks, and is 3-10× faster than other rewriting scrubbers.

Based on the Reviewers’ feedback, we were able to significantly improve our paper. Below, we summarize the main highlights of our rebuttal:

  1. Downstream utility evaluation (Reviewer dh7e): we showed that PII-RADAR preserves downstream performance, while masking hurts perplexity.
  2. Replacement strategy ablations (Reviewers nv8r & HPsE): we showed that replacement consistency, diversity, and quality matter for protecting missed PII from leakage. Our PII-RADAR offers protection similar to oracle-level replacements.
  3. CTC frame expansion ablations (Reviewer nv8r): we showed that 10× expansion covers all tested replacement lengths, 91%-96% of PII-RADAR’s outputs match the target length, and the expansion rate does not significantly affect replacement quality.
  4. MIA Leakage Gains in new settings (Reviewer nv8r): we showed that MIA Leakage Gains trends hold for a bigger model, other backbones, and PII densities.

We improved positioning of our work against HIPS [1] and InferDPT [2]. We restructured Sec. 4, expanded Sec. 5 with timing results, and added Appx. C.6-9, F.4-5, Figs. 10-13, 21-24, and Tabs. 3, 4, 10-17 with new results.

We believe these changes address the Reviewers’ concerns and strengthen the paper. If there are any remaining questions, we are happy to answer them.

References

[1] Carrell, D. et al. (2013). Hiding in plain sight: use of realistic surrogates to reduce exposure of protected health information in clinical text. Journal of the American Medical Informatics Association. pp. 342-348.

[2] Tong, Meng, et al. (2025). InferDPT: Privacy-Preserving Inference for Closed-box Large Language Model. IEEE Transactions on Dependable and Secure Computing 22.5 (2025): 4625-4640.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.