Does Brain-Signal Complexity Predict Which People an EEG Model Leaks?

Deep models trained on electroencephalography (EEG, scalp recordings of brain electrical activity) are increasingly pooled across many people and shared, sometimes as synthetic-data generators that are explicitly offered as a privacy defense. Privacy research has shown two things that matter here. Large models memorize individual training examples well enough that an attacker can tell whether a given record was used in training, which is called membership inference. And atypical records are memorized most. The EEG literature has separately shown that people differ stably in the nonlinear "texture" of their signals: how irregularity scales from fast to slow timescales and how predictable the dynamics are. These differences are captured by multifractal and predictive-complexity measures. Nobody has asked whether that texture determines who gets exposed.

This note proposes a person-level privacy audit of EEG models and tests one specific claim. A person's atypicality in nonlinear complexity, measured from a short recording before any model is trained, predicts how easily an attacker can detect that person's participation. It does so beyond what simpler atypicality measures (power spectrum, spatial covariance, artifact load) already predict. If the claim holds, data curators get a cheap way to flag high-risk participants before release and an honest way to build stress-test forget sets. If it fails, and spectral or covariance outlierness explains everything, that is also worth knowing, because the much cheaper measures would then be enough.

Core Claim

The attack unit is the person, not the trial. An adversary holds fresh recordings from someone, taken in a session the model never saw, and wants to know whether that person's other sessions were in the training set. Our hypothesis is that per-person membership advantage, defined as the attacker's true-positive rate minus false-positive rate, rises with that person's distance from the population in multifractal-cumulant space. We further hypothesize that this relation survives two controls. The first is conditioning on spectral and covariance atypicality. The second is replacing the person's signals with phase-randomized surrogates: copies with an identical power spectrum but scrambled nonlinear structure. Under surrogates, the predictive power should collapse.

A second, sharper prediction comes from how Euclidean Alignment works. Euclidean Alignment (EA) is a per-person linear whitening that sets each person's mean spatial covariance to the identity. It should erase the covariance fingerprint. Because it is a fixed linear map, though, it should leave much of each channel's scaling structure intact. After EA, covariance atypicality should stop predicting leakage while complexity atypicality keeps predicting it.

Grounding in Prior Work

Two papers motivate the problem. The theory in Parameters or Privacy proves that in the regime with more parameters than samples, an optimal likelihood-ratio attacker approaches perfect membership inference, and that ridge regularization makes this worse. The paper checks the result empirically on generic time series, but it never asks which samples are most exposed or how biosignal structure enters. Underestimated Privacy Risks for Minority Populations shows that rare identifiers in text suffer at least 20% higher PrivLeak than random forget sets, which it calls an "average-case fallacy" in evaluation. We carry its question into a continuous domain, where "minority" is not a discrete identifier but a position in a complexity space.

Three papers supply the measurement tools. Exploring two novel features for EEG-based BCIs defines multifractal cumulants of wavelet leaders and predictive complexity. Its pattern of results hints at strong person heterogeneity: an average gain of 1.4% over band power, but 9.06% for a single subject. Multiscale Fractal Analysis on EEG for Emotion Recognition supplies multiscale fractal dimension and multifractal detrended fluctuation analysis features on DEAP. Those features beat power-spectral features by about 5% for arousal and reach 63% accuracy subject-independently. That suggests they carry person-transferable structure rather than noise. Graph Neural Networks on SPD Manifolds supplies the covariance geometry we use for the strongest competing predictor.

Combining Euclidean Alignment and Data Augmentation both constrains and motivates the design. It shows that high-capacity models such as Deep4Net and EEG-Conformer memorize person-specific patterns. It also shows that EA combined with segment-swapping augmentation raises accuracy by 8.41% overall and by 13.60% for EEG-Conformer. However, it treats memorization only as a generalization problem. Read alongside the regularization paradox, it is unclear whether EA protects participants or merely moves their exposure into features that EA does not touch. We test that directly.

Finally, Let EEG Models Learn EEG presents the JET flow-matching generator on TUH. It argues that synthetic EEG eases privacy restrictions and uses a heavy-tailed L1 prior specifically to keep sharp transients. It runs no membership attack. Diffusion-model membership work (Membership Inference of Diffusion Models) gives reason to expect leakage from such generators.

Research Program

We build three things.

The first is the complexity atypicality score. For each person we take a calibration session that is never used for training or attack, and compute three feature sets per channel: the first five wavelet-leader cumulants across scales, the width of the multifractal detrended fluctuation spectrum, and predictive complexity. A person's score is their Mahalanobis distance from the population mean, computed with a shrinkage covariance estimate. We compute matched scores the same way for log band power in the five classical bands, for Riemannian distance of the person's mean covariance to the population mean, and for amplitude and artifact rate.

The second is a person-level attack harness, run on models trained on random halves of the participant pool. For each model, every person has some sessions in training and others held out. The attack scores a person's held-out-session trials under the target model, either by loss or by calibrated confidence. It compares that score against in and out distributions estimated from shadow models trained with and without that person. This is the likelihood-ratio construction that the overparameterization paper analyzes. Per-person membership advantage and true-positive rate at 1% false-positive rate are the outcome variables.

The third is the intervention arm, which trains the same models with and without EA, with segment-swapping augmentation, with a weight-decay sweep, and with a 25 Hz low-pass filter.

Classifiers are EEGNet, Deep4Net and EEG-Conformer on MOABB motor imagery. Cho2017 (52 subjects) is primary and Shin2017A is the replication; BNCI2014001 is too small for person-level statistics. DEAP (32 subjects) serves as an affective replication, since it is where the fractal features were validated. For generators we train a reduced flow-matching model in the style of JET on a TUAB patient subset, keeping the L1 and spectral-slope terms. Its attack score is the per-window flow-matching loss.

Baselines

The naive reference point is the assumption of uniform risk. Under it, every participant is equally exposed and the model's average membership advantage describes everyone. This is the average-case evaluation that Underestimated Privacy Risks argues hides the people at risk. Beating it means showing statistically significant between-person variance in advantage, tested by permutation over shadow-model draws.

The strongest competing predictor is covariance atypicality: Riemannian distance of a person's mean spatial covariance to the population mean, using the geometry of Graph Neural Networks on SPD Manifolds. It is the strongest competitor because covariance is the dominant source of person identity in motor-imagery EEG, and because EA exists precisely to remove it. Spectral atypicality is the second competitor, since Multiscale Fractal Analysis uses power-spectral features as its reference. All predictors are computed from the same calibration session, with the same shrinkage, and are given the same feature dimensionality through principal-component truncation, tuned on a disjoint fold.

We found no reported result on Lacuna for membership inference, at any granularity, on EEG models. So no published number exists for this competitor, and its values will be established here. Complexity atypicality beats it if adding it to a regression that already contains covariance, spectral and artifact atypicality raises explained variance in per-person advantage by a pre-registered margin. We fix that margin at an increase in held-out Spearman correlation of 0.15, under nested cross-validation over people.

A model-based baseline is the per-person loss of a single reference model trained on everyone. It needs training, but only once. The proposed score must match it to be useful as a pre-release screen.

The key ablation is the surrogate control. We recompute complexity atypicality on phase-randomized copies of each calibration session, which leaves the power spectrum exactly in place. If the surrogate score predicts leakage as well as the real one, the "complexity" effect is spectral and the claim fails.

Why not just…?

Why not just use spectral outlierness, which is cheaper? This is the real threat. Multifractal features and band power are correlated, and a high-amplitude, high-beta person may simply be an outlier on both. The surrogate ablation and the incremental-variance test settle the question. If spectral atypicality explains all the variance, we report that, and recommend that curators screen with band power.

Why not just apply Euclidean Alignment, since it already removes person-specific structure? Combining Euclidean Alignment and Data Augmentation shows that EA improves transfer. It never measures privacy, and the regularization paradox in Parameters or Privacy warns that shrinking population variance can sharpen membership signals. Our EA arm measures exactly this, both on average and for the top decile of atypical people.

Why not just low-pass filter or smooth the data? Smoothing may well lower leakage. Let EEG Models Learn EEG reports that removing heavy-tailed transients degrades clinical event recognition, so the cost is utility. We report leakage and accuracy together under the 25 Hz filter. We also ablate the generator's L1 prior against an L2 prior to test whether keeping spikes increases exposure for patients with frequent transients.

Why not just run the attack itself instead of predicting it? Shadow-model attacks are the ground truth here, and we run them. They need dozens of trained models and access to the final architecture, however. A curator deciding which consenting participants need extra protection before release has neither. The predictor matters only if it approaches the single-reference-model loss baseline. If it does not, we say so.

Decisive Checks

The proposal is confirmed if three conditions hold on Cho2017 and replicate on Shin2017A or DEAP. First, between-person variance in membership advantage is significant. Second, complexity atypicality adds predictive power over covariance and spectral atypicality, and that power vanishes under phase-randomized surrogates. Third, after EA, covariance atypicality loses its predictive power while complexity atypicality retains it.

For the generator, the analogous check asks whether patients in the top complexity-atypicality decile of TUAB show higher flow-matching-loss attack success than the median patient, and whether the L1-versus-L2 ablation moves that gap.

As a final consequence, we compare forget sets chosen by complexity atypicality against random forget sets. If the claim holds, the selected sets should show higher post-unlearning leakage under bounded unlearning (Negative Preference Optimization) than random ones. That would extend the minority-population finding to brain signals. Any failure of the first three checks is reported as a negative result about which EEG properties actually drive exposure.