Interfaces that read the user's body — EEG headsets, heart-rate straps, eye trackers — are moving onto cheap edge hardware like the Raspberry Pi. There they run a closed loop: sense physiology, infer mental workload, and simplify or automate the interface before the user overloads. But the memory of such a device does not forget when the power is cut. In a cold-boot attack, someone with brief physical access chills the memory chips with an inverted air duster, reboots the device, and dumps whatever charge remains. This attack has been shown to steal deep-learning models from edge hardware with near-perfect fidelity, but only for ordinary image classifiers. Nobody has asked what a cold boot reveals when the memory holds brain and body signals, per-user calibration data, and a classifier of the owner's mental state. That question changes the harm: the stolen artifact is not just a model but a record of a person's moment-to-moment cognition.
Our claim has three parts. First, a commodity cold boot on a neuroadaptive edge device leaks three distinct artifact classes — raw physiological window buffers, user-specific calibration matrices, and the cognitive-load classifier — and from the raw buffers an attacker can reconstruct the victim's past mental states and re-identify the victim across sessions, a threat with no analogue in prior edge-model theft. Second, because the biocybernetic loop already computes a sensitive inference (the user's current cognitive load) every cycle, that same signal can drive tiered memory protection: physiological buffers and calibration state get the strongest lifetime management at all times, while less sensitive state is protected only when inferred load or task stakes are high, keeping the security tax inside the loop's latency deadline. Third, the escalation thresholds and the severity taxonomy for leaked artifacts are value-laden questions with no experimental ground truth, which is precisely the regime where structured expert consensus — the Delphi method — is the established instrument, and where, in a twist we also test, the panelists themselves can be physiologically monitored to detect the fatigue-driven false convergence that Delphi practitioners already worry about.
Core Claim
The file's own survey confirms the intersection is empty. DeepFreeze executed cold boots against vision models on exactly the hardware class we target (an Intel Neural Compute Stick 2 on a Raspberry Pi), recovering architectures with a "100% success rate" and weights with a "0.04%" error rate — but it never touched physiological models or the buffers streaming through them. T-Edge builds a TrustZone-backed trusted execution environment for edge SoCs and openly concedes that "cold-boot attacks on RAM are mentioned as a potential vulnerability if memory encryption is not available." On the other side, the neuroadaptive literature warns about "neuro-hacking" and mental-state leakage (Embracing Trustworthy Brain-Agent Collaboration; CogErgLLM) but treats security purely at the algorithmic or network layer, assuming the hardware beneath is safe. And Delphi has been used to validate clinical scales and privacy metrics for synthetic data (A Consensus Privacy Metrics Framework for Synthetic Data; VISION-Cog) — never to set security policy for adaptive physiological systems. We propose to occupy that void with a concrete, measurable system study rather than another position paper.
Research Program
Workstream 1 builds and attacks the target system. We train cognitive-load classifiers on the CL-Drive dataset — 21 subjects, four sensor modalities (EEG, ECG, EDA, eye gaze), nine scenario complexities — where the published XGBoost baseline reaches "83.67%" binary and "74.08%" ternary accuracy under 10-fold cross-validation (Multimodal Brain-Computer Interface for In-Vehicle Driver Cognitive Load Measurement), and a transfer-learning variant reaches "74.07%" accuracy with an F1 of "70.28%" (EEG-based Cognitive Load Classification using Feature Masked Autoencoding and Emotion Transfer Learning). We deploy the loop on a Raspberry Pi with the physiological stream replayed in real time, then run the DeepFreeze protocol verbatim — air-duster cooling, reboot into an attacker image, memory dump — at multiple extraction temperatures, since DeepFreeze documents steep recovery degradation between cryo-cooled baselines and −20°C/−10°C extraction. The new measurement is post-extraction inference: using only the dumped bytes, can the attacker (a) classify the victim's cognitive states per window, (b) re-identify which subject the buffers came from across sessions, and (c) rebuild a working classifier via the Dropout knowledge-distillation recovery DeepFreeze introduced, which reduced relative accuracy drop to "around 0.01-0.04%" using 10% unlabeled data on image models? The CL-Drive cross-subject gap — accuracy falling from "83.67%" under cross-validation to "76.17%" under leave-one-subject-out — is our handle on (b): user-specific physiological "accents" that make buffers fingerprintable are the same phenomenon that hurts generalization, and Embracing Trustworthy Brain-Agent Collaboration reports "15-30% of users fail to achieve reliable control" for related reasons.
Workstream 2 builds the defense: load-aware data lifetime management. The key observation is that the strongest existing hardware mechanism, the per-row live-flag gating of Secure DRAM (0.1% performance degradation versus 5.1% average and 12.8% allocation-heavy slowdown for software zeroing), only clears dead rows — it does nothing for live streaming buffers at the moment a running device is seized, which is exactly the cold-boot scenario. Our scheme therefore combines always-on protection for the narrow class of physiological artifacts (raw windows held only in on-chip secure memory or encrypted in DRAM with keys confined to the trust boundary, exploiting the latency savings of in-situ approaches like FeFET in-situ encryption, which reports a "90% reduction in latency" versus external AES engines) with adaptive, tiered handling of everything else, gated by the loop's own inferred cognitive state. The design question is whether the encryption and zeroing overhead fits the loop's real-time deadline, against the known costs of counter-mode encryption on persistent memory documented by SecPM.
Workstream 3 runs the Delphi study, and instruments it. A panel of roughly 15–20 experts spanning BCI, HCI, hardware security, and privacy regulation completes three anonymous rounds with controlled feedback, using Content Validity Ratio scoring as in VISION-Cog and Content Validity Index filtering as in the traumatic-injury follow-up instrument of Wake et al.. The panel ranks the severity of each recoverable artifact class and sets the load thresholds at which the defense escalates. Separately, we monitor panelists' own physiology (pupillometry and heart rate via egocentric cameras, in the spirit of egoPPG, for which no reported estimation-error figure was found on Lacuna, so we will validate the monitor separately) to test whether detected panelist fatigue predicts drift toward the group median without defensible justification — the false-convergence failure mode that Delphi's own literature flags as an open problem (AI Agents and the Future of Deliberation).
Baselines
The naive reference point is an unmodified Raspberry Pi stack with no memory protection, exactly reproducing the DeepFreeze configuration. On this baseline the attack side is essentially settled by prior numbers: architecture recovery at a "100% success rate" and weight error of "0.04%" with at most a "0.5%" accuracy drop (DeepFreeze); our contribution is the new leakage measurements (victim-state inference and re-identification) on top of it. The strongest defense competitor is always-on protection with no adaptivity, run in two variants: blanket software zeroing of all buffers, whose cost is documented at "5.1%" average and "11.2%" peak slowdown on SPEC CPU 2006 and "12.8%" on forkbench (DRAM architecture for efficient data lifetime management); and a hardware live-flag scheme in the Secure DRAM style, run on an emulator at its reported "0.1%" overhead. Both are given the same hardware, the same CL-Drive replay, and the same tuning budget as our tiered scheme. Beating them means matching or improving their residual leakage while keeping window-to-adaptation latency inside the loop's deadline and producing a strictly lower worst-case latency for the protected tiers; against the live-flag variant we additionally predict and will demonstrate that it leaves live streaming buffers exposed, which is the specific gap our design closes. The final comparison is an ablation of our own system with the load-aware gating replaced by a fixed, Delphi-independent tier assignment — same mechanisms, same compute — which isolates whether adaptivity, rather than the underlying protection, earns the latency savings. On the attacker side, the inference baseline is an attacker who recovers the classifier but no raw buffers, scored against the majority-class and chance reference points; only if the buffer-equipped attacker substantially exceeds it does the new threat exist, and only if our defense pushes it back toward it does the defense work.
Why not just…?
Why not just encrypt all memory, always? This is the most serious alternative, and we treat it as the headline defense baseline rather than arguing around it. The evidence against blanket encryption as a free lunch is the security tax itself: counter-mode encryption on persistent memory imposes round-trip latency and write amplification that accelerates cell wear (SecPM), and even TrustZone-backed designs concede unencrypted RAM exposure (T-Edge). Our experiment E2 settles it empirically: if always-on encryption fits the loop's deadline with margin to spare, the adaptive tiering is unnecessary complexity and we will say so; the proposal is falsifiable on exactly this point.
Why not just reuse Secure DRAM's live flags? Because the mechanism targets the wrong lifetime. A live flag physically gates sense amplifiers only for rows whose flag has been cleared — i.e., deallocated data — whereas the cold-boot attacker seizes a running device whose physiological buffers are, by definition, live. Our live-flag baseline in E2 is designed to expose this directly: we expect it to show near-zero overhead and near-zero protection for the artifacts that matter, which is precisely the result that motivates a different design rather than a retuning.
Why not just skip the Delphi and pick thresholds ad hoc, or with a one-shot survey? For the attack and defense engineering, an ad hoc threshold would indeed suffice to get numbers on the board, and we will report results under several fixed thresholds as the ablation. But the deliverable that practitioners need — which leaked artifact is unacceptable at what latency cost — is a contested value judgment across disciplines with no ground truth, which is the exact condition under which Delphi with Content Validity scoring has been validated in medicine and privacy metrics (A Consensus Privacy Metrics Framework for Synthetic Data). A one-shot survey cannot do this job: it lacks iteration and outlier justification, the two features that prevent status bias and groupthink from dominating the aggregate. The physiologically instrumented panel is the one part of the program that is genuinely exploratory; if fatigue markers do not predict false convergence, we will report that as a negative result, which would itself be the first evidence on a question the deliberation literature currently only speculates about.
Decisive Checks
The program stands or falls on three numbers. First, post-extraction victim-state inference accuracy from dumped buffers must exceed both chance and the classifier-only attacker baseline, and re-identification must exceed chance across the 21 CL-Drive subjects — otherwise the "physiological cold boot" is no worse than ordinary model theft and the threat framing collapses. Second, the tiered defense must hold residual leakage at or below the always-on encryption baseline while meeting the loop's per-window adaptation deadline where blanket software zeroing, with its documented 5.1–12.8% slowdowns, does not; and the live-flag ablation must confirm that live buffers leak through it. Third, the Delphi panel must reach statistical stability on the severity taxonomy, with Content Validity scores that discriminate artifact classes rather than flattening them — if every item "passes," the instrument is not measuring anything. A successful outcome is a quantified, temperature-dependent leakage map for neuroadaptive edge memory, a defense that pays for itself in latency, and the first expert-consensus policy for a hardware class that is already shipping with our bodies inside it.
References
- DeepFreeze: Cold Boot Attacks and High Fidelity Model Recovery on Commercial EdgeML Device
- T-Edge: Trusted Heterogeneous Edge Computing
- Embracing Trustworthy Brain-Agent Collaboration as Paradigm Extension for Intelligent Assistive Technologies
- CogErgLLM: Exploring Large Language Model Systems Design Perspective Using Cognitive Ergonomics
- A Consensus Privacy Metrics Framework for Synthetic Data
- Feasibility, comprehensibility and acceptability of the VISION-Cog, a novel tool to assess cognitive impairment in visually impaired older adults: a cross-sectional pilot study in Singapore
- Multimodal Brain-Computer Interface for In-Vehicle Driver Cognitive Load Measurement: Dataset and Baselines
- EEG-based Cognitive Load Classification using Feature Masked Autoencoding and Emotion Transfer Learning
- DRAM architecture for efficient data lifetime management
- Embedding security into ferroelectric FET array via in situ memory operation
- SecPM: a Secure and Persistent Memory System for Non-volatile Memory
- Follow-up after major traumatic injury: a survey of services in Australian and New Zealand public hospitals
- egoPPG: Heart Rate Estimation from Eye-Tracking Cameras in Egocentric Systems to Benefit Downstream Vision Tasks
- AI Agents and the Future of Deliberation: Designing Human-AI Collaboration for Democratic Dialogue