Open neuroimaging consortia such as the Human Connectome Project (HCP) protect participants by removing faces and demographic details before release. Their own users then show that a person's functional connectome, the matrix of correlations between brain regions' activity over time, works as a fingerprint. It picks out the same individual across scan days with near-perfect accuracy. The distributor removes identifiers and assumes downstream researchers will not rebuild them. Downstream researchers apply hub normalization, rank correlations or eigenspace projections to improve their models, and assume the release was already de-identified. Security researchers call this pattern a responsibility gap: each party believes the adjacent party handles a protection, so no one does. It has been documented in Node.js file-upload middleware and in enterprise machine learning teams. No one has measured it in a scientific data pipeline.
This proposal makes the gap measurable. For any released artifact (raw time series, a parcellated connectome, a projected embedding), we define its release robustness: the highest cross-session identification rate that any downstream pipeline in a realistic family can recover from it. We then ask whether identification risk can be split into per-stage contracts. In such a contract, each party promises a bound on how much its stage can raise identifiability, and the end-to-end bound follows from the local promises. Our hypothesis is that it often cannot. Standard graph-construction choices interact, so identification risk is a jointly owned property of the whole pipeline. Release policy therefore has to be written against the worst case over downstream processing, not against the processing the distributor happened to apply.
Core Claim
Two papers in the HCP literature treat the uniqueness of functional connectomes purely as a biomarker to maximize. GEFF: Graph embedding for functional fingerprinting projects connectomes into a group-level eigenspace and identifies subjects with a nearest-centroid rule. Improving Functional Connectome Fingerprinting with Degree-Normalization suppresses dominant hub regions to sharpen individual signatures. We read both papers as attack methods. They motivate this work because they show identifiability is a moving target: it depends on transformations the distributor never applies.
On the responsibility side, Information-flow interfaces supplies the vocabulary. It formalizes three kinds of requirement. Assumptions are what a component expects of its environment. Guarantees are what the component itself delivers. Closed-guarantees are properties of the whole system that the component and its environment share responsibility for. That theory is binary: a flow is either forbidden or not. It explicitly leaves quantitative leakage open, and it has no notion of statistical identifiability.
We make the contracts quantitative and test them empirically rather than prove them. Let $A$ be a released artifact and $\mathcal{F}$ a family of downstream pipelines. Let $\mathrm{ID}(\cdot)$ be the fraction of subjects whose day-2 connectome is matched correctly to their own day-1 connectome. Release robustness is then
$$R(A) = \max_{f \in \mathcal{F}} \mathrm{ID}(f(A)).$$
A distributor guarantee is a statement $R(A) \le \rho$, valid for the family $\mathcal{F}$. A downstream party that applies a transform outside $\mathcal{F}$ inherits responsibility for any identifiability above $\rho$. A stage $s$ satisfies a local contract with gain $g_s$ if, for every input $X$ in a test pool,
$$R(s(X)) \le g_s \, R(X).$$
The compositional question is whether $\prod_s g_s \, R(\text{input})$ bounds the measured end-to-end identifiability. Wherever it fails, the bound is a closed-guarantee that no single party can certify alone.
Research Program
We build $\mathcal{F}$ from published HCP pipelines rather than inventing attacks. The data-centric design space of Defining and Benchmarking a Data-Centric Design Space for Brain Graph Construction varies three stages. The first is temporal filtering, for example keeping only high-amplitude signal. The second is how edges are defined, using Pearson, Spearman or Kendall correlation, with or without one topology shared across all subjects. The third is lagged-correlation features; that paper finds HCP prefers lags of 3–8. Its codebase is GeQinwen/DataCentricBrainGraphs. We add degree-normalization and GEFF eigenspace projection as further stages. Ge et al. treat all of these purely as utility hyperparameters; we score each one for both utility and identifiability.
The analysis has three parts.
Attribution. For every pipeline in the grid we measure identification rate and differential identifiability. We then estimate each stage's contribution with a Shapley-style decomposition over stage choices, which tells us whether risk is additive or dominated by interactions.
Composition test. We fit local gains $g_s$ on held-out subjects and check whether their product predicts end-to-end identification on new subjects.
Release interventions. We evaluate candidate release forms under the worst case over $\mathcal{F}$. GEFF reports that early eigencomponents carry task variance and later ones carry individual idiosyncrasy. This suggests releasing only the task-dominant subspace, which is a concrete, testable form of declassification.
Utility is measured throughout so that a release form cannot win simply by destroying the data. Following A simple but tough-to-beat baseline for fMRI time-series classification, we use a meanMLP sex and age classifier, plus GEFF-style task identification.
One consequence of the Popov et al. result deserves a test in its own right. If meanMLP works equally well on time-shuffled data, then temporal shuffling preserves the connectome, and shuffling offers no privacy protection at all. The dynamics-focused work of Behavioural relevance of spontaneous, transient brain network interactions in fMRI raises the opposite worry. Releases that keep transient, time-varying network interactions could leak identity through a channel that static-connectome audits never see. We therefore include a dynamic-connectivity attacker in $\mathcal{F}$.
Baselines
Do-nothing reference. We match day-1 and day-2 sessions using raw Pearson correlation connectomes on the HCP 1200 release: 409 unrelated subjects, parcellated with the Glasser multimodal parcellation into 374 regions, and compared within the same task. This shows how much risk the current release form carries before anyone tries to amplify it.
Strongest existing attacks. The first is GEFF. On 100 unrelated HCP 900 subjects it reports 100% subject identification for resting state with one learning task and about 98% with two learning tasks (GEFF). The second is degree-normalization, which reports an average 16% gain in identification rate and 9.6% in differential identifiability over baseline connectomes (Degree-Normalization).
To keep the comparison fair, every attack runs on the same subject splits, parcellations and sessions. Every attack gets the same hyperparameter search budget, selected on a disjoint subject set. We first reproduce the published numbers within a few points before trusting our harness.
A release intervention counts as beating these baselines under the following condition. It must cut the worst-case identification rate over all of $\mathcal{F}$, not over any single attack, to near chance (1/409) or to a pre-registered threshold. It must do so while keeping meanMLP AUC and task identification within a pre-registered margin of the unprotected release. We found no reported Lacuna result for meanMLP's specific HCP AUC, so we will measure it ourselves.
Key ablations. The first evaluates each release form only against the pipeline the distributor itself applied, not the worst case. This isolates the value of worst-case auditing. The second replaces the learned eigen-subspace truncation with random-subspace truncation of the same dimension. This isolates whether the task/subject split in the eigenspace is doing the work.
Why not just…?
Why not just audit the final artifact once and gate release on its identification rate? This is the most serious alternative. If end-to-end identifiability under the distributor's own pipeline predicted identifiability after downstream processing, stage contracts would be unnecessary. The degree-normalization result already argues against it, because a cheap downstream transform raises identification by about 16% on average. Our first ablation measures the gap between single-pipeline and worst-case audits directly. If that gap is small across the design space, we will report that a simple audit suffices and drop the contract machinery.
Why not just add noise or apply differential privacy to the connectomes? We include Gaussian edge noise and differentially private release, a formal guarantee that limits how much any one person's data can change the output, as interventions evaluated under the same worst-case protocol. We expect noise strong enough to defeat attacks across sessions to also hurt sex and age prediction, because both rely on the same spatial structure that Popov et al. show carries the discriminative signal. That expectation is not established. If noise matches subspace truncation on the privacy–utility curve, that is a legitimate finding. The responsibility analysis would still be needed to say which party must apply the noise.
Why not just rely on controlled-access data use agreements? Agreements govern the people who receive the data, not the derived artifacts they later publish, such as embeddings, group eigenspaces or trained models. That is exactly the downward shifting of liability without tooling that Voice App Developer Experiences with Alexa and Google Assistant documents on voice-assistant platforms. Agreements and technical bounds complement each other, and we do not test legal mechanisms.
Proposed Experiments
The primary experiment computes $R(A)$ for five release forms: time series, Pearson connectome, degree-normalized connectome, full GEFF projection, and task-subspace projection. Each is evaluated against all pipelines in $\mathcal{F}$ on HCP 1200, using resting state and seven tasks, matched across days.
First decisive check: composition. If the product of local gains underpredicts end-to-end identification by a wide margin on held-out subjects, identifiability is a closed-guarantee and per-party certification is insufficient. That would be the central result.
Second decisive check: stability across parcellations. Stage attributions must hold when the 374-region parcellation is swapped for a coarser one. Otherwise any assignment of responsibility is an artifact of the parcellation.
Third decisive check: transfer. We apply the best release form to the 100-subject HCP 900 configuration and to ABIDE. On ABIDE, Ge et al. report that high-amplitude filtering moves GCN accuracy from 60.00% to 65.86%, which gives a utility anchor. The protection should hold up beyond the dataset it was designed on.
We will also report where the theory of Information-flow interfaces maps cleanly onto the measured contracts and where quantitative leakage breaks its binary assumptions.
References
- GEFF: Graph embedding for functional fingerprinting
- Improving Functional Connectome Fingerprinting with Degree-Normalization
- Information-flow interfaces
- Defining and Benchmarking a Data-Centric Design Space for Brain Graph Construction
- A simple but tough-to-beat baseline for fMRI time-series classification
- Behavioural relevance of spontaneous, transient brain network interactions in fMRI
- Voice App Developer Experiences with Alexa and Google Assistant: Juggling Risks, Liability, and Security