Noise That Respects the Genome: Linkage-Disequilibrium Metric Privacy for Encrypted Federated GWAS

Hospitals can now run a genome-wide association study (GWAS) across their combined biobanks without any of them seeing another's raw genotypes: systems like SF-GWAS keep the data encrypted throughout using multiparty homomorphic encryption (HE) — encryption that lets computers add and multiply on ciphertexts without decrypting them. But when the analysis finishes, these systems decrypt and publish the exact association statistics. The literature is blunt about what that costs: exact GWAS summaries and beacon responses are vulnerable to membership inference, in which an attacker determines whether a specific person's genome is in the study (Differential Privacy Protection Against Membership Inference Attack on Machine Learning for Genomic Data; Enabling tradeoffs in privacy and utility in genomic data Beacons and summary statistics), and genomic data is a permanent identifier that also implicates relatives (Practical Solutions for Protecting Individual Genomic Privacy).

The obvious fix — add differential-privacy noise to the released statistics — is worse than it sounds for two reasons. First, standard differential privacy treats the genome as independent columns: it adds noise separately to each of roughly 650,000 SNP statistics, destroying the subtle correlation structure GWAS depends on, while simultaneously leaking information, because SNPs are not independent and an attacker can combine correlated statistics to cancel noise (Blowfish Privacy: Tuning Privacy-Utility Trade-offs using Policies). Second, generating noise inside a cryptographic pipeline is expensive: the state of the art for MPC-friendly Laplace noise costs 21.6 to 26.6 milliseconds per sample with 14,397 to 19,781 AND gates at 128-bit security (A Failure-Free and Efficient Discrete Laplace Distribution for Differential Privacy in MPC) — multiplicative depth and interaction that biobank-scale encrypted pipelines cannot afford. This proposal builds a private-release layer for encrypted federated GWAS whose privacy guarantee is defined by, and whose noise is shaped by, the actual correlation geometry of the human genome.

Core Claim

The claim is that metric differential privacy — a version of differential privacy where the strength of the guarantee scales continuously with a distance $d$ between data points rather than treating all differences as equal (A Decade of Metric Differential Privacy: Advancements and Applications) — can be instantiated with a genomic metric built from linkage disequilibrium (LD, the correlation between nearby SNPs) and kinship, and that this mechanism both closes the correlation-leakage hole of per-SNP noise and reduces total noise, because the noise budget should scale with the number of effectively independent haplotype blocks rather than the number of SNPs. Moreover, the mechanism can be evaluated inside the encrypted pipeline at near-zero added cost, because it is a linear perturbation: one plaintext-ciphertext matrix multiplication in the HE layer, with no bootstrapping and no extra MPC rounds.

No existing work connects these pieces. SF-GWAS, SF-Relate, SAFETY, and SVAT are encrypted genomic systems with zero differential privacy on outputs; PrivSTRAT is differentially private GWAS in plaintext with standard, non-metric noise that protects only phenotypes; Revisiting ML Training under Fully Homomorphic Encryption combines HE with standard DP but only for low-dimensional convex tabular tasks. The decade survey of metric DP explicitly lists principled geometry-aware mechanisms for high-dimensional biological data as an open problem, and notes that adversarial side information — exactly what LD and family structure provide — distorts naive metrics. This proposal is aimed at that gap.

The Mechanism

Concretely, we define a distance between two individual genomes $g, g' \in \{0,1,2\}^M$ (each SNP takes value 0, 1, or 2 copies of the minor allele) as an LD-weighted block distance:

$$d(g, g') \;=\; \sum_{b=1}^{B} w_b \cdot \delta_b(g, g'),$$

where the genome is partitioned into $B$ haplotype blocks using a public reference panel (1000 Genomes, the same resource used by SAFETY and SVAT), $\delta_b$ measures how distinguishable the two genomes are in block $b$ (zero when they carry haplotypes that are nearly interchangeable given the block's LD structure, following the elastic-distinguishability idea of setting $d=0$ inside protected zones from Constructing elastic distinguishability metrics for location privacy), and $w_b$ weights blocks by their effective information content. A mechanism $K$ then satisfies metric DP when

$$\frac{P(K(g) \in O)}{P(K(g') \in O)} \;\le\; e^{\epsilon \cdot d(g, g')}.$$

Relatives, who share long chromosome segments identical by descent, become close under this metric — an attacker cannot tell which family member contributed, which directly addresses the relative problem identified by Practical Solutions for Protecting Individual Genomic Privacy. Kinship distances can be computed securely with the KING-robust estimator machinery of SF-Relate, which already demonstrates that relatedness filtering is both feasible and necessary at biobank scale (it cut cross-site comparisons by over 99.8% and reduced false positives from 5.14% to 5.01% in a simulated six-party cohort).

The release layer applied to the encrypted GWAS statistic vector $s \in \mathbb{R}^M$ is then simply

$$\tilde{s} \;=\; s + A\,\eta,$$

where $\eta \in \mathbb{R}^B$ is block-level noise and $A$ is a public $M \times B$ map from blocks back to SNPs derived from the reference LD matrix. The crucial engineering observation is that every operation here is linear. Each institution locally samples its share of $\eta$, encrypts it under the multiparty HE scheme, and the shares sum homomorphically; the aggregate distribution is calibrated to the metric target. The $A$ multiplication is a plaintext-ciphertext operation, depth one, SIMD-friendly in CKKS in the same way CryptoNets amortizes cost across packed slots. No comparisons, divisions, or bootstrapping are introduced — we deliberately inherit the no-clipping philosophy of Revisiting ML Training under Fully Homomorphic Encryption (which cut circuit depth from 24 to 9 by making bounds analytic rather than computed), but apply it to output perturbation, where genotype bounds $\{0,1,2\}$ give analytic sensitivity with no encrypted comparisons at all. Because noise is added under encryption, no party — not even the one holding the decryption key — ever sees the exact pooled statistics; the decryption oracle sees only already-perturbed values, which incidentally also mitigates the CKKS decryption-leakage concern raised by Securing Approximate Homomorphic Encryption Using Differential Privacy.

We adopt the semi-honest adversary model deliberately, matching SF-GWAS and SF-Relate: participants follow the protocol but inspect everything. The DP layer is orthogonal defense-in-depth against what semi-honest encryption cannot cover — what the output reveals. Malicious-security upgrades such as Helen's zero-knowledge verification are explicitly out of scope; our design adds no new verification surface.

Research Program

The work has three parts. First, construct the metric: partition the genome into haplotype blocks from 1000 Genomes LD structure, measure the effective covering number of the genotype space under $d$ (this determines the noise scaling — the hypothesis that noise scales with $B \ll M$ independent blocks rather than $M$ SNPs is a measurable empirical claim, not an assumption), and define the kinship component of the metric using IBD sharing so that close relatives are assigned small distances. Second, derive the mechanism: prove the metric-DP guarantee for the block-noise construction, work out the aggregate distribution of summed party-generated noise shares and its calibration, and establish utility bounds (noise variance per SNP statistic versus the per-SNP independent baseline at matched worst-case guarantee). Third, integrate into an SF-GWAS-style pipeline: implement the release layer in CKKS as an additive share-plus-matrix-multiply step, profile the added wall-clock time, multiplicative depth, and communication, and verify that the non-linear MPC phases (divisions, square roots) are untouched.

Proposed Experiments

Evaluation uses the benchmarks documented in this literature. The realistic-scale testbed is the AMD GWAS cohort ($n = 22{,}683$), on which SF-GWAS reports 4.6 hours versus 64.3 hours for its baseline (14x faster) — our target is to keep the added release layer under a small fraction of that. 1000 Genomes populations (CHB, CHS, JPT, MXL, as used by SAFETY and SVAT) serve for the metric construction and for simulated multi-site cohorts of the kind SF-Relate evaluated (100,000 samples across six parties). UK Biobank scale (409,547 individuals for LMM, where SF-GWAS finished in about 6 days) is the stretch goal for the overhead measurement, since the release layer's cost is what we are testing there, not the full LMM.

Privacy is measured by attack, not by $\epsilon$ alone: membership-inference success on released statistics using the beacon-style and Bayesian attacks of Differential Privacy Protection Against Membership Inference Attack on Machine Learning for Genomic Data and Enabling tradeoffs in privacy and utility in genomic data Beacons and summary statistics, run by an attacker who knows the true LD structure and has access to relatives' genomes — the strong-prior adversary that the metric-DP survey flags as the model's weak point. Utility is measured three ways: effect-size fidelity, reported as squared Pearson correlation against plaintext REGENIE/PLINK, where SF-GWAS reports $r^2$ typically exceeding 0.99 (we aim to stay above roughly 0.95 at the strongest attack-matched noise setting, accepting an explicit power loss); proportion of true genome-wide-significant associations recovered; and downstream polygenic risk score performance under PUMAS-style summary-statistic fine-tuning, which reports an average 205% improvement in predictive $R^2$ over naive summary methods and runs in 8.3 seconds per trait on a standard CPU — making it a cheap, standardized utility probe for perturbed summaries. Overhead is reported as added runtime, depth, and communication relative to the unmodified encrypted pipeline, and against the MPC-sampling alternative, whose per-sample cost of 21.6 to 26.6 milliseconds and 14,397 to 19,781 AND gates (A Failure-Free and Efficient Discrete Laplace Distribution for Differential Privacy in MPC) we extrapolate to 650K-SNP releases as a projected comparison point.

Baselines

The "do nothing" reference point is the existing encrypted federated pipeline with exact release, i.e., SF-GWAS as published: it achieves $r^2$ typically exceeding 0.99 against PLINK/REGENIE with zero added noise, and it is the utility ceiling. It will be run unmodified on the same cohorts and compute so that any utility loss we report is attributable to our layer alone. Beating it means holding attack success near chance while keeping effect-size correlation and association recovery within a stated tolerance — if our mechanism drives genome-wide power to uselessness at attack-resistant noise levels, the proposal fails, whatever the theory says.

The strongest existing competitor is standard per-SNP differential privacy applied to the released statistics — independent Laplace or Gaussian noise per SNP, in the spirit of PrivSTRAT's calibrated noise on EIGENSTRAT statistics, but applied at the release layer on top of the encrypted pipeline. This is the strongest competitor because it is the mature, widely implemented default, and any reviewer will demand it. It will be given the same data, the same encrypted pipeline, and an equal tuning budget over $\epsilon$ and noise scale, with the comparison made at matched attack success under the strong-prior LD-aware attacker rather than at matched $\epsilon$ (since the entire claim is that $\epsilon$ values are not comparable across metrics). We beat it if, at equal attack resistance, the LD-metric mechanism recovers significantly more true associations and higher PRS predictive $R^2$ — driven by the lower effective noise dimension — and if it additionally resists the correlation-combination attack that per-SNP noise provably does not.

Two ablations isolate the ingredients. The first replaces the LD-structured noise with i.i.d. per-SNP noise of the same total variance inside the same encrypted release layer, isolating the metric as the sole changed variable. The second runs the identical LD-metric mechanism but perturbs after decryption by a single output party, isolating the encrypted-release ingredient and directly measuring what the no-trusted-curator property costs in runtime (expected: very little, since the layer is linear). If the first ablation matches the full method, the contribution collapses to an engineering note about HE integration, and we will report it as such.

Why not just…?

Why not just run GWAS locally at each site and meta-analyze the summaries? This is the cheapest architecture and must be addressed first. SF-GWAS benchmarks meta-analysis directly against joint secure computation and reports that meta-analysis is not enough: it yields statistical discrepancies and lower power in heterogeneous, ancestrally diverse populations. PUMAS further shows that naive summary combination forfeits large predictive gains that modeling the underlying LD structure recovers. Our proposal targets exactly the regime where local analysis is insufficient, and adds to it an output-privacy guarantee that per-site exact summaries lack entirely. For collaborations where meta-analysis genuinely suffices, this proposal is not the right tool, and we say so.

Why not just use standard differential privacy on the outputs? This is the most serious alternative, and we do not argue around it — we test it head-to-head as the primary baseline. The evidence for expecting it to fall short is indirect but consistent: Blowfish Privacy shows that ignoring correlation structure forces excessive perturbation, with policy-aware mechanisms cutting clustering error by up to 100x on spatial data, and the metric-DP literature (A Decade of Metric Differential Privacy; Enhancing Scalability of Metric Differential Privacy via Secret Dataset Partitioning and Benders Decomposition) argues uniform noise obliterates utility for structured data. But none of that is genomic evidence, and it is entirely possible that at biobank cohort sizes the sample count is large enough that per-SNP noise is affordable and the LD advantage is marginal. The attack-matched comparison on the AMD cohort and simulated multi-site cohorts is designed precisely to settle this, and we commit to reporting whichever way it goes.

Why not just decrypt to a trusted output curator and perturb in plaintext? If a single trusted party is acceptable, this is simpler and cheaper, and the plaintext-perturbation ablation will quantify exactly how little the encrypted version costs. The reason to do it under encryption is the trust model itself: in cross-institution biobank collaborations there is no party all sites accept as curator, and the exact pooled statistics are precisely the artifact no single site should see. Our design makes the no-trusted-curator property nearly free rather than expensive, which is the honest framing — the ablation will show whether "nearly" holds.

Why not just sample the noise in the MPC phase the pipeline already has? SF-GWAS already drops into MPC for divisions and square roots, so one could run Tjuawinata et al.'s discrete Laplace samplers there. Two problems: at 21.6 to 26.6 milliseconds per sample, per-SNP sampling across hundreds of thousands of SNPs adds hours of interactive MPC across WAN links, the exact overhead the HE literature identifies as the bottleneck; and discrete per-grid-cell noise is independent, so it inherits the correlation-leakage hole our metric is designed to close. Our layer instead costs one depth-one HE multiplication and no interaction. This is a projected comparison in our experiments rather than a fully implemented one, and we will label it as such.

Why not just use secure enclaves (Intel SGX) instead of HE? SAFETY already answered this empirically: its own hybrid design was 1.7x faster than pure SGX for linkage disequilibrium and 1.93x to 4.82x faster for Hardy-Weinberg tests, because every owner's inputs otherwise require independent in-enclave decryption, and pure SGX remains exposed to side-channel attacks. Our release layer is enclave-agnostic in principle, but the encrypted setting is the one the leading biobank-scale systems actually use, so that is where we build.

Limitations and Decisive Checks

The design's central vulnerability is the metric-trust assumption the decade survey warns about: the guarantee holds only if $d$ faithfully captures what an attacker can distinguish. Our answer is empirical — every privacy evaluation is run by an adversary equipped with the true LD structure and relatives' genomes, the worst case the model admits — but if that attacker breaks the mechanism at noise levels that preserve utility, the genomic metric formalism fails here and the honest conclusion is that metric DP does not transfer to this domain. The second decisive check is the noise-scaling hypothesis: we will measure the effective block covering number from 1000 Genomes panels, and if noise does not scale down relative to per-SNP independence at matched worst-case guarantees, the utility case for the whole construction evaporates. The third is overhead: if the encrypted release layer costs more than a small fraction of the AMD-cohort pipeline's 4.6-hour runtime, the "nearly free linearity" claim is wrong in practice and the trusted-curator simplification becomes the better engineering recommendation. Each check has a pre-stated failure condition, and the proposal is falsifiable on any of them.