- 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
- 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.
- 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
- 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.
- 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
- 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.
- 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?
- 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?
- 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.
- 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
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.
| Claim | Share | #matching / # total replaced spans | Source |
|---|---|---|---|
| Privasis Cl. 4B writes REDACTED for phones | 29.0% | 10,880 / 37,537 | Tab. 19 |
| Anonymizer 4B writes XXXX XXXX XXXX XXXX for cards | 26.3% | 3,374 / 12,850 | Tab. 19 |
| OpenPipe e-mails at example.com/.net/.org | 91.9% | 77,401 / 84,199 | Fig. 15; Tab. 19 |
| OpenPipe valid card numbers | 0.0% | 0 / 6,436 | Fig. 15; Tab. 19 |
| Valid, not a placeholder: PII-RADAR (ours) | 92.3% | 250,661 / 271,672 | Figs. 5, 15 |
| Valid, not a placeholder: Anonymizer 4B | 38.2% | 102,282 / 267,625 | Figs. 5, 15 |
| Valid, not a placeholder: Privasis Cl. 0.6B | 1.4% | 2,106 / 145,284 | Figs. 5, 15 |
| Valid, not a placeholder: Privasis Cl. 4B | 0.7% | 1,072 / 153,071 | Figs. 5, 15 |
| Ten most frequent values: PII-RADAR (ours) | 0.07% | 192 / 271,672 | Figs. 5, 14; Tab. 19 |
| Ten most frequent values: Privasis Cl. 0.6B | 69.5% | 100,943 / 145,284 | Figs. 5; 14, Tab. 19 |
| Ten most frequent values: Privasis Cl. 4B | 74.1% | 113,446 / 153,071 | Figs. 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.
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.
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:
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:
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):
We added these results to Sec. 5.3 (Tab.4).
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.
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.