Machines that read emotion and cognition from scalp EEG — a brain-computer interface measuring the brain's electrical rhythms through the scalp — lean overwhelmingly on the fastest of those rhythms. Across many studies, the slow oscillations (delta, theta, alpha, below roughly 14 Hz) carry little affective information, while Beta (14–30 Hz) and Gamma (above 30 Hz) activity in temporal, prefrontal, and parietal cortex carries most of it (Exploring EEG Features in Cross-Subject Emotion Recognition, Instance-Adaptive Graph for EEG Emotion Recognition). In parallel, adversarial machine learning has learned a hard lesson: defenses evaluated only against fixed, pre-built attacks create an illusion of safety, and collapse once the adversary explicitly models the defense (DeepCloak: Masking Deep Neural Network Models for Robustness Against Adversarial Samples, MART: Improving LLM Safety with Multi-round Automatic Red-Teaming). Nobody has brought these two lessons together: there is no account of what a defense-aware attacker does to a frequency-specific biomarker decoder.
This proposal's core idea can be stated in two sentences. The frequency bands that carry the emotion signal are also the best place to hide an attack: an attacker who confines perturbations to Beta and Gamma, at amplitudes matching each subject's natural per-band statistics, can flip emotion predictions with signals that no band-pass filter or visual inspection can distinguish from genuine neural activity, because the attack lives inside the biomarker rather than beside it. The only defense with a structural reason to survive such an attacker is one that randomizes the spectral processing itself per trial, so the attacker faces a moving filter rather than a fixed one.
Core Claim
Two claims, one offensive and one defensive. Offensively: a spectrally constrained, defense-aware test-time attacker — optimizing perturbations in the frequency domain under per-band power budgets drawn from the subject's own natural band statistics — will achieve attack success rates on affective EEG decoders comparable to unconstrained white-box attacks, because it exploits the same high-frequency features the decoder depends on. Prior test-time EEG attacks are either unconstrained and non-adaptive (Universal Adversarial Perturbations for CNN Classifiers in EEG-Based BCIs, which leaves emotion graph models and adaptive evaluation explicitly open) or naturally pooled their energy in low frequencies on regression tasks (White-Box Target Attack for EEG-Based BCI Regression Problems, whose drowsiness and vigilance perturbations concentrated in 3–10 Hz). Neither has asked what happens when the attacker is forced into the informative bands and knows the defense. Defensively: deterministic defenses — fixed band-pass filtering, static channel or frequency masking, adversarial training against a frozen attack — fail by construction against this attacker, mirroring DeepCloak: Masking Deep Neural Network Models for Robustness Against Adversarial Samples's finding that static masks merely shrink the attacker's search space, and the DIS reward in Professor X: Manipulating EEG BCI with Invisible and Robust Backdoor Attack, which shows that scattering energy across bands defeats static filters. What should survive is per-instance randomized spectral processing, extending the randomized-geometry intuition of Adversarial Defense without Adversarial Defense : Enhancing Language Model Robustness via Instance-level Principal Component Removal — whose authors themselves flag adaptive evaluation as unproven — from embedding spaces to the spectral domain.
Research Program
The first build is the attack tool. A perturbation $\delta$ is optimized in the Fourier domain under two constraints: a band projection $\mathcal{P}_B$ that confines its energy to an attacker-chosen set of frequency bins (typically Beta and Gamma), and a per-band physiological budget $\|\mathcal{P}_b \delta\|^2 \le c\,\sigma_b^2$, where $\sigma_b^2$ is the subject's natural power in band $b$ and $c$ is a small multiplier. Concretely, the attacker solves
$$ \max_{\delta}\; \mathcal{L}\big(f(x+\delta),\, y_{\text{target}}\big) \quad \text{s.t.} \quad \delta = \mathcal{P}_B\delta,\;\; \|\mathcal{P}_b\delta\|^2 \le c\,\sigma_b^2 \;\text{for each band } b, $$
where $f$ is the target decoder, $x$ the clean trial, and $\mathcal{L}$ a target-class loss in the spirit of the Carlini–Wagner objective. The defense-aware variant wraps the objective in an expectation over the defense's randomness, $\mathbb{E}_{\theta\sim p(\theta)}[\mathcal{L}(f_\theta(x+\delta), y_{\text{target}})]$, so the attacker optimizes against the defense distribution rather than one realization — the honest version of adaptive evaluation that the survey Interpretable and Robust AI in EEG Systems: A Survey explicitly notes is missing from EEG robustness research.
The second component is a vulnerability map. We run the attack against the strongest affective decoders in the file: the Instance-Adaptive Graph, which builds a per-sample brain graph from all five bands and reports 95.44% subject-dependent and 86.30% subject-independent accuracy on SEED (Instance-Adaptive Graph for EEG Emotion Recognition); the Regularized Graph Neural Network with its biologically wired adjacency and subject-invariance training, at 94.24% subject-dependent on SEED and 73.84% subject-independent on SEED-IV (EEG-Based Emotion Recognition Using Regularized Graph Neural Networks); a static CNN in the EEGNet family; and the linear Support Vector Machine on Hjorth Beta mobility features that reached 83.33% under leave-one-subject-out evaluation (Exploring EEG Features in Cross-Subject Emotion Recognition). Two open questions get settled here. First, whether Beta/Gamma-constrained attacks match unconstrained attack success on affective tasks, which would establish that the informative bands are the vulnerable bands. Second, whether the per-instance graph generation acts as an intrinsic moving target that resists perturbations — a possibility neither graph paper tests, since both evaluate only clean accuracy — or instead hands the attacker a differentiable surface it can manipulate directly.
The third component is the defense: per-instance randomized spectral partitioning. Instead of fixed band boundaries at 4, 8, 14, and 30 Hz, each trial's band edges, and the seed of any graph coarsening downstream, are drawn from a defender-held distribution. The attacker must then average over the randomization, and we measure how much its attack cost grows — drawing on the cost-asymmetry framing of Optimizing Adaptive Attacks against Content Watermarks for Language Models, where a surrogate attacker stripped watermarks with 96–98% evasion at under $10 of compute, and on the black-box query regime of Query-Based Adversarial Prompt Generation. Finally, we run a multi-round loop in the spirit of MART: Improving LLM Safety with Multi-round Automatic Red-Teaming: the attacker adapts to the deployed defense, the defender updates, and we track whether security claims made after round one survive round four — a discipline the adaptive-attack literature shows round-one evaluations routinely fail to provide.
Baselines
The naive reference point is an undefended model under random noise. This matters because it calibrates whether the threat is real: on the Driving regression dataset, random Gaussian noise achieved an attack success rate of 0.00% while the crafted CW-R and IFGSM-R attacks reached nearly 100% (White-Box Target Attack for EEG-Based BCI Regression Problems). Any defense we propose must be judged against crafted attacks, not noise, and we adopt that distinction throughout.
The strongest attack baseline is an unconstrained white-box attack with no spectral constraint — a classification adaptation of the CW-R and IFGSM-R optimizers, which reported nearly 100% attack success on both the Driving and PVT regression datasets, and whose sibling result, the Total Loss Minimization universal perturbation, dropped P300 speller CNN accuracy from approximately 81% to 17% with near 100% targeted success (Universal Adversarial Perturbations for CNN Classifiers in EEG-Based BCIs). This baseline answers the question our title asks: if the constrained in-band attack reaches, say, eighty percent of the unconstrained attack's success at equal perturbation budget, the informative bands are indeed the privileged attack channel. We will run both attacks against identical models on identical splits with the same tuning budget, and count the constrained attack as established only if its success rate and perturbation norms are within a small factor of the unconstrained ones.
The strongest defense competitors are the field's defaults. First, static spectral preprocessing: the band-pass and notch filtering that every pipeline in the biomarker literature applies, which directly removes out-of-band energy and is therefore the natural first defense; its predicted failure mode is exactly the distribution-scattering evasion that Professor X: Manipulating EEG BCI with Invisible and Robust Backdoor Attack already engineered against filters, and our defense-aware attacker will test it head-on. Second, adversarial training on raw waveforms, whose cost profile is documented at roughly 3–4 times normal training with overfitting to the training attack in Adversarial Defense without Adversarial Defense : Enhancing Language Model Robustness via Instance-level Principal Component Removal. Third, a port of that paper's PURE method — per-instance principal component removal — to EEG feature spaces, which we implement ourselves since no reported result found on Lacuna for PURE applied to EEG. All defenses get the same training data, the same compute and tuning budget, and the same clean-accuracy requirement: within about one percent of the Instance-Adaptive Graph's 95.44% subject-dependent SEED accuracy, so that robustness is not bought by crippling the decoder.
The key ablation isolates the defense's single ingredient: the same randomized-partitioning architecture run with a fixed, deterministic partition — same model, same compute, only the randomness removed. If the deterministic variant collapses under the defense-aware attacker while the randomized one holds, the randomness itself, not the architecture, is doing the defensive work. A second ablation runs the same attack twice, once against a frozen defense and once with the defense in the loop, so that any robustness claim we report is explicitly tagged as non-adaptive or adaptive.
Why not just…?
Why not just filter harder? Aggressive preprocessing — independent component analysis for ocular artifacts, wavelet denoising, strict band-pass filtering — is the standard answer in the biomarker literature, and it genuinely removes out-of-band noise. But it cannot remove an in-band attack: the perturbation occupies the same bins as the biomarker, at the subject's own natural power levels, so a filter strong enough to remove it also removes the signal the decoder needs. The frequency-domain injection route was already demonstrated by Professor X: Manipulating EEG BCI with Invisible and Robust Backdoor Attack, whose distribution-scattering reward exists precisely to defeat band-stop filters, and the regression attacks of White-Box Target Attack for EEG-Based BCI Regression Problems landed inside physiological 3–10 Hz rhythms for the same reason. Our experiments make this failure measurable rather than rhetorical: we will report the static filter's attack success rate under the defense-aware attacker, and expect it near the undefended level.
Why not just adversarially train the model? Standard adversarial training on raw waveforms is the obvious off-the-shelf defense, and we will run it as a baseline. Two documented problems temper expectations. It costs roughly 3–4 times normal training and overfits the specific attack used during training (Adversarial Defense without Adversarial Defense : Enhancing Language Model Robustness via Instance-level Principal Component Removal), and an attacker who reoptimizes against the defended model — rather than replaying the training attack — routinely bypasses it, as multi-round red-teaming and surrogate-based evasion repeatedly show (MART: Improving LLM Safety with Multi-round Automatic Red-Teaming, Optimizing Adaptive Attacks against Content Watermarks for Language Models). If adversarial training turns out to hold against our defense-aware attacker, that is a real finding and we will report it as such; our closed-loop protocol is designed to make that test fair rather than rigged.
Why not just use simple linear models? The feature bake-off of Exploring EEG Features in Cross-Subject Emotion Recognition showed that a Support Vector Machine on Hjorth Beta mobility reaches 83.33% under leave-one-subject-out evaluation, so one might avoid deep models entirely. The evidence says simplicity does not confer immunity: the regression attacks achieved nearly 100% success against linear Ridge Regression specifically (White-Box Target Attack for EEG-Based BCI Regression Problems), and universal perturbations crafted on CNNs transferred to Common Spatial Patterns plus Logistic Regression pipelines (Universal Adversarial Perturbations for CNN Classifiers in EEG-Based BCIs). We include the SVM pipeline as an attack target precisely to test this directly on affective data.
Why not just mask or drop channels randomly at inference? Static masking is mapped and bypassed by adaptive attackers who shift energy into the unmasked subspace (DeepCloak: Masking Deep Neural Network Models for Robustness Against Adversarial Samples), and "mini-UAPs" confined to small channel subsets remained potent in the BCI setting (Universal Adversarial Perturbations for CNN Classifiers in EEG-Based BCIs). This is a genuine near-neighbor of our defense, and the difference matters: channel masking randomizes which features are seen, while our scheme randomizes how the spectral axis is carved up before any feature exists, so the attacker cannot know which bins belong to which band on a given trial. Our deterministic-partition ablation exists to separate these, and we will also run a channel-dropout defense under the adaptive attacker to check whether our reasoning about the distinction is empirically right.
Proposed Experiments and Decisive Checks
The vulnerability map runs on SEED, SEED-IV, MPED, and DEAP — 62-channel affective EEG across three and four emotion classes, seven emotion classes, and continuous valence-arousal ratings respectively — covering subject-dependent and subject-independent protocols as defined in the graph papers (Instance-Adaptive Graph for EEG Emotion Recognition, EEG-Based Emotion Recognition Using Regularized Graph Neural Networks). On MPED the Instance-Adaptive Graph reports 74.77% on the harder seven-class problem, versus approximately 42% for the Attention-LSTM of MPED: A Multi-Modal Physiological Emotion Database for Discrete Emotion Recognition, giving us a difficulty gradient. Metrics are attack success rate under a fixed perturbation budget, per-band perturbation energy, a spectral-plausibility score (the distance between attacked and natural per-band statistics, since the survey warns that Beta/Gamma models can already cheat on electromyographic contamination — Interpretable and Robust AI in EEG Systems: A Survey), and cross-subject and cross-context attack transfer, the latter using the acquisition-context variation documented in A Multi-Context Emotional EEG Dataset for Cross-Context Emotion Decoding. A secondary analysis tests whether the subject-invariance training of the Regularized Graph Neural Network smooths away individual anatomical quirks that currently block attack transfer — its NodeDAT and EmotionDL modules cost 2–4% clean accuracy when removed on SEED-IV, so the security side of that trade-off deserves measurement.
Three checks decide the proposal. First, on the offensive claim: the Beta/Gamma-constrained attack must reach a large fraction of the unconstrained attack's success rate at equal budget on affective tasks; if it does not, the informative-bands-as-attack-channel hypothesis is wrong and the frequency-specific angle collapses to ordinary adversarial robustness. Second, on the defensive claim: under the defense-aware attacker, static filtering and the deterministic ablation must fail while the randomized partition holds, with clean accuracy within about one percent of the undefended model; if the deterministic variant also holds, randomness is not the active ingredient and the defense story must be rewritten. Third, the multi-round loop must report security after each round, so that no claim depends on a single frozen adversary — the discipline that separates this work from the round-one evaluations the adaptive-attack literature has repeatedly shown to overstate safety.
References
- Exploring EEG Features in Cross-Subject Emotion Recognition
- Instance-Adaptive Graph for EEG Emotion Recognition
- DeepCloak: Masking Deep Neural Network Models for Robustness Against Adversarial Samples
- MART: Improving LLM Safety with Multi-round Automatic Red-Teaming
- Universal Adversarial Perturbations for CNN Classifiers in EEG-Based BCIs
- White-Box Target Attack for EEG-Based BCI Regression Problems
- Professor X: Manipulating EEG BCI with Invisible and Robust Backdoor Attack
- Adversarial Defense without Adversarial Defense : Enhancing Language Model Robustness via Instance-level Principal Component Removal
- Interpretable and Robust AI in EEG Systems: A Survey
- EEG-Based Emotion Recognition Using Regularized Graph Neural Networks
- Optimizing Adaptive Attacks against Content Watermarks for Language Models
- Query-Based Adversarial Prompt Generation
- MPED: A Multi-Modal Physiological Emotion Database for Discrete Emotion Recognition
- A Multi-Context Emotional EEG Dataset for Cross-Context Emotion Decoding