acceptodds
Under review as a conference paper at ICLR 2027

GenoFA: quantifying privacy leakage of genomic language models under dependent data

Abstract

Genomic language models (gLMs) enable large-scale training on personal genomic variants and advance personalized medicine. However, they also introduce unique privacy risks. Human genomes are inherently correlated across familial relationships due to Mendelian inheritance. Such correlation breaks the assumption of independence required by most differentially private algorithms, preventing straightforward adoption in the context of genomic privacy. We leverage Mendelian inheritance to create a family-aware genotype reconstruction attack, called GENOFA. gLMs can impute on non-training data points, but this should not count as privacy leakage as it occurs for seen and unseen data alike; we propose imputation-corrected reconstruction risk, a novel metric to remove the model’s imputation impact from privacy accounting. Our results demonstrate that familial correlations can amplify individual privacy leakage, motivating the need for family-level privacy protections. We therefore propose to treat the family as an atomic unit for privacy protection and apply client-level differential privacy to bound the privacy leakage of the entire family. We demonstrate that with a privacy budget of 6.0, client-DP improves the model utility by 22.9%, reducing the RMSE from 0.35 (group-DP) to 0.27. Together, these contributions provide a unified framework that identify, quantify, and mitigate privacy leakage in gLMs despite the inclusion of familially related individuals.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine learning, differential privacy, dependent data, genomics language models

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_BaqZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies genotype reconstruction from genomic language model trained on related individuals. GenoFA combines model predictions with Mendelian transmission probabilities to infer a missing genotype. They measure the separation between reconstruction scores from models trained with and without a target's entire family, aiming to distinguish training-related leakage from imputation. They evaluate a fine-tuned Nucleotide Transformer and a MLP including real trios and larger families with simulated offspring. They then apply family-level client-DP and audit family membership with group-LiRA.

Critical strengths
  1. This paper addresses a privacy issue: relatives can contribute information about the same individual and genomic training can divide one person's data across several records. Their family-aware reconstruction attack connects this problem to a concrete inference procedure. Evaluating both a genomic model and a MLP helps establish that the concern extends beyond one architecture.

  2. The comparison with nonmember reconstruction is useful. An attacker can recover genotypes through imputation even when the family did not contribute to training. This gives a more informative assessment than reporting reconstruction accuracy on training examples alone.

Critical weaknesses
  1. The privacy-utility comparison does not match the full family-level privacy guarantee. Under the bound used in this paper, Client-DP achieves better reported utility despite the tighter delta, which is encouraging, but the comparison does not establish the claimed gain at the same privacy guarantee. The authors also describe dependence as breaking an independence assumption of standard DP. Standard DP bounds changes under its specified adjacency without requiring independent records. The distinction between record-level protection, family-level protection, and inferential privacy needs to be stated correctly.

  2. The evidence for leakage increasing with family size relies on a substantial change to the biological setting. The author explains that this creates diversity within short windows, but do not establish how this choice affects the attack advantage. The real-trio experiments support a narrower empirical claim. Without a sensitivity analysis, the stronger results on simulated families provide limited evidence about the effect of family size in human cohorts.

Questions for authors
  1. Can you compare client-DP and group-DP at matched family-level epsilon and delta, and specify the adjacency and accountant used for each? Please also distinguish the limitations of record-level protection from the validity of the standard DP guarantee.

  2. Does GenoFA retain its advantage, and does leakage increase with family size, when offspring generation uses a human-scale recombination rate? A sensitivity check within the existing simulation setup would help determine how much of the reported effect depends on this choice.

4 comments

Weakness 1: Record-DP limitations against the DP definition, and DP against inferential privacy:

  1. Our claim in the abstract concerns DP training algorithms (such as DP-SGD with privacy budget ε), where each data point’s gradient is clipped, and the guarantee is at record granularity. An individual’s genomic information that spans several correlated records (a family) is then protected only at the weaker group-privacy level (kε), not at the nominal ε. The DP definition requires no independence, as client-DP itself shows, since all of a client’s data is correlated. We agree the wording introduces this confusion and have replaced “assumption of independence required by most differentially private algorithms” with a statement about commonly used algorithms that protect record-level DP and are insufficient here because of the correlation structure.
  2. Inferential privacy bounds how far an adversary’s belief about an individual moves from its prior, before release, to its posterior, after release [1]. DP instead bounds the change in the mechanism’s output when one unit of input changes under the stated adjacency, which is a bound between two posteriors [1]. Section 5 now talks about the scope of privacy protection of DP and how inferential privacy is out of the scope of this work. Appendix A.12.4 discusses Pufferfish privacy, a form of inferential privacy, Bayesian DP as an instantiation of Pufferfish, and the conditions under which Bayesian DP and DP coincide.

[1] Shiva P. Kasiviswanathan and Adam Smith. On the ‘semantics’ of differential privacy: A Bayesian formulation. Journal of Privacy and Confidentiality, 6(1), Jun. 2014. doi: 10.29012/jpc.v6i1.634.

Weakness 1 and Question 1: Matching epsilon and delta for group-DP and client-DP: The table below gives matching family-level epsilon and delta for client-DP and group-DP (epsilon and delta scaled from record-DP) with maximum family size capped at 3 (k=3); the manuscript reported the case where client-DP uses the stricter delta in Figure 3a. Both record-DP training with DP-SGD (Algorithm 1) and client-DP training (Algorithm 2) use a Rényi DP accountant. Under record-DP, adjacency is D, D′ differing in one record (Definition 1). Under client-DP (Definition 5), D and D′ differ in all data points of the entire family, following prior implementations (McMahan et al., 2018; Geyer et al., 2018) that treat all of a family’s data points as one client.

Group-DP and client-DP comparison (maximum family size 3):

εrε_rεΦε_Φ and εgε_ggroup-DP RMSE, (δgδ_g) from Figure 3aclient-DP RMSE, (δΦδ_Φ) from Figure 3agroup-DP RMSE, (δgδ_g) matched δδ
1/31.00.634 +/- 0.029 (4.34e-05)0.735 +/- 0.033 (1e-5)0.815 +/- 0.034 (1e-5)
1.03.00.488 +/- 0.021 (1.11e-04)0.427 +/- 0.014 (1e-5)0.506 +/- 0.029 (1e-5)
2.06.00.353 +/- 0.022 (6.30e-04)0.268 +/- 0.010 (1e-5)0.591 +/- 0.018 (1e-5)
3.09.00.393 +/- 0.028 (4.25e-03)0.263 +/- 0.014 (1e-5)0.404 +/- 0.019 (1e-5)

Weakness 2 and Question 2: Does larger family size widen the gap between GenoFA and naive leakage, and what happens at human-scale recombination rates?

Both points are fair, and we agree the rate is a substantial deviation from biology. We ran two follow-up experiments: one isolating family size at a fixed recombination rate, one varying the recombination rate. Our simulations use 10,000 cM/Mb against a human-scale rate of roughly 1 cM/Mb because the short windows permitted by the gLM’s context length do not, at human rates, yield enough haplotype diversity to separate in-distribution from out-of-distribution statistically. The family-aware leakage is nonetheless not a simulation artifact: the real-trio experiments (Table 3, 1000 Genomes) use actual average human recombination, and the target’s relatives being in training still raises reconstruction accuracy there. The two experiments are updated in Appendix A.8 and Appendix A.21.

Experiment 1 holds the recombination rate at 10,000 and varies the number of correlated family members k, with all children’s genomes simulated at that rate; the high rate ensures enough unique genomic sequences for members and non-members. The gap between GenoFA and naive leakage widens on average as family size grows. Results are averaged across all 6 attack sites used in the manuscript.

Number of kidsAverage GenoFA leakage measureAverage naive leakage measureGap
10.5910.579+0.0114
20.5830.568+0.0149
40.5740.557+0.0174
80.5570.538+0.0194
160.5510.532+0.0192

Experiment 2 fixes one child per family and varies the recombination rate from 1 to 10,000. The effect is minor, so the comparison in the paper is not invalidated. We acknowledge that recombination rates vary across genomic regions, and that estimating them from HapMap rather than using an average rate of 1 may produce more pronounced differences for longer sequences, as noted among the simplifications in Appendix A.2. This is worth exploring in future work.

Recombination rate (cM/Mb)GenoFA accuracy %, membersNaive accuracy %, membersGenoFA accuracy %, non-membersNaive accuracy %, non-members
196.990.672.663.7
1096.990.672.663.7
10097.090.972.963.7
100097.191.072.463.0
1000097.191.370.361.2

Thank you again for your detailed review. As the discussion closes today, we wanted to check whether our responses have addressed your questions. We compared client-level and group-level DP at matched ε and δ, and clarified the scope of the DP guarantee. We also added a recombination-rate sensitivity analysis, which shows that GenoFA retains its advantage at human-scale rates. If any concerns remain, we would be glad to address them. Otherwise, we would be grateful if you would consider updating your score.

Thank you for the rebuttal and the additional experiments. I appreciate the authors' detailed responses to my questions, particularly the new tables provided.

After considering the rebuttal and the other reviewers' comments, I think my current score is still appropriate, so I will keep it unchanged.

My primary concerns regarding the theoretical alignment of the privacy-utility comparison and the biological realism of the recombination rates in the simulations remain partially unaddressed. The new experiments do not fully change my assessment of the paper's core claims.

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

This work investigates the privacy risks for Genomic language models (gLMs). In particular, it argues that in human genomes there is a strong correlation between relatives and family members due to Mendelian inheritance, which invalidates the per-record differential privacy protection and limits the effectiveness of some of the attacks. To address this, this work proposes a GenoFA a family-aware genotype reconstruction attack that exploits this correction between family members and apply the model inversion likelihood attacks with the Mendelian genetic prior. Additionally, this work proposes a differential privacy defense based on DP-SGD and tailored to family level protection.

Critical strengths
  • The analysis of the privacy risks of gLMs is a relevant and important problem, as gLMs are increasingly fine-tuned on personal variant genomes, and family members are routinely present in such cohorts.
  • It provides an empirical evaluation across 2 models (a gLM and a MLP) and 2 configurations. The same qualitative behaviour appears for both architectures and for both small and large families, so the reported effects are not an artifact of a single setup.
Critical weaknesses
  • The current results do not provide strong enough empirical evidence of the effectiveness of the proposed approaches. For GenoFA (attribute inference), a standard metric such as TPR @ 1%FPR is missing, while group-LiRA is instead evaluated with AUROC and TPR @ 0.1% FPR. Moreover, this attack has access to more information than the baseline making the comparison unfair (genotypes and the labels of the relatives). The same is true for group-LiRA, which aggregates the scores of several records and is therefore easier to detect than a single record regardless of any genetic correlation. It is unclear whether the improvement is due to the Mendelian inheritance correlation or some other factors.
  • It does overstate the novelty and contribution. In particular, it does not cite some important related work that finds similar issues in reconstructing the genotypes of relatives using Mendel’s laws (e.g. [1]). Moreover, the distinction between attribute inference and imputation, which the paper presents as the motivation for a novel metric is a well known phenomenon as described by [2], which is also not cited. Additionally, the DP-SGD with family-structure aware adaptation of DP for family-level privacy seems quite straightforward.

[1] Humbert, Mathias, et al. "Addressing the concerns of the lacks family: quantification of kin genomic privacy." CCS 2013

[2] Jayaraman, Bargav, and David Evans. "Are attribute inference attacks just imputation?" CCS 2022

Questions for authors
  1. What is the result of GenoFA when the auxiliary information comes from an unrelated trio instead of the real family? This would analyze whether the gain comes from the Mendelian correlation or from the additional observed data.
  2. Can the comparison between GenoFA and the naive attack also report TPR @ low FPR, as done for group-LiRA in Table 5?
  3. How does GenoFA differ from [1], beyond replacing the known genotype-phenotype association with a trained gLM? And how does the family-level protection differ from the adaptation attributed to Almadhoun et al. (2020) in Appendix A.12.4?
2 comments

Weakness 1 and Question 1: Does the improvement come from Mendelian correlation or from more data points? It comes from the Mendelian correlation: unrelated trios do not improve attack success. Below we report accuracy for the naive attack, GenoFA with real trios, and GenoFA with an unrelated trio (control), for rs53576 (large) in MLP mode, averaged over 3 train-test splits and across the same 6 attack sites as the manuscript. Results are updated in Appendix A.20 (referred to in Section 4.2). For members (training data points):

Attack siteNaiveGenoFA (real trio)GenoFA (unrelated trio)
87626850.8720.9950.841
87636800.6790.9480.707
87619710.5710.8040.542
87615910.8230.8280.821
87620140.6820.9340.728
87621250.7300.9680.656

For non-members (non-training data points):

Attack siteNaiveGenoFA (real trio)GenoFA (unrelated trio)
87626850.8120.9810.822
87636800.6550.8820.673
87619710.5050.6570.467
87615910.8120.7800.781
87620140.6700.8820.703
87621250.7000.9090.617

Weakness 1 and Question 2: Fairness of the comparison, and TPR at low FPR for reconstruction attacks

On fairness: GenoFA and the naive attacks both have access to the same inputs (known features and labels of an individual); GenoFA additionally uses the family structure that the naive attack ignores. That is, the only difference is that GenoFA leverages the family structure while the naive attacks treat individuals independently, so the comparison measures exactly what that correlation information buys an adversary. The same holds for GroupLiRA against LiRA, where the family structure is given as meta information to both. The reviewer is right that GroupLiRA exploits group structure rather than the Mendelian correlation specifically; that is by design, since the family is the privacy granularity and the point is to evaluate the family’s aggregate contribution to membership signal, for which the grouping can be arbitrary.

TPR at FPR = 0.1%:

rs53576 (large), MLP, non-DP, TPR at FPR = 0.1%, averaged across all attack sites:

SplitGenoFANaive
Training data0.6640.328
Non-training data0.2060.087

rs53576 (small), MLP, non-DP, TPR at FPR = 0.1%, averaged across all attack sites:

SplitGenoFANaive
Training data0.4930.353
Non-training data0.1650.084

Weakness 2 and Question 3: Novelty and positioning relative to [1] and Almadhoun et al. (2020), and the use of “imputation” in [2]: We thank the reviewer for the references and cite [1] (Humbert et al., 2013) in Appendix A.12.1. Excess leakage from Mendel’s laws is well known and we claim no novelty there. The distinction is that this literature, [1] included, concerns raw genotype and query release, whereas we quantify and bound leakage from trained ML models. The attacker must therefore infer genotypes through prediction likelihood rather than from genetic records, and the setting calls for metrics and guarantees that isolate leakage attributable to training-set membership, which information-theoretic measures for raw genotype publication (for example Humbert et al., 2013) do not capture. To our knowledge, no prior work addresses these questions together while considering Mendelian correlation of genomic data in model training. We therefore (i) define an imputation-corrected metric that isolates the membership-attributable component from the correlation-heightened imputation floor, and (ii) bound that component with a family-level (εΦ,δΦ)(ε_Φ, δ_Φ)-DP guarantee against arbitrary adversaries. We also respectfully disagree that adopting family-level privacy is straightforward: a privacy analysis gives only an upper bound, so we also establish lower bounds through empirical family-level membership inference attacks (Figure 3b). Analysis and attacks together give what is, to our knowledge, the first complete privacy profile for Mendelian-correlated data for ML model training.

Almadhoun et al. (2020) propose family-level protection but calibrate query sensitivity heuristically against observed Mendelian attack success, without accounting for baseline imputation or providing worst-case guarantees. We clip and noise family-level aggregate gradients during training, which yields a formal (εΦ,δΦ)(ε_Φ, δ_Φ)-DP guarantee against arbitrary adversaries.

On “imputation” we use the standard definition from statistical analysis with missingness, replacing missing attributes, and in our case, the model trained on the same data does the replacing. The authors of [2] use the term for inferring missing attributes from auxiliary data. Appendix A.12.5 notes the difference to avoid confusion.

Thank you for the detailed response and the additional experiments, in particular the unrelated-trio control, which addresses my main concern about the source of GenoFA's improvement. I have increased my score accordingly.

Reviewer_LXSEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies privacy leakage in genomic language models (gLMs) when the training data contain genetically related individuals. The authors propose GenoFA, a family-aware genotype reconstruction attack that exploits Mendelian inheritance to infer unknown genomic attributes. They also introduce an imputation-corrected reconstruction risk metric to distinguish leakage associated with training-set inclusion from the model's general imputation capability. To mitigate the identified risks, the paper applies client-level differential privacy (DP), treating an entire family as the unit of privacy protection. Experiments on genomic datasets demonstrate increased reconstruction accuracy when family information is available and an improved privacy-utility trade-off compared with the group-DP baseline.

Critical strengths
  1. The paper addresses an important privacy problem in genomic language models. The study of how familial correlations affect privacy leakage during model training is relevant and practically motivated.

  2. The imputation-corrected reconstruction risk metric is an interesting contribution. Distinguishing training-related leakage from general genotype imputation is useful for understanding the privacy risks of genomic models.

  3. The experimental evaluation covers different model architectures and family configurations.

Critical weaknesses
  1. The novelty of exploiting familial information in genomic privacy attacks is limited. My main concern is the novelty of the proposed family-aware attack. Exploiting genetic relationships and family information to infer sensitive genomic information is not new. In fact, this has been a well-established research direction for more than a decade. For example, Identifying Personal Genomes by Surname Inference (Science, 2013) demonstrated how genetic inheritance and genealogical information can be exploited to re-identify individuals. Preserving Privacy for Distributed Genome-Wide Analysis Against Identity Tracing Attacks (IEEE TDSC, 2023) investigated identity tracing attacks that exploit genetic relationships to identify genomic data owners. More directly related to the proposed genotype reconstruction attack, Deznabi et al., An Inference Attack on Genomic Data Using Kinship, Complex Correlations, and Phenotype Information (IEEE/ACM TCBB, 2018), already studied how kinship, Mendelian inheritance, genomic correlations, and phenotype information can be combined to reconstruct hidden genomic attributes. These studies show that exploiting familial dependencies for genomic inference is a mature research direction. While I understand that GenoFA considers a different setting involving genomic language models, the core idea of incorporating family relationships into genotype reconstruction is not fundamentally new. The paper should more clearly distinguish its technical contribution from these previous studies. In particular, what is new beyond adapting existing family-based inference techniques to gLMs? Is the main novelty the attack itself, the imputation-corrected leakage metric, or the use of family-level DP? I think a more comprehensive discussion of the related literature is necessary to properly position the contribution.

  2. The scope of the differential privacy guarantee is not sufficiently clear. The paper proposes treating the entire family as an atomic unit and applying client-level DP. This is a reasonable approach, and the definition of client-DP is consistent with standard differential privacy when the training mechanism and privacy accounting are implemented correctly. However, I am concerned about the claim in Section 3.5 that protecting the entire family also bounds the privacy leakage of every family member. This statement needs to be qualified. Family-level DP bounds the additional information revealed through the inclusion or exclusion of an entire family's records in the training dataset. It does not necessarily guarantee that an adversary cannot infer the sensitive genomic attributes of an individual. For example, even if a target's family is excluded from training, an attacker may still infer that individual's genotype using publicly available genomic information, population-level correlations, or genetic information from relatives outside the protected family. Such inference may be accurate regardless of whether the target or their immediate family participates in model training. Therefore, the proposed DP mechanism provides a meaningful guarantee about the family's contribution to the trained model, but this is different from protecting the confidentiality of every individual's genomic attributes. The paper should make this distinction explicit, particularly given its claims about protecting individual genomic privacy. Furthermore, the theoretical framework assumes that different families are independent and that their boundaries are known. In practice, genetic correlations can extend beyond immediate family groups. It is unclear how the proposed protection should be interpreted when individuals across different families are genetically related. The authors should discuss whether the claimed guarantee still holds and, more importantly, what it means for genotype inference risks under such dependencies.

  3. The formal privacy analysis and accounting need further clarification. The paper introduces definitions of record-level, group-level, and client-level DP and describes the corresponding training procedures. However, the theoretical contribution appears to rely primarily on applying existing client-level DP mechanisms rather than establishing a new privacy guarantee specifically for dependent genomic data. I did not find a separate formal theorem proving the claimed privacy bound for the implemented training algorithm. In particular, the guarantee depends on correct client-level gradient clipping, Poisson subsampling, Gaussian noise calibration, and composition across training steps. The paper should more explicitly specify the privacy accountant used to obtain the reported privacy budgets and clarify the adjacency definition underlying the guarantee. Another concern is the comparison with group-DP. Section 3.5 describes the group privacy budget mainly through the scaling of epsilon by the maximum group size. However, under approximate DP, delta must also be considered. Although the appendix acknowledges this through a group-privacy expression, the experimental comparison should clearly explain how both epsilon and delta are accounted for to ensure that group-DP and client-DP are compared under equivalent family-level guarantees. I do not think these concerns necessarily invalidate the proposed client-DP mechanism. Rather, the paper should provide a more rigorous and transparent justification of its privacy accounting, especially because theoretical privacy protection is one of its central claims.

  4. The proposed reconstruction risk metric should not be interpreted as a general guarantee of genomic privacy. The imputation-corrected reconstruction risk is interesting, but its interpretation needs to be more precise. Definition 4 measures the probability that an attack assigns a higher reconstruction score when the target is included in training than when the target is excluded. This measures the additional reconstruction advantage associated with training participation, rather than the absolute difficulty of inferring a sensitive genotype. For instance, an attacker might reconstruct a target's genotype with high accuracy regardless of whether the target appears in the training set. In this case, the proposed metric could indicate little additional privacy leakage, even though the individual's genomic information remains highly predictable. This is particularly important in the genomic setting, where family relationships and population correlations may already reveal substantial information about an individual independently of the trained model. The authors should clarify that the proposed metric quantifies membership-dependent reconstruction risk rather than overall genomic attribute confidentiality. It would also be helpful to report absolute reconstruction accuracy alongside the proposed metric to better understand the actual inference risk faced by individuals.

Questions for authors
  1. How does GenoFA differ technically from existing genomic inference attacks that already exploit kinship, Mendelian inheritance, and phenotype information, particularly Deznabi et al. (2018)? What is the central novelty beyond applying such attacks to genomic language models?

  2. Does the proposed client-level DP mechanism guarantee the confidentiality of an individual's genomic attributes, or does it only bound the additional leakage attributable to the family's participation in training? Can the authors clarify the scope of their privacy claims?

  3. Can the authors provide a more explicit formal analysis of the privacy guarantee for Algorithm 2, including the accountant, sampling assumptions, adjacency definition, and handling of both epsilon and delta?

  4. How does the framework handle genetic correlations between different families? Does the DP guarantee remain valid, and what limitations arise when interpreting it as protection against genomic attribute inference?

  5. How should the proposed imputation-corrected reconstruction risk be interpreted when the attack can accurately reconstruct a genotype even without the target being in the training set? Could the authors report absolute reconstruction accuracy together with the proposed metric?

3 comments
AuthorsOpenReview ↗

Novelty, DP analysis and scope

Weakness 1 and Question 1: Novelty and positioning of our work: We claim no novelty in Mendelian genomic inference. Our contribution is family-based inference against a new target: a model trained on genomic data rather than the genotypes themselves. While the Bayesian framework of Mendelian and LD priors is shared, the likelihood function is fundamentally different: Deznabi et al. compute P(phenotype | genotype) from known associations, whereas GenoFA computes P(observed | model(sequence)), which requires querying the released model, tokenizing candidate genotypes into DNA sequences, and extracting the prediction likelihood from a forward pass before combining it with genetic priors. Every work cited concerns released genomic data (surname inference, identity tracing, reconstruction from records); here only the model is released and the adversary has query access, which changes both what an attack must do and what must be protected. Related work and A.12.1 now position the novelty as the threat model and the metric, not Mendelian inheritance.

Weakness 3 and Question 3: DP analysis, and the epsilon and delta values used for group-DP and client-DP: The formal guarantee for Algorithm 2 (accountant, adjacency, sensitivity, sampling): We introduce no new privacy analysis. We instantiate client-level (user-level) DP-SGD from decentralized training (McMahan et al., 2018; Geyer et al., 2018) with the family as the client, which lets us use a mature accountant: the Rényi DP accountant (Mironov, 2017; Opacus, Yousefpour et al., 2021), applied per iteration and composed over T = epochs/p iterations. Adjacency is two datasets differing in all data points of one family (Def. 5), so ε bounds the worst-case leakage for the entire family. Each iteration Poisson-samples families at rate p < 1; each family’s aggregated update is clipped to ℓ2 norm C regardless of family size, and Gaussian noise σC is added, with σ set to meet the target ε. We set δ < 1/N^1.1 following standard implementation, where N is the number of clients. Algorithm 2 (Appendix A.5) and Appendix A.8 are updated accordingly with full details. The comparison with group-DP under equivalent family-level guarantees: in the manuscript we take a conservative approach that favours the group-DP baseline, using a tighter δ for client-DP than for group-DP. Section 4.3 and Appendix A.8 now justify this in detail. Below we give the (εg,δg)(ε_g, δ_g) and (εΦ,δΦ)(ε_Φ, δ_Φ) values behind Figure 3a, together with additional group-DP runs at δ matched to client-DP. Values are macro RMSE +/- std.

εrε_rεΦε_Φ and εgε_ggroup-DP RMSE, (δgδ_g) from Figure 3aclient-DP RMSE, (δΦδ_Φ) from Figure 3agroup-DP RMSE, (δgδ_g) matched δδ
1/31.00.634 +/- 0.029 (4.34e-05)0.735 +/- 0.033 (1e-5)0.815 +/- 0.034 (1e-5)
1.03.00.488 +/- 0.021 (1.11e-04)0.427 +/- 0.014 (1e-5)0.506 +/- 0.029 (1e-5)
2.06.00.353 +/- 0.022 (6.30e-04)0.268 +/- 0.010 (1e-5)0.591 +/- 0.018 (1e-5)
3.09.00.393 +/- 0.028 (4.25e-03)0.263 +/- 0.014 (1e-5)0.404 +/- 0.019 (1e-5)

Weakness 2 and Question 2: What does client-DP protect, and is it confidentiality of the genotype? Client-DP bounds the leakage attributable to a family’s participation in training: it bounds how much the model weights can change when one family is included rather than excluded, a posterior-to-posterior change in the adversary’s belief [1]. It does not bound what an adversary can infer from auxiliary information such as population structure or relatives outside the protected family. That is the domain of inferential privacy (e.g., Pufferfish privacy, Bayesian DP; Appendix A.12.4), which bounds the move from the adversary’s prior belief before the release to the posterior after the release. DP mechanism cannot bound such leakage: if the adversary already knows a target’s genotype from auxiliary data, DP training changes nothing. The reviewer is correct that a target’s genotype may be inferable whether or not their family participated; this is an active concern in bioethics [2] but outside the scope of our work. In this work, we study the leakage of the training data from the trained model. We also avoid “confidentiality,” which typically denotes cryptographic privacy (only authorized parties can access the data) rather than the DP notion that client-DP provides. Sections 3.5 and Section 5 have been updated accordingly.

[1] Shiva P. Kasiviswanathan and Adam Smith. On the ‘semantics’ of differential privacy: A Bayesian formulation. Journal of Privacy and Confidentiality, 6(1), Jun. 2014. doi: 10.29012/jpc.v6i1.634.

[2] Deborah R. Gordon and Barbara A. Koenig. “If relatives inherited the gene, they should inherit the data.” Bringing the family into the room where bioethics happens. New Genetics and Society, 2022.

AuthorsOpenReview ↗

Absolute reconstruction accuracy and correlation between families

Weakness 4 and Question 5: Interpreting the metric, and absolute reconstruction accuracy:

Why the metric subtracts imputation: A perfectly private model answers randomly and is useless, so a privacy-preserving mechanism cannot be evaluated apart from its utility. Utility here is generalization, and imputation, meaning how well the model fills in a missing feature of an unseen data point, measures exactly that. Consider a model trained to map rs334=T to sickle cell disease or trait: observing the disease lets an adversary guess rs334=T whether or not the target was used in training. That inference reflects generalization, not membership, so it must be removed from any quantification of training-data leakage. We agree the metric should be read as membership-driven reconstruction risk rather than overall genotype reconstruction success, and Section 3.5 and the new scope paragraph in Section 5 now say so. The defence is correspondingly not aimed at driving imputation accuracy to zero: as Table 4 shows, client-DP bounds membership-driven leakage (ℒⁱ), not absolute reconstruction accuracy.

Absolute reconstruction accuracy: Table 1 of the manuscript reports this for rs53576 (small, real trios) at site 8761971 for NT on one train-test split. The table below covers all 6 attack sites (NT), averaged over 50 train-test splits. The excess risk column isolates the gain from the family-aware attack; the gap between the two rows is the membership-driven component (member − non-member).

SplitNaiveGenoFAExcess risk (GenoFA − Naive)
Train (member)0.7100.793+0.083
Imputation (non-member)0.6330.689+0.056

Question 4: Correlation between different families: The family structure is an input to our algorithm and the privacy granularity follows it. If cross-family or higher-order correlation matters for an application, that correlation should be factored into the privacy unit, so that a larger group becomes the definition of one client; the definition and the guarantee carry over unchanged. The cost is that model utility may drop as we increase the size of clients, which we note in Section 5 (Limitations).

Thank you again for your thoughtful review. As the discussion closes today, we wanted to check whether our responses have addressed your questions. We clarified our novelty and the scope of the client-level DP guarantee, detailed the privacy accounting with comparisons at matched δ, and added absolute reconstruction accuracy. If any concerns remain, we would be glad to address them. Otherwise, we would be grateful if you would consider updating your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.