Rerandomized Connectome Releases: Unlinkable Brain Data That Keeps Its Science

A person's resting-state brain connectivity pattern works like a fingerprint. In the Human Connectome Project (HCP), a scan from one day picks out the same person on another day with near-perfect accuracy. Because consortium data are copied into many studies, anyone holding two copies, or one copy plus a scan from elsewhere, can link a participant across them. Encrypting the files does not help, since researchers must decrypt them to use them, and the decrypted matrix is itself the identifier. We borrow one idea from cryptography. Ciphertext rerandomization turns an encrypted message into a fresh-looking encryption of the same message, so an observer cannot tell that two ciphertexts belong together. Our proposal is the analogue in plaintext for brain data. Each time a dataset is released, every subject's connectivity matrix is rewritten so that the scientific content is unchanged but the identifying part is freshly randomized. Two releases then cannot be linked.

Concretely, we split each connectivity matrix into two parts. One part is explained by group structure, covariates, and a small set of directions tied to behavior, and it is kept. The other part is the individual residual. For each release, we replace every subject's residual with the residual of a different subject from the same demographic stratum, using a new random assignment each time. This swap preserves the distribution of residuals across the release exactly, which parallels "perfect" rather than "computational" rerandomization. It also cuts the link between the residual and the person. The research question is whether the kept part can be made small enough to defeat the best fingerprinting attackers while group maps, task decoding, and brain–behavior prediction still hold up. We answer it against a security target set by what an attacker could already infer from age, sex, and family structure alone.

Core Claim

Existing fingerprinting work on HCP treats identifiability as evidence of reliability. GEFF projects functional connectomes, the region-by-region correlation matrices of brain activity, into a group principal-component space and identifies subjects by nearest centroid. Degree-normalization rescales edges to down-weight hub regions and raises identification rate by 16% on average. The autoencoder plus sparse dictionary approach of Cai et al. removes shared population activity and reaches up to 99.5% rest-to-rest identification. Taken together, these papers are a catalog of attacks: none of them asks how to release data that resists them. They also supply the mechanism we need. GEFF reports that early eigenvectors carry task-dominant variance and later ones carry subject-dominant variance. Cai et al. find that the discriminative signal and the cognition-predictive signal both live mostly in association cortex. The second finding means the trade-off is real, and it cannot be assumed away.

Our claim is that identity information and scientific content take up different amounts of the connectome. Brain–behavior prediction depends on a few directions, while fingerprinting draws on thousands of weakly informative edges. So there is a release mechanism, residual swapping within covariate strata with fresh randomness per release, that pushes cross-release linking accuracy down to the level of a covariate-only attacker while keeping most of the predictive correlation. Write $F_i$ for subject $i$'s vectorized connectome, $\mu(c_i)$ for the covariate-conditional group mean, and $P$ for the projector onto a kept subspace made of group, task, and behavior-relevant directions. The release is then

$$\tilde F_i^{(r)} = \mu(c_i) + P\,(F_i - \mu(c_i)) + (I-P)\,(F_{\pi_r(i)} - \mu(c_{\pi_r(i)})),$$

where $\pi_r$ is a random permutation for release $r$ that only swaps subjects within the same covariate stratum and never maps a subject to itself. The consortium keeps the permutations secret. This loosely echoes updatable public keys in QuisQuis, where the holder of the secret can still connect refreshed identities that outsiders cannot.

The statistical frame comes from On statistical tests of functional connectome fingerprinting. That paper shows that under random exchangeability the number of correct matches is Poisson(1), so more than 4 correct matches rejects the null at $p<0.01$. It also shows that this null is easily rejected by demographic or familial clustering alone, even when 90% of the data is noise. This constrains our design: "matches no better than chance" is the wrong security target, and one we could never honestly meet. Our target is instead the linking accuracy of an attacker who sees only the released covariates and family labels.

Research Program

The first component is the kept subspace $P$. We build it from three sources. The first is the leading group principal components, following the GEFF finding about task-dominant early eigenvectors. The second is the task-discriminating directions. The third is a low-rank set of directions fitted, on a held-out split, to predict the behavioral variables the consortium releases, such as fluid intelligence and language comprehension. We sweep the rank of $P$ to trace a curve of privacy against utility, rather than reporting a single operating point.

The second component is the strata. We match on age band, sex, and scanner session. We also study family-aware strata on the HCP twin and sibling cohort, because the similarity hierarchy Self > monozygotic twin > dizygotic twin > sibling > stranger means that residual swapping between unrelated subjects could still leave twins linkable through their kept part.

The third component is the attacker suite, which follows Kerckhoffs' principle: the attacker knows the mechanism and $P$, and lacks only the permutations. It includes nearest-neighbor matching and Hungarian-assignment matching on raw and degree-normalized connectomes, the GEFF attacker, and the Cai et al. attacker. We add an adaptive attacker that projects both copies onto $P$ before matching, since that is the best use of the part that survives. We also add a learned metric attacker trained on some linked pairs from the attacker's own scans.

The fourth component is utility. We measure preservation of group-level edge statistics, task identification accuracy, and out-of-sample prediction of fluid intelligence and language comprehension from the released data. We state the scope up front. Heritability and individual-trajectory studies need the residual, so they are served by the unmodified controlled-access data, not by this release. The method targets the broadly shared open tier.

Baselines

The naive reference point is releasing raw connectomes, which is current practice. Here linking is essentially solved. Cai et al. report up to 99.5% rest-to-rest identification on HCP S1200 with the Shen 268-node atlas. GEFF reports 100% subject identification for resting state with one learning task and 98% with two. The statistical-tests paper reports about 75% on its HCP cohort. We rerun all three attackers on raw data in our own splits so that every defense is compared against the same starting accuracy. This also gives the utility ceiling: fluid-intelligence prediction on raw versus refined connectomes is reported to rise from 0.481 to 0.548 correlation in Cai et al. That number serves as our reference for what good utility looks like.

The strongest competing defense is truncation, built on GEFF's own eigenvector finding. It projects each connectome onto the leading group components and discards the subject-dominant tail. The resulting release looks like ours with the residual set to zero, which makes truncation also the ablation that isolates our key ingredient. Because it is both the closest competitor and the key ablation, it needs a fair comparison. It gets the same $P$ (same rank sweep, same tuning split, same compute) as our method, so the only difference is zeroing versus swapping the residual. Two outcomes would count as beating it. The first is equal or lower adaptive-attacker linking accuracy at the same rank. The second is better preservation of edge variances and of downstream regression calibration, since zeroing shrinks every subject toward the mean and biases any statistic that depends on spread.

The second defense baseline is additive noise in the style of differential privacy, following the Laplacian edge perturbation used in DP-FedRec. We sweep the noise scale over the same range of achieved linking accuracies, which lets us compare the utility of the two methods at matched privacy. Beating it means higher behavior-prediction correlation at every matched linking accuracy.

A further ablation draws the swapped residual from the whole cohort instead of from covariate strata. This tests whether stratification is what keeps demographic structure intact and keeps the released distribution realistic.

Why not just…?

Why not just encrypt the data with an oblivious or enclave-based store? Systems such as Enclage, Anonymous RAM, and the Melbourne Shuffle hide stored content and access patterns from an untrusted server. They are evaluated on key-value and abstract memory workloads. None of them protects the plaintext once an authorized researcher reads it, and in open science that happens constantly. The linking threat comes from the content, not from storage. Encryption is complementary for the controlled-access tier and is not a substitute here.

Why not just add calibrated noise? This is a real alternative, and the matched-privacy comparison above settles it empirically. There is a reason to expect it to lose. The weak-null result shows that a small structured fraction of the data is enough to drive significant matching. Isotropic noise therefore has to swamp the identifying signal in every direction, including the behavior-relevant ones. Residual swapping, by contrast, spends its randomness only in the directions the analysis does not need. If noise matches swapping on the utility curve, we will report that the simpler method suffices.

Why not just truncate to the group eigenspace, as GEFF's structure suggests? This is the most serious threat. If an adaptive attacker links truncated and swapped releases equally well, and if truncation's shrinkage turns out harmless for prediction, the extra machinery is not justified. The zeroing-versus-swapping ablation at matched rank decides it. We expect swapping to win mainly on distributional fidelity, meaning variances, calibration, and realistic-looking matrices. It may also win on privacy, because shrunken subjects can remain separable within the kept dimensions. We will say plainly if that does not happen.

Proposed Experiments

The main experiment uses the HCP 1200-participant release with 409 unrelated subjects and the Glasser 374-region atlas, as in the degree-normalization study. We use one rest session for release A and the other day's session for release B, each processed with an independent permutation. We report linking accuracy for every attacker against the covariate-only attacker, together with correct-match counts relative to the Poisson(1) null, as a function of the rank of $P$.

We repeat the experiment on the HCP S1200 cohort with the Shen atlas, where the Cai et al. attacker was reported, and on the 466-subject twin and sibling cohort to test family-aware strata. A cross-atlas test releases the data in Glasser while the attacker holds Shen-parcellated scans, which reflects the realistic case of a participant scanned by another study. The Baltimore Longitudinal Study of Aging cohort, where raw identification is reported at only about 40%, tests whether protection holds when scans are years apart.

For utility, we measure task identification on the eight HCP conditions used by GEFF, whose reported task identification saturates around 99%, and fluid-intelligence and language-comprehension prediction against the 0.481 and 0.548 references. We also add a sex-classification check with the meanMLP model, a simple permutation-invariant time-series classifier, on connectomes rebuilt from released data. That file reports no specific AUC on Lacuna, so we will establish the raw-data number ourselves.

The decisive check is whether some rank of $P$ keeps behavior prediction within a few hundredths of the raw correlation while the adaptive attacker's cross-release linking accuracy stays statistically indistinguishable from the covariate-only attacker. If no such rank exists, the conclusion is itself useful and publishable: identity and behavior are entangled in the same few directions, and open connectome release cannot be both useful and unlinkable.