- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper studies how hospitals can work together on rare-disease diagnosis when patient text cannot be shared. The authors argue that sending text is not allowed, and that sending raw model states can still leak private data. The main challenge is to pass useful diagnostic information between hospitals while limiting this leakage, including when hospitals use different LLMs. The authors propose to send a short, trained latent block instead of text or raw states, and they build a benchmark that measures both accuracy and leakage. They report accuracy close to text sharing, with much lower leakage than sharing raw model states.
Critical strengths
S1. The method keeps most of the accuracy of text sharing while leaking much less.
- Accuracy is only slightly below a trained host that reads the records directly (Table 21).
- Leakage is far lower than when raw KV caches are sent, under all five attacks (Tables 3 and 29).
- A control with the same training and size leaks much more, so the gain is not just from compression (Table 22).
Why this matters to my score: This is the core result and the main reason I lean towards acceptance.
S2. The experiments are thorough.
- Each attack has a control where nothing is sent (Table 3).
- Trained controls show that evidence from other hospitals helps (Table 23).
- The results hold across settings and LLM families, at a small extra cost (App. C–D, Table 2).
Why this matters to my score: These controls make the results believable, so I see the weaknesses below as fixable.
Critical weaknesses
W1. The privacy claims go further than the attacks support.
- The blocks can be recomputed from known inputs and public weights, so an attacker could check a candidate record directly. This attack is not tested (Sec. 3.1, App. F.3).
- Only a passive listener is studied. A host or hospital that sends its own queries is not, and the query itself is sent in plain text (Sec. 3.1).
- There is no noise or differential privacy, and under learned attacks leakage is still above the level seen when nothing is sent (Sec. 4.5, Table 3).
Why this matters to my score: Privacy is half of the contribution. If the simple attack works, the membership result fails. This is why my score is 3, not 4.
W2. Part of the accuracy gain comes from training, and the benchmark is easier than the real setting.
- All test diseases appear in training, but the main baselines are untrained. On Llama, a trained model that sees only the query already beats them; on Qwen, LatentDx shows no clear gain (Tables 1 and 23).
- Two of the data sources probably share cases, and duplicates are removed only by exact match (App. A.1).
- The module is trained on the hospitals' "private" records, and each hospital already covers almost all diseases (Sec. 3.2, Table 5).
Why this matters to my score: It is unclear how much of the gain comes from the method, and whether it would hold in real hospitals.
Questions for authors
- In the membership test, is the query case removed from retrieval? Is any part of block construction random? How well does the "compute and compare" attack in W1 work?
- Why is it acceptable to send the query's phenotypes in plain text but not the retrieved records? How would you handle a host that sends its own queries to probe the hospitals?
- How do baselines that share only a disease label or a local diagnosis perform (for example, Struct. RAG + LoRA), and what do they leak?
- How well does the method work on diseases that are in the hospitals' records but were not used to train the module?
2 comments
Response to Reviewer 5ymK [2/2]
W2: Source of the Gain and Benchmark Realism
We thank the reviewer for this careful analysis and address its three points in turn.
Role of training. We compare LatentDx with trained query-only controls that receive the same supervision but no hospital evidence, and the revision brings these trained comparisons into the main text (Section 4.4, Table 3). Removing the evidence lowers accuracy from 0.73 to 0.64 on Llama and from 0.64 to 0.41 on Qwen (Appendix E.2), so the hospital evidence adds substantially on top of training. To also match the latent budget, in the revision we train a query-only model that sends three blocks of the same size without retrieved cases. Averaged over three training seeds, LatentDx exceeds it by 0.07 accuracy and is better on all three metrics (Appendix E.2, Table 30). Trained references that read the shared evidence directly are discussed in our response to Q3.
Near-duplicates across sources (new). Appendix A.1 now reports a near-duplicate audit. No test case shares its HPO set with a training case of the same disease, and 11 of the 401 test cases (2.7%) have one with HPO Jaccard similarity of at least 0.9. Removing these 11 cases changes the accuracy of every method by at most one point. Likewise, when hospitals never return a case with Jaccard of at least 0.9 (or 0.8) to the query, the accuracy of LatentDx changes by less than one point.
Disease coverage across hospitals (new). In the revision each source cohort becomes its own hospital (Appendix D.4). The three hospitals differ in size (5,894 / 804 / 522 cases), disease coverage (219 / 38 / 64 diseases) and annotation practice. LatentDx is best on every cohort and metric, with 0.76 overall accuracy against 0.47 for the strongest baseline. The disease-skewed partition already in the original Appendix D.3 concentrates each disease in one hospital, which holds on average 65% of its cases, and LatentDx keeps the highest accuracy there on both backbones.
Records never used in training (new). We retrain the distiller without any Zenodo record, removing them from the training queries, the validation queries and the training-time hospital databases. Zenodo then joins as its own hospital only at test time. LatentDx still has the highest accuracy of all methods, both on the Zenodo patients (0.55 vs. 0.41 for the strongest baseline) and over all patients (Appendix D.4, Table 21). Records that were never used to train the module therefore remain usable at test time.
Q3: Baselines Sharing Only a Disease Label or a Local Diagnosis
We thank the reviewer for this suggestion. Struct. RAG already shares only each hospital's nearest-case diagnosis and its similarity score, which amounts to a local nearest-neighbor diagnosis (0.42 accuracy on Llama, Table 1).
Utility with host tuning (new). As suggested, we trained Struct. RAG with host LoRA (Appendix D.9, Table 27). It reaches 0.74 accuracy against 0.73 for LatentDx, with a lower macro F1, while tuning 24.3M parameters of the host LLM itself, whereas LatentDx keeps every backbone frozen and trains only its 9.45M-parameter distiller.
What it leaks. Its utility is therefore comparable, but it sends each retrieved patient's diagnosis in plaintext, so any observer of the messages reads it. From LatentDx's blocks, the inversion decoder recovers the retrieved disease in 38% of cases on Llama, against 27% when nothing is transmitted (Table 4).
Q4: Records and Diseases Unseen in Training
We thank the reviewer for this question and conducted additional experiments to answer it. We distinguish records that were not used in training from diseases that were absent from training altogether.
Records unseen in training (new). In the setting of W2, Zenodo is held out from training entirely. No Zenodo record is used as a training query or stored in the training-time hospital databases, and Zenodo joins as its own hospital only at test time. LatentDx still keeps the highest accuracy, both on the Zenodo patients and over all patients (Appendix D.4, Table 21).
Diseases unseen in training (new). We further built a strict split (Appendix D.6) in which 58 diseases are removed from all training queries and training-time hospital databases, while the test-time databases still contain them. We find this setting genuinely challenging, since the interface never saw these diseases during training. To address it, we add name-anonymized copies of half of the training queries, in which every disease name is consistently replaced by an unrelated one so that the answer must be read from the evidence rather than recalled. With this change, LatentDx reaches 0.47 accuracy on unseen diseases, nearly double its original accuracy, while accuracy on seen diseases remains high (Table 23).
Response to Reviewer 5ymK [1/2]
We sincerely thank the reviewer for the careful review and for recognizing the utility–leakage result as the main reason for leaning towards acceptance. We also appreciate the reviewer's clear articulation of which concerns matter most for the assessment, in particular the scope of our privacy claims. We took this feedback seriously and address it first below. New experiments are marked (new) below and in blue in the revised PDF.
W1 & Q1: Scope of the Privacy Claims and the Compute-and-Compare Attack
As the reviewer identified the scope of our privacy claims as the key remaining issue, we implemented the proposed attack, developed a differentially private variant, and now state precisely which claims hold under which attacker.
Answers to Q1. (1) The candidate is not removed from retrieval. By design, the membership attack queries with the candidate's own phenotypes, so a member hospital retrieves the candidate itself, which favors the attacker. (2) Block construction involves no randomness. Retrieval, the frozen backbone and the distiller are deterministic at inference, so the same query always yields the same block. (3) As the reviewer anticipated, the compute-and-compare attack is effective. With all weights and the complete candidate record, it reaches a TPR of 1.00 at 1% FPR.
What this attack shows. We agree with the reviewer. Against an attacker who holds a candidate's complete record and all weights, the deterministic block does not hide membership, as is also the case for Raw KV and LatentMAS. The attack verifies a record the attacker already has rather than recovering unknown content. With 90% of the candidate's phenotypes its TPR falls to 0.43, and with 70% it falls to 0. Our reconstruction, attribute and inversion results, which measure the recovery of unknown content, therefore stand, while the membership result does not hold against this attacker. We have revised the threat model to state that the attacker may know all weights, including the interface, and qualified the membership row of Table 4 accordingly.
Protecting membership with LatentDx-DP (new). To protect membership against such an attacker, and to address the reviewer's point on noise and differential privacy, we developed LatentDx-DP (Appendix F.5). Each hospital still sends a latent block of the same size, and the host LLM is unchanged. Only the distiller's input changes, and the distiller is retrained on such inputs with the same recipe. Instead of the raw record, the hospital encodes a differentially private summary of its local evidence, released with the analytic Gaussian mechanism (Balle & Wang, 2018). Each query satisfies (4.1,10−5)-DP with respect to adding or removing one record, and the block is post-processing, so the guarantee holds even against an attacker with all weights. The table below summarizes membership leakage under both attackers on Llama-3.2-3B (TPR at 1% FPR, chance is 0.01).
LatentDx-DP reduces compute-and-compare membership to near chance at a cost of 0.01 accuracy.
Residual leakage under learned attacks. Beyond membership, the reviewer also noted that leakage under the learned attacks remains above the no-transmission control. This residual is concentrated in disease-level information that the diagnostic task requires, while the inversion decoder's recovery of the retrieved record's phenotype terms stays close to the control (HPO micro-F1 0.27 vs. 0.23 on Llama-3.2-3B, Table 37).
Q2: Plaintext Query and Actively Probing Hosts
We thank the reviewer for this question.
Plaintext query. The query phenotypes belong to the patient of the diagnosing hospital that initiates the consultation, whereas LatentDx protects the historical records held by the responding hospitals. As stated in Section 3.1, we assume that sharing the query is authorized, as in a standard inter-institutional consultation.
Actively probing hosts. A host that issues its own probing queries is the setting of our membership attacks, in which the attacker chooses the candidate's phenotypes as the query (Table 4 and Appendix F.5). Without the complete candidate record, even the compute-and-compare attack weakens sharply (see W1). With the exact record, LatentDx-DP limits what each query reveals about any single record to (4.1,10−5)-DP (Appendix F.5). Since the guarantee is per query, it composes over the queries a hospital answers, which each hospital can bound with a query budget.