Brain scans can be summarized as a functional connectome, a matrix recording how strongly the activity of each pair of brain regions rises and falls together. These matrices are useful for predicting sex, disease, or the task a person is performing. They are also a fingerprint: a few dozen connections can pick a person out of a cohort. Training a predictor on connectomes under differential privacy, which mathematically limits how much any one person can change the trained model, is therefore attractive. The standard tool, differentially private stochastic gradient descent (DP-SGD), clips each person's gradient and adds Gaussian noise. That noise grows with the number of input dimensions, and a 400-region connectome has tens of thousands of them, while cohorts have a few hundred to a thousand people. Most of that noise lands on connections that never carried a reliable signal.
Hardware security engineers face a similar problem when they build physical unclonable functions (PUFs), circuits whose manufacturing quirks serve as a device key. Their answer is enrollment. Before a cell is trusted, it is measured repeatedly across temperatures and voltages, and cells that flip are thrown out or marked as a third "unstable" state. We propose the same discipline for private brain-network learning. A disjoint, consented enrollment cohort with repeated scanning sessions is used to measure, for every connection, how large its task signal is compared with how much it drifts from session to session within a person. Only high-margin connections enter DP-SGD, quantized to three levels so that marginal values abstain. Gradients are clipped per subject after averaging across that subject's sessions, which mirrors the PUF practice of majority voting over repeated reads. The claim is that this improves accuracy at a fixed privacy budget. It may also lower empirical re-identification risk, and the program tests that separately.
Core Claim
At matched privacy budget $(\epsilon,\delta)$ with the subject as the privacy unit, DP-SGD on connectomes screened by cross-session reproducibility margin should reach higher held-out accuracy than DP-SGD on the full connectome. It should also beat variance-based or leverage-based selections of the same size and selection by principal component analysis (PCA) on public data. The specific new ingredient is the stability measurement: repeated sessions play the role of PUF operating points. Selecting on stability is different from selecting on variance, which is dominated by hub regions, and from selecting on identifiability, which favors exactly the connections that make a person recognizable.
For a binary label, the margin of edge $e$ is
$$m_e = \frac{|\mu_e^{(1)} - \mu_e^{(0)}|}{s_e^{\text{within}}},$$
where $\mu_e^{(c)}$ is the mean connectivity of edge $e$ in class $c$ over the enrollment cohort, and $s_e^{\text{within}}$ is the pooled within-subject standard deviation of that edge across sessions. This is the analogue of the arbiter-PUF "margin of victory" relative to environmental jitter. Retained edges are then ternarized, in the spirit of discarding indeterminate PUF states:
$$\tilde{x}_e = \operatorname{sign}(x_e - \tau_e)\ \text{if}\ |x_e - \tau_e| > \kappa\, s_e^{\text{within}},\ \text{else}\ 0.$$
Here $\tau_e$ is the enrollment midpoint between class means and $\kappa$ sets the width of the abstain band. Ternarization also bounds input magnitudes, which tightens the norm of first-layer gradients before clipping. For multi-class targets such as 7-task identification, the numerator becomes the between-class standard deviation of edge means.
How This Sits Against Prior Work
The PUF literature supplies the method. Majzoobi et al.'s FPGA arbiter PUF shows that challenges with large delay margins resist temperature and voltage shifts, and that majority voting over repeated evaluations shrinks the metastability window to 2.5 ps (Automated Design, Implementation, and Evaluation of Arbiter-based PUF on FPGA using Programmable Delay Lines). Baek et al.'s SRAM PUF discards bits that flip across operating points. This cuts bit error rate from 13.7% to 0.9%, at a discard rate of up to 47% (A Reconfigurable SRAM Based CMOS PUF With Challenge to Response Pairs). Ternary schemes mark marginal cells as unusable instead of forcing a bit (Bertrand Cambou). These papers measure stability in hardware. We transplant the measurement protocol, not the circuits.
The connectomics literature both motivates and constrains the design. Ravindra et al. show that identity concentrates in 61 to 95 high-leverage edges, reaching 95.5% identification on HCP (Constructing Compact Brain Connectomes for Individual Fingerprinting). Degree normalization shows that hubs mask idiosyncratic structure, and normalizing them away raises identification rate by 16% (Improving Functional Connectome Fingerprinting with Degree-Normalization). Both papers optimize for identifiability. Under privacy, that is the wrong objective: the edges they keep are the ones an attacker wants. Our margin divides task signal by within-subject drift, so it does not reward between-subject idiosyncrasy. We will measure how much the retained set overlaps the leverage-score set. Kim and Ye's graph isomorphism network reaches 84.61% non-private sex classification on HCP with 400 regions (Understanding Graph Isomorphism Network for rs-fMRI Functional Connectivity Analysis). That gives the non-private ceiling.
Three papers fix the privacy side. Approx-Clip-SVD privatizes connectivity matrices by output perturbation for visualization and group tests on FBIRN, BLSA, and ADNI, but trains no predictor (Privacy-Preserving Visualization of Brain Functional Connectivity). Node-level DP-GNN bounds sensitivity through degree caps on sparse graphs (Node-Level Differentially Private Graph Neural Networks). Its setting differs from ours: each brain graph is one person, so we need graph-level privacy, which ordinary per-example DP-SGD gives directly without degree caps. Bagdasarian et al. show that DP-SGD hurts atypical subgroups most (Differential Privacy Has Disparate Impact on Model Accuracy). This constrains us, because discarding edges could amplify that harm, so we track subgroup accuracy throughout.
The research file floats a different PUF-to-DP translation: discarding "marginal" per-sample gradients inside DP-SGD. We reject it. A gradient-dependent discard rule must itself be privatized, and Zhang et al. show that throwing away clipped mass creates bias (Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach). Moving the stability test to an enrollment phase on disjoint data sidesteps both problems.
Research Program
The first step builds the enrollment protocol on HCP. HCP subjects have multiple resting-state and task sessions, and subjects are split into a disjoint enrollment set and a private training set. Every edge gets three quantities: the margin $m_e$, a within-subject stability score, and a between-subject identifiability ratio. The analysis then maps how these relate, and in particular whether high-margin edges for sex or task are separable from Ravindra-style fingerprint edges.
The second step trains under subject-level DP-SGD. Each training subject's gradient is the average over that subject's sessions and is clipped once, so the guarantee covers the person rather than a single scan. Rényi accounting is applied to the subject count. Models are a logistic or two-layer MLP on screened edges and a GIN on the screened, ternarized graph.
The third step evaluates three things. Utility is accuracy across $\epsilon \in \{1,2,4,8\}$ with $\delta = 1/n$. Empirical leakage is measured two ways: membership-inference AUC, and a linkage attack in which an adversary holding a held-out session of a training subject tries to match it to that subject using model embeddings. Disparate impact is accuracy for the smallest demographic subgroups. A variant replaces public enrollment with a privately computed margin, using noisy per-edge statistics under a share of the budget. This covers settings where no consented enrollment cohort exists.
Baselines
The naive reference is subject-level DP-SGD on the full upper-triangular connectome with standard clipping, using the same models, epochs, and hyperparameter grid. Tuning trials are counted toward reporting, or tuned on the enrollment set to avoid spending private budget. We also report non-private training on the full connectome as a ceiling. The non-private GIN figure on HCP sex classification is 84.61% accuracy (Understanding Graph Isomorphism Network for rs-fMRI Functional Connectivity Analysis). For 7-task identification, Ravindra et al. report 91.93% with sparse features and 84.95% dense (Constructing Compact Brain Connectomes for Individual Fingerprinting). I found no reported result on Lacuna for DP-SGD on HCP connectome classification, so all private numbers will be produced here under one shared protocol.
The strongest competitor is public-data dimensionality reduction followed by DP-SGD. PCA is fit on the enrollment cohort, connectomes are projected to $k$ components, and training proceeds privately. It attacks the same dimension-versus-noise problem with no PUF idea at all, so it is the most likely to match us. A second strong competitor is DiceSGD, error-feedback clipping, on the full connectome. Its reported CIFAR-100 ViT-small accuracy at $\epsilon=2$ is 86.5%, versus 78.9% for standard clipped DP-SGD (Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach). That is a different domain, so we rerun it on our data. Third, we test leverage-score edge selection and degree-normalized input, each computed on the enrollment cohort, at matched feature count $k$.
The key ablation holds the edge count at $k$ and changes only the selection rule. One arm ranks edges by between-class difference over total cross-subject variance from a single session, which needs no repeated measurement. The other arm ranks by $m_e$, which uses cross-session drift. If the single-session rule matches $m_e$, the PUF ingredient adds nothing beyond ordinary univariate screening. Two further ablations isolate the rest of the pipeline: ternarized versus continuous screened inputs, and session-averaged versus single-session per-subject gradients. Beating the baselines means a higher accuracy curve across the $\epsilon$ range, with non-overlapping confidence intervals over 10 seeds and repeated subject splits at $\epsilon \le 4$, and no worse subgroup gap.
Why not just…?
Why not just use public-data PCA before DP-SGD? This is the real threat. It reduces dimension just as screening does, and it is cheap. Variance-ranked components on connectomes tend to follow hub structure, which the degree-normalization study shows dominates raw matrices (Improving Functional Connectome Fingerprinting with Degree-Normalization). Hub structure may be neither label-relevant nor session-stable. That is an argument, not evidence. The matched-$k$ comparison settles it. If PCA ties, the honest conclusion is that dimension alone explains the gain, and the contribution shrinks to the subject-level, session-averaged protocol.
Why not just fix clipping with DiceSGD? Error feedback removes clipping bias. It does not reduce the per-coordinate Gaussian noise that dominates when $d$ is in the tens of thousands and $n$ is around a thousand. It also triples memory and needs a larger noise multiplier (Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach). The two methods are complementary. We therefore also run DiceSGD on screened inputs, and the result is meaningful in either direction.
Why not just privatize the matrices once and train non-privately? Approx-Clip-SVD releases noisy, denoised matrices for visualization and group tests (Privacy-Preserving Visualization of Brain Functional Connectivity). Every subject's full matrix is perturbed, so the noise is paid on all $d$ edges per person. That is worse scaling than DP-SGD's noise on one aggregated gradient per step. We include its released matrices, followed by non-private training at matched $\epsilon$, as a baseline. Approx-Clip-SVD reports no exact MSE values, so its numbers are generated here.
Proposed Experiments and Decisive Checks
The primary benchmark is HCP. We use 400-region Schaefer connectomes for sex classification and task-fMRI connectomes for 7-task identification. HCP's multiple sessions supply both the enrollment drift estimates and the linkage-attack probes. The secondary benchmark is FBIRN schizophrenia versus control (1,056 subjects), with ADNI as a further check. These datasets offer fewer clean repeat sessions, so within-subject drift will come from split-half time series. That is a weaker proxy, and we will report it as such.
The first decisive check is the matched-$k$ ablation: margin selection versus single-session univariate selection versus enrollment PCA at $\epsilon = 2$ and $\epsilon = 4$. The second is the linkage attack. If margin-screened models leak no less than full-connectome models at the same $\epsilon$, the empirical-leakage claim fails, even though the formal guarantee still holds. The third tracks the subgroup accuracy gap under screening. Bagdasarian et al.'s results suggest that discarding edges tuned to the majority could widen it (Differential Privacy Has Disparate Impact on Model Accuracy). If it does, the margin will be recomputed per subgroup on the enrollment set. Finally, sweeping the discard fraction traces a curve analogous to the SRAM PUF trade-off between bit error rate and discard rate. A discard fraction near 50% that still helps would mirror the hardware finding that aggressive discard pays off in high-noise regimes. A peak at low discard would say connectome signal is more diffuse than PUF entropy.
References
- Automated Design, Implementation, and Evaluation of Arbiter-based PUF on FPGA using Programmable Delay Lines
- A Reconfigurable SRAM Based CMOS PUF With Challenge to Response Pairs
- A Proof of Concept SRAM-based Physically Unclonable Function (PUF) Key Generation Mechanism for IoT Devices
- Constructing Compact Brain Connectomes for Individual Fingerprinting
- Improving Functional Connectome Fingerprinting with Degree-Normalization
- Understanding Graph Isomorphism Network for rs-fMRI Functional Connectivity Analysis
- Privacy-Preserving Visualization of Brain Functional Connectivity
- Node-Level Differentially Private Graph Neural Networks
- Differential Privacy Has Disparate Impact on Model Accuracy
- Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach