Brain wiring is partly inherited. Identical twins have measurably more similar connectomes (maps of which brain regions are connected, and how strongly) than fraternal twins or unrelated people. This is why twin cohorts are among the most valuable datasets in neuroimaging, and also why they are sensitive: sharing them can expose who is related to whom. A popular proposal for multi-institution data sharing is "hybrid" blockchain storage. The scans stay encrypted in cheap off-chain storage. A public ledger records only small fingerprints of the data (cryptographic hashes) and the log of who was granted access.
We argue that this design has a leak that encryption cannot fix. Heritability studies must upload, anchor, and request relatives together, because the analysis compares twins with twins. The public ledger therefore records family structure in its timing, its ordering, and its access grants, even when every scan is perfectly encrypted. The proposal measures this leak on real Human Connectome Project (HCP) twin data, designs commitment and access schemes that hide family structure, and reports how much heritability-estimation accuracy and on-chain cost those schemes give up.
Core Claim
The claim has two parts. First, in hybrid designs of the kind proposed by Zyskind et al. and benchmarked by Delgado-Mohatar et al., an adversary who sees only the public ledger can reconstruct a large fraction of a twin cohort's family structure. That ledger holds anchoring transactions, the positions of records within a Merkle tree (a hash tree whose single root commits to many records), and access-grant transactions. The adversary needs no scan content to do this.
Second, the leak comes from the workload, not from the cryptography. A classical twin estimate of heritability makes the reason clear:
$$h^2 \approx 2\,(r_{MZ} - r_{DZ}),$$
Here $r_{MZ}$ and $r_{DZ}$ are the within-pair correlations of a connectivity measure in identical and fraternal twin pairs. Any analysis of this kind must retrieve records in labeled family units, and the retrieval pattern reveals those units. Variance-component tools such as SOLAR-Eclipse, used by Kochunov et al., need the whole pedigree, so the problem gets worse.
What is new is the target. Existing work either secures template content or secures model parameters. Delgado-Mohatar et al. anchor hashes and reduced face templates. Goel et al. chain CNN layers and Merkle-organize templates. BPDS encrypts medical-record payloads under attribute policies. MetisFL encrypts federated model updates. None of them treats the relational structure among data subjects as the asset at risk, and none studies a workload whose access pattern is itself relational.
Why Connectomes Make This Acute
The neuroimaging literature supplies both the motivation and the attack tools. Connectome fingerprints are small and identify people almost perfectly. Fiberprint compresses about 50,000 tractography streamlines per subject into a 50-dimensional vector that reaches 100% precision@1 over 4,255 scan instances. Ravindra et al. identify subjects from only 61 to 95 of more than 64,000 functional connections, with 93.11% test accuracy. Chauvin et al. match identical twins with near 100% accuracy from 3D keypoint sets using a soft Jaccard index.
These compact signatures are exactly what a cost-conscious hybrid design would place on-chain or in transaction payloads. Delgado-Mohatar et al. compress face features by 97.5% specifically to cut gas cost. So the "reduced template" storage strategy is a direct kinship leak for connectomes. It is not only an identity leak, because heritable signal concentrates in these features: Pourmotabbed et al. identify identical twins from network topology at up to 88% accuracy with an area under the ROC curve (AUC) of 0.94–0.96. The content leak is known in outline. The metadata leak is the part nobody has measured.
Research Program
Building the design space. We build a testbed of hybrid storage designs on a local Ethereum-compatible chain, instantiated on HCP cohorts. Each design is a choice along three axes.
The first axis is what is anchored. Options are a per-subject SHA-256 hash of the raw scan, a per-subject hash of a quantized compact signature, a reduced template in plain form, or a single Merkle root per data release.
The second axis is how anchoring is scheduled. Records can be anchored as acquired, which in twin studies tends to group families together in time. Alternatively they can be batched per release, with leaf order randomized and each leaf salted (a random value mixed into each hash so it cannot be guessed).
The third axis is how access is granted. Options are per-subject grants in the style of Zyskind's $T_{access}$ transactions, cohort-block grants, or per-subject grants padded with decoy requests drawn to match family-size statistics.
Simulating the workload. Access logs are generated from realistic analysis workloads: twin-pair designs of the kind above, full-pedigree variance-component analyses, and non-genetic analyses such as sex classification as a control.
HCP open data does not give us upload timestamps. The anchoring-order experiments therefore simulate acquisition schedules at several levels of family clustering, ranging from fully family-sequential to fully random. We report results as a function of that clustering rather than claiming a single real-world number.
The adversary. The adversary sees the public ledger. Optionally, it also holds one scan of a known person, modeling a relative who participated in another study. It outputs a predicted relatedness graph over pseudonymous ledger identities. We score it by pairwise relatedness AUC, by F1 on recovered family clusters, and separately on identical-twin pairs.
Defenses and their costs. For each defense we measure three things. The first is utility: bias and variance of heritability estimates relative to the undefended analysis. The second is on-chain cost: gas, and dollars at the per-operation cost reported by Delgado-Mohatar et al. The third is latency.
Baselines
Naive reference. The do-nothing reference is per-subject hash anchoring with per-subject access grants, anchored in acquisition order. This follows the Data Hashing strategy of Delgado-Mohatar et al. combined with the access log of Zyskind et al.. It is what a straightforward deployment would build, and it sets the leakage ceiling. We expect it to be the worst case. Any defense that fails to lower pedigree-recovery F1 substantially below this reference, at matched utility, has failed.
Strongest existing privacy design. This is attribute-based encryption of off-chain payloads with policy-governed access, following BPDS. It is the most complete confidentiality mechanism in the file. The file lists no reported result on Lacuna for BPDS latency or leakage; its authors give no quantitative benchmarks. We run it on identical data, workload, and adversary.
The hypothesis is sharp. Its pedigree-recovery AUC should be statistically indistinguishable from the naive reference, because encryption leaves timing and grant structure untouched. If BPDS-style encryption alone drives leakage to chance, our central claim is wrong.
Strongest content-level attack. For the reduced-template anchoring strategy, the attack baselines are the published fingerprint matchers. The reported figure to reproduce is identical-twin identification accuracy of up to 88% at AUC 0.94–0.96 from Pourmotabbed et al.. We also use the near-100% twin matching of Chauvin et al..
We apply these matchers to on-chain Fiberprint and leverage-score signatures. Every attack method receives the same HCP subjects, the same train/test family splits (entire families on one side), and the same hyperparameter search budget. The file's figures come from different pipelines. What we aim to establish is whether ledger-exposed compact signatures reach comparable kinship AUC; we do not claim to match those numbers exactly.
Trusted-broker reference. This is a centralized trusted repository with no public log, in the spirit of the central controller of MetisFL or the cloud backend of CAVE. By construction it leaks nothing to outsiders, and it marks the privacy a public ledger must approach to be worth using.
Ablations. The ablations isolate each ingredient. We run batching alone, leaf shuffling and salting alone, and decoy padding alone. We also vary decoy calibration: decoys matched to family-size statistics versus uniform random decoys. This separates whether hiding the anchoring order or hiding the access pattern does the work, and whether decoys must mimic pedigree structure to be effective.
Why not just…?
Why not just use a trusted central repository and data-use agreements? This is a real threat to the whole enterprise, and we treat it as one. If multi-institution parties are willing to trust one operator, the trusted-broker reference gives the best privacy at the lowest cost, and a ledger adds only risk.
The experiments settle the question quantitatively. They report how close the best ledger design comes to zero leakage, and at what gas and utility price. A negative finding is also useful: public audit trails would then be a poor fit for kinship-bearing cohorts, and we would publish it as such.
Why not just encrypt everything with attribute-based encryption? Encryption hides what is in a record, not which records are touched together or when. The BPDS baseline is the direct test. The prediction is that its pedigree-recovery AUC matches the unencrypted naive design.
Why not just add differential privacy to released templates or heritability estimates? Differential privacy adds calibrated noise to published outputs so that no single person's data changes them much. It protects released numbers, but grant transactions are not outputs, and the grant pattern is where the relational leak sits. Noise on the estimates also costs the most in exactly the small cohorts where heritability matters: the MEG cohort in Pourmotabbed et al. has 89 subjects.
We include a differentially private release arm. It lets us show that output noise leaves access-log leakage unchanged, and it quantifies the loss in heritability accuracy at practical privacy budgets.
Proposed Experiments
The data are the HCP Q3 diffusion release, the HCP Q4 structural release, and the HCP S1200 release. Q3 has 861 subjects with 82 identical-twin pairs, 82 fraternal-twin pairs, and 166 non-twin sibling pairs. Q4 has 1,010 subjects in 434 families, including half-siblings. S1200 provides functional MRI on 1,003 subjects and MEG on 89. Signatures are computed with the published Fiberprint, leverage-score, and keypoint pipelines.
The first decisive check tests whether metadata alone recovers pedigrees. Under the naive design and a full-pedigree workload, family-cluster F1 should rise well above chance, and it should grow with acquisition clustering and with the number of heritability studies run. We will report the whole curve. If recovery stays near chance even for family-sequential schedules, the thesis fails.
The second check tests whether defenses restore privacy without breaking the science. With batched, shuffled, salted Merkle-root anchoring plus calibrated decoys, recovery should fall toward the trusted-broker reference. Tract-level fractional-anisotropy heritability recomputed through the defended pipeline must also reproduce the undefended estimates. The reported values to reproduce, from Kochunov et al., are 0.88 whole-brain average, 0.90 corpus callosum, 0.66 corticospinal tract, and 0.53 fornix. We will verify agreement within the estimates' confidence intervals.
The third check is cost. Delgado-Mohatar et al. report about $0.0122 per Merkle root operation, and roughly $12,200 versus more than $610,000 per million templates for hybrid versus full on-chain storage. Per-release Merkle roots should be cheaper than per-subject anchoring. Decoy padding adds cost, and we will report it explicitly in gas units, so the privacy–cost trade-off is visible rather than asserted.
The fourth check covers the known-relative attack on content. If the adversary holds one relative's scan, per-subject hashes of quantized compact signatures may be vulnerable to dictionary search when quantization is coarse. We measure the quantization level at which salting becomes necessary, and we show that plain reduced templates on-chain allow twin identification at AUC comparable to published matchers.
References
- Decentralizing Privacy: Using Blockchain to Protect Personal Data
- Biometric Template Storage With Blockchain: A First Look Into Cost and Performance Tradeoffs
- Heritability of fractional anisotropy in human white matter: A comparison of Human Connectome Project and ENIGMA-DTI data
- Securing CNN Model and Biometric Template using Blockchain
- BPDS: A Blockchain Based Privacy-Preserving Data Sharing for Electronic Medical Records
- Secure Federated Learning for Neuroimaging
- Fiberprint: A subject fingerprint based on sparse code pooling for white matter fiber analysis
- Constructing Compact Signatures for Individual Fingerprinting of Brain Connectomes
- Efficient Pairwise Neuroimage Analysis Using the Soft Jaccard Index and 3D Keypoint Sets
- Genetic fingerprinting with heritable phenotypes of the resting-state brain network topology
- CAVE: Connectome Annotation Versioning Engine