Slow-Band Attacks: Hiding Adversarial Perturbations in the Brain's Own Frequency Range

When a person lies still in an MRI scanner, the recorded blood-oxygen signal drifts in slow rhythms — roughly one cycle every 12 to 100 seconds. The amplitude of these low-frequency fluctuations (ALFF), and its normalized cousin fALFF, is one of the most widely used biomarkers in clinical neuroimaging: classifiers built on it separate Alzheimer's stages, autism, and depression with high accuracy. Separately, in computer vision, adversarial examples — tiny input perturbations that flip a model's prediction — famously transfer: a perturbation computed against one model fools a completely different, unseen model, with no access to its weights. This proposal asks what happens when these two facts collide. Every model trained on functional MRI reads the same slow band, because that is where the biological signal lives and where every standard preprocessing pipeline deliberately keeps energy (the clinical bandpass filter passes 0.01–0.08 Hz and removes everything else). That shared dependence is exactly the kind of shared representation that makes adversarial perturbations transfer between models.

The core claim is that an adversarial perturbation confined to the 0.01–0.08 Hz band of the fMRI time series will both survive the standard clinical preprocessing pipeline — because the pipeline's bandpass filter cannot distinguish it from genuine neural signal — and transfer across diagnostic model architectures at far higher rates than unconstrained perturbations of equal energy, which the same filter largely destroys. If true, this means the growing fleet of deep models deployed on resting-state fMRI, none of which has ever been adversarially evaluated, is exposed to a zero-query attack that public datasets (ADNI, HCP, ABIDE) make cheap to mount.

Core Claim

The proposal has three testable claims. First, band-constrained transfer: perturbations whose energy lies entirely within slow-5 (0.01–0.027 Hz) and slow-4 (0.027–0.073 Hz) sub-bands transfer across architecturally dissimilar clinical classifiers — 3D convolutional networks, support vector machines on regional spectral features, and sequence models on raw time series — while full-spectrum perturbations of matched energy transfer poorly after preprocessing. Second, the fALFF question: the ratio normalization that makes fALFF robust to physiological noise could either neutralize broadband adversarial shifts (an intrinsic defense) or concentrate vulnerability into a low-dimensional spectral manifold that small band-tuned perturbations flip at will; the literature gives arguments for both sides and no experiment. Third, extension to biometrics: the same mechanism should let an adversary erase or impersonate a subject's functional-connectome fingerprint across unknown identification pipelines, which no work has attempted.

Research Program

The method to be built is a temporal band-projected transfer attack. Writing the perturbation as $\delta$ added to the BOLD time series, each attack step computes a gradient direction $g$ on a surrogate model and then projects the accumulated perturbation back into the slow band:

$$\delta \leftarrow \Pi_{\text{slow}}\!\left(\delta + \alpha\, g\right), \qquad \|\delta\|_2 \le \epsilon$$

Here $\Pi_{\text{slow}}$ is an ideal bandpass filter retaining only 0.01–0.08 Hz (per-voxel, applied along time), $\alpha$ is the step size, and $\epsilon$ is a total-energy budget on the time series. This is the temporal analogue of what Low Frequency Adversarial Perturbation did spatially with a 2D DCT for query-based image attacks, and what LFAA: Crafting Transferable Targeted Adversarial Examples with Low-Frequency Perturbations did with a Gaussian low-pass split for targeted transfer — but neither touches time series, and neither confronts a defense that is itself a temporal bandpass filter matched to the attack band.

The mechanism study matters as much as the attack. The transferability survey Devling into Adversarial Transferability on Image Classification: Review, Benchmark, and Evaluation attributes transfer to shared representations and flat loss minima between models trained on the same distribution, and Adversarial Vulnerability from Interference Between Features in Superposition links vulnerability to features sharing representational capacity. In the fMRI setting the shared representation has a precise spectral address: every pipeline extracts power from the same narrow band that Altered intrinsic brain activity in patients with familial cortical myoclonic tremor and epilepsy: An amplitude of low-frequency fluctuation study established as the canonical ALFF band and that Dissociation of Regional Activity in Default Mode Network in Medication-Naive, First-Episode Somatization Disorder formalized for fALFF. We will measure whether transfer success scales with the fraction of perturbation energy inside that band, which turns the vision-side intuition into a testable spectral law.

Proposed Experiments

Experiment 1 runs on ADNI, the Alzheimer's dataset used by both anchor papers. We train four surrogate families on the public release: a 3D multimodal CNN in the style of Voxel-wise Fusion of Resting fMRI Networks and Gray Matter Volume for Alzheimer's Disease Classification using Deep Multimodal Learning, a support vector machine on regional fused ALFF/fALFF features following Gradual Disturbances of the Amplitude of Low-Frequency Fluctuations (ALFF) and Fractional ALFF in Alzheimer Spectrum, a 1D temporal CNN, and a Transformer over voxel time series. Attacks are computed on each surrogate against held-out target models, and the perturbed time series are passed through a realistic clinical pipeline (motion correction, scrubbing, 0.01–0.08 Hz bandpass, spatial smoothing) before reaching the target. The primary metric is diagnostic flip rate: fraction of d-AD scans pushed to healthy-control (and vice versa), plus a targeted variant forcing a chosen stage. A second dataset, ABIDE for autism — where Gradient of White Matter Functional Variability via fALFF Differential Identifiability links fALFF to identifiability — tests generality; no reported result found on Lacuna for adversarial transferability on ABIDE classifiers, so we establish that baseline ourselves.

Experiment 2 is the band ablation: identical attack optimizer and energy budget, with $\Pi_{\text{slow}}$ replaced in turn by projection onto slow-5 only, slow-4 only, the full classic band, a high-frequency-only band, and no projection at all. The prediction is a monotone relationship between in-band energy fraction and post-preprocessing transfer success, with high-frequency perturbations nearly annihilated by the bandpass filter. A spatial-frequency control — restricting the perturbation to low spatial frequencies of each volume via DCT, mimicking the vision attacks — tests whether spatial smoothness alone, without temporal band placement, transfers; the claim predicts it does not, because the fMRI threat model filters time, not space.

Experiment 3 resolves the fALFF tension directly: identical targets trained on raw ALFF, on ratio-normalized fALFF, and on the spatial iVIP features, all attacked with the same slow-band perturbation. If fALFF's normalization is a defense, its targets flip less; if it concentrates vulnerability, they flip more. Either outcome is informative, and it also answers the robustness question that Voxel-wise Fusion of Resting fMRI Networks and Gray Matter Volume for Alzheimer's Disease Classification using Deep Multimodal Learning left open when it judged ALFF inferior on clean accuracy alone.

Experiment 4 extends to biometrics on HCP-YA, where Functional connectome fingerprinting: Identifying individuals and predicting cognitive function via deep learning reports individual identification accuracy up to 99.5% for the rest-rest pair. We attack a surrogate fingerprinting autoencoder with a universal slow-band perturbation and test whether it degrades identification or impersonates a target subject across unseen pipelines, exploiting the shared-versus-individual subspace decomposition that paper introduced. Functional Connectome Fingerprint Gradients in Young Adults found fingerprints saturate after roughly 13 minutes of scan, which conveniently bounds how much temporal data the attacker must perturb.

Baselines

The "do nothing" reference point is each target model's clean diagnostic performance, because an attack only matters relative to it. On ADNI, the strongest published reference points are the fused ALFF+fALFF support vector machine of Gradual Disturbances of the Amplitude of Low-Frequency Fluctuations (ALFF) and Fractional ALFF in Alzheimer Spectrum, which reports 91.83% accuracy separating d-AD from normal controls (and 81.94% for aMCI, 76.44% for SCD), and the iVIP multimodal 3D-CNN of Voxel-wise Fusion of Resting fMRI Networks and Gray Matter Volume for Alzheimer's Disease Classification using Deep Multimodal Learning, which reports 94.12% test accuracy and 97.79 AUC for AD versus CN. We will reproduce both model families as targets, so flip rates are measured against known clean ceilings rather than against our own re-implementations alone.

The strongest competitor is the best existing transfer attack adapted to this domain. On the ImageNet benchmark, Devling into Adversarial Transferability on Image Classification: Review, Benchmark, and Evaluation reports that DRA (Distribution Relevant Attack), which alters surrogate backpropagation paths, achieved an average success rate of 86.4% across victim models, with the ensemble strategy MBA at 74.1%; DRA is therefore the strongest published transfer method and the one to beat. We will port DRA and momentum-based MI-FGSM to the fMRI surrogates unmodified apart from input shape, giving them the same surrogate models, the same perturbation energy budget $\epsilon$, the same compute, and the same hyperparameter search budget (equal number of tuning trials) as our attack. Beating the baseline means a higher post-preprocessing flip rate at matched $\epsilon$, with the gap widening as preprocessing strictness increases; if DRA matches the slow-band attack even after bandpass filtering, the band-constraint claim is wrong and the proposal's contribution collapses to a negative result.

The key ablation isolates the single ingredient that distinguishes our method: the projection $\Pi_{\text{slow}}$. The same optimizer, surrogate, and budget run with no projection, with a high-frequency projection, and with a spatial-DCT projection gives three controls. The proposal's specific claim is vindicated only if the slow-band projection beats all three after clinical preprocessing; if unconstrained PGD transfer survives the bandpass filter equally well, then the bandpass-defense premise was wrong and that too should be reported.

Why not just…?

Why not just use the existing low-frequency image attacks? Low Frequency Adversarial Perturbation and LFAA restrict perturbations to low spatial frequencies of 2D images, and LFAA's success — an 80.1% targeted transfer rate from ResNet-50 to DenseNet-121 versus 55.2% for C-GSP, and successful fooling of Google Cloud Vision — motivates this proposal rather than duplicating it. The fMRI threat model is different in kind: the defense is a temporal filter tuned to the biological band, and the diagnostic signal is temporal power, not spatial texture. A spatially smooth perturbation with broadband temporal energy is removed by the bandpass exactly like any other noise. This is not just an argument — the spatial-DCT control in Experiment 2 settles it empirically, and if it transfers as well as the temporal version, the vision results carry over and the temporal framing was unnecessary.

Why not just bandpass-filter or adversarially train the classifier? Bandpass filtering cannot defend here by construction, because the attack lives inside the pass band: this is the fMRI-specific replay of what Low Frequency Adversarial Perturbation showed in vision, where low-frequency-restricted attacks defeated low-pass defenses and cut query counts from a 4,000-query median to 1,000. Adversarial training is a genuine threat to the proposal's practical significance, and we do not argue around it: Experiment 3 includes adversarially trained targets, and if training on slow-band perturbations closes the gap, that becomes the paper's defensive contribution. The fALFF-normalization question is similarly left open by design — Altered Spontaneous Brain Activity in Children with Early Tourette Syndrome: a Resting-state fMRI Study shows the ratio suppresses broadband physiological noise, but whether that suppression extends to adversarial shifts is precisely what nobody has measured.

Why not just query the black-box model directly? Query-based attacks like the boundary attack need thousands of model accesses — the 4,000-query median that Low Frequency Adversarial Perturbation reduced to 1,000 — and in clinical deployments scan access is logged, audited, and expensive per query. Zero-query transfer avoids the interface entirely, which is why Model Extraction and Adversarial Transferability, Your BERT is Vulnerable!, QFA2SR: Query-Free Adversarial Transfer Attacks to Speaker Recognition Systems, and Zero-Query Transfer Attacks on Context-Aware Object Detectors all adopt it: perturbations are computed offline on public data (open ADNI and HCP releases) and applied to private clinical models that the attacker never touches. Query attacks remain a complementary baseline worth one comparison row, but they model the wrong adversary for hospital systems.

Why study ALFF at all if iVIP beats it? Voxel-wise Fusion of Resting fMRI Networks and Gray Matter Volume for Alzheimer's Disease Classification using Deep Multimodal Learning showed iVIP features significantly outperform ALFF and fALFF for clean classification ($p < 0.000001$), and we take that result seriously — which is why iVIP-style models are among our targets, not ignored. But that paper's comparison says nothing about robustness, and ALFF/fALFF remain the working clinical biomarker in recent studies of adolescent depression (Altered resting-state prefrontal activity and network topology in adolescents with depression: an fNIRS study), preschool autism (Early identification of autism spectrum disorder in preschoolers by static and dynamic amplitude of low-frequency fluctuations features), and Parkinson's subtypes (MRI structural and functional axial asymmetry in the brain-first versus body-first subtypes of Parkinson’s disease). The open question is whether the "inferior" spectral features are the robust ones — Experiment 3 answers it either way.

Decisive Checks

The proposal stands or falls on three numbers. First, the band ablation must show that post-preprocessing transfer success tracks in-band energy fraction, with slow-5 and slow-4 projections beating unconstrained, high-frequency, and spatial-only projections at matched $\epsilon$. Second, the slow-band attack must beat DRA and MI-FGSM adapted to the same surrogates with the same budget, against targets whose clean ceilings are known (91.83% for the fused-feature SVM, 94.12% for the iVIP-style CNN on ADNI; 99.5% identification for the HCP fingerprinting pipeline). Third, the fALFF-versus-ALFF-versus-iVIP comparison must yield a clear ordering, whichever direction it goes — a defense finding and a vulnerability finding are both publishable, but an inconclusive one is not. If the slow-band attack transfers below the unconstrained baseline, the correct conclusion is that shared spectral bands do not function like shared image statistics, and we will report that as a boundary on how far the vision transferability story extends into biomedical time series.

References

  1. Low Frequency Adversarial Perturbation
  2. LFAA: Crafting Transferable Targeted Adversarial Examples with Low-Frequency Perturbations
  3. Devling into Adversarial Transferability on Image Classification: Review, Benchmark, and Evaluation
  4. Adversarial Vulnerability from Interference Between Features in Superposition
  5. Altered intrinsic brain activity in patients with familial cortical myoclonic tremor and epilepsy: An amplitude of low-frequency fluctuation study
  6. Dissociation of Regional Activity in Default Mode Network in Medication-Naive, First-Episode Somatization Disorder
  7. Voxel-wise Fusion of Resting fMRI Networks and Gray Matter Volume for Alzheimer's Disease Classification using Deep Multimodal Learning
  8. Gradual Disturbances of the Amplitude of Low-Frequency Fluctuations (ALFF) and Fractional ALFF in Alzheimer Spectrum
  9. Gradient of White Matter Functional Variability via fALFF Differential Identifiability
  10. Functional connectome fingerprinting: Identifying individuals and predicting cognitive function via deep learning
  11. Functional Connectome Fingerprint Gradients in Young Adults
  12. Altered Spontaneous Brain Activity in Children with Early Tourette Syndrome: a Resting-state fMRI Study
  13. Model Extraction and Adversarial Transferability, Your BERT is Vulnerable!
  14. QFA2SR: Query-Free Adversarial Transfer Attacks to Speaker Recognition Systems
  15. Zero-Query Transfer Attacks on Context-Aware Object Detectors
  16. Altered resting-state prefrontal activity and network topology in adolescents with depression: an fNIRS study
  17. Early identification of autism spectrum disorder in preschoolers by static and dynamic amplitude of low-frequency fluctuations features
  18. MRI structural and functional axial asymmetry in the brain-first versus body-first subtypes of Parkinson’s disease