A fully implanted brain–computer interface that drives a robot arm has to squeeze its data through a radio link without heating brain tissue by more than one degree Celsius. One appealing way to save power is to transmit only the electrodes that matter at each moment. That is plausible because neighboring electrodes record largely the same signal, and decoding accuracy often peaks with a small subset. In a reach-and-grasp task, though, which electrodes matter, and how many, should change with the phase of the movement. The fast gross reach is one phase. The slow terminal correction is another, and it happens just when the robot's own gripper blocks its camera's view of the target. This proposal starts from a simple observation. If the implant adapts how much it sends to the movement phase, the size and timing of its packets reveal what the user is doing, even when the contents are perfectly encrypted. An eavesdropper does not need to decrypt anything to learn "the user is now closing on an object."
The proposal is to keep the transmitted amount constant and adapt only which channels fill it. The spatial redundancy that makes channel pruning work also implies that many differently placed, equally sized subsets carry similar information. A phase-dependent choice of a fixed number $K$ of channels could therefore recover most of the decoding benefit and power savings of adaptive masking, while the traffic looks identical across phases. We test this on recorded macaque motor-cortex data that drives a simulated manipulator. We measure decoding quality, estimated implant power, and how well an eavesdropper can infer movement phase and target from encrypted traffic alone.
Core Claim
Phase-adaptive channel selection at a fixed channel count $K$ can match the decoding accuracy of phase-adaptive selection with variable counts. It also clearly beats a single static subset of the same size during the terminal, visually occluded phase. Because the packet stream has constant length and rate, an eavesdropper's traffic-analysis classifier is pushed to chance, whereas variable-count masking leaks phase. Two claims could fail, and the experiments are built to expose either failure. The first is that phase adaptation helps at all over a static subset. The second is that the fixed count costs little relative to variable counts.
Stated compactly, for each movement phase $\phi$, which takes a value such as reach or terminal correction, we choose an electrode subset $S_\phi$ of the $N$ recorded channels:
$$ S_\phi = \arg\max_{S \subset \{1,\dots,N\},\ |S| = K} \; R^2_\phi(S), $$
Here $R^2_\phi(S)$ is the held-out velocity-decoding $R^2$ on phase-$\phi$ data using only channels in $S$. The constraint $|S| = K$ for all $\phi$ is what makes the packet length phase-independent.
Where This Sits
Three lines of work motivate the gap. Work on spatial redundancy shows that more channels can hurt. Methods Towards Invasive Human Brain Computer Interfaces prunes intracranial electrodes by recursive elimination because irrelevant regions cause overfitting. EEG channel selection based on sequential backward floating search reports a U-shaped accuracy curve, with one dataset going from 63.0% with all channels to 90.9% with a searched subset. EEG-ARNN learns channel graphs end to end and keeps 87.9% accuracy on its TJU dataset with 10 of 64 channels. All three choose one static subset offline, and EEG-ARNN names dynamic switching as unresolved. We reuse their selection machinery but condition the subset on movement phase.
Implant telemetry work supplies the power model and the motivation for masking. The 96-channel Networked Neuroprosthesis module computes spiking band power on the implant. Spiking band power is the average magnitude of the 300–1000 Hz signal per channel. The module draws 33.6 mW for 96 channels at 2 kSps, against 31.2 mW for a single raw channel, retains about 95% of decoding information, and explicitly proposes channel masking as future work. tinyRSNN shows a 64-neuron spiking decoder reaching $R^2 \approx 0.66$ on macaque finger velocity within about 27 KB. Neither considers what the link reveals.
The security literature supplies the threat model, but it constrains our design rather than providing it. The Model Inversion Eavesdropping Attack and Secure Deep-JSCC Against Multiple Eavesdroppers show that intercepted learned features can be inverted, or used to infer private attributes without full reconstruction. Defenses such as keyed permutation (same paper) and diffusion-based receiver denoising protect content in analog semantic links. We deliberately do not build an analog semantic encoder. Implants today send digital, quantized features, and content can be encrypted cheaply there. Our contribution is to show that attribute inference survives content encryption once compression becomes phase-adaptive, and to remove that leak.
Manipulation work defines why the terminal phase is special. ContactGuard and DECO document that the gripper blocks the camera during final approach, and they compensate with latent prediction or touch. In a brain-driven arm without tactile arrays, the user's fine corrective intent is the remaining guidance signal in that window. That is why we evaluate decoding quality separately in the terminal phase and why that phase is where a static subset is most likely to fail.
Research Program
The neural side uses the macaque cortical spike-train reaching data used by tinyRSNN and the 96-channel Utah-array recordings from the Networked Neuroprosthesis work. From those we compute spiking band power per channel. Movement phase is not labeled in these datasets, so we define it from kinematics. The terminal phase is the final window, in which hand speed falls below a threshold near the target. This kinematic definition is a stated proxy for "the period when a robot's camera would be occluded." We select $S_\phi$ per phase with two methods: the floating backward search and an EEG-ARNN-style learned aggregation score. We then train a phase-conditioned decoder. Phase itself must be inferred online from the decoded signal, so switching errors are part of the evaluation.
We compare four transmission schemes, all with content encrypted by a counter-mode stream cipher. Counter mode turns a bit error into a single flipped bit rather than a corrupted block, which addresses the error-propagation concern raised by Cross-Layer Encrypted Semantic Communication. The schemes are all channels, a static subset of size $K$, phase-adaptive subsets of variable size, and phase-adaptive subsets of fixed size $K$. Power is estimated by scaling the 33.6 mW per-96-channel figure by channel count. This linear scaling is an assumption, since some front-end power is fixed, and we report sensitivity to a fixed-overhead term.
The eavesdropper is a classifier that sees only packet lengths and timing. It is trained to predict movement phase and reach target. A content-level attacker on unencrypted streams serves as a reference for how much intent the features carry.
For the robot side, decoded velocities drive the end effector in shared control on RLBench reaching and grasping tasks. During the final approach, the visual policy's input is masked to mimic occlusion, so the neural command carries the terminal correction. This tests whether decoding differences in the terminal phase translate into task success.
Baselines
The naive reference is transmitting spiking band power from all 96 channels with content encryption. It sets the power ceiling and the "no selection" decoding level. If spatial redundancy is as strong as the cited work suggests, it should not be the best decoder.
The strongest competitor for decoding is a single static subset of size $K$, chosen by the floating search or EEG-ARNN aggregation on all phases pooled. This is what the channel-selection literature would deploy. It is constant-rate by construction, so it leaks nothing through traffic. It is the alternative most likely to make our proposal unnecessary. For the decoder, the reference number is tinyRSNN's $R^2 \approx 0.66$, with bigRSNN at 0.698, on the macaque finger-velocity benchmark. We use the same spiking decoder family, training data and hyperparameter search budget across all schemes. The only thing that changes is which channels are visible. Beating this baseline means a higher terminal-phase $R^2$ for phase-switched fixed-$K$ subsets at equal $K$, at matched or lower overall power.
The strongest competitor for privacy is phase-adaptive variable-count masking, which is the natural realization of the masking the Networked Neuroprosthesis paper proposes. It should give the best power–accuracy trade-off and the worst leakage. Our scheme beats it if leakage drops to near chance while the $R^2$ loss stays within a margin we fix in advance, such as 0.02. For traffic-analysis inference on neural telemetry, no reported result was found on Lacuna, so the attacker's accuracy on this baseline is itself a first measurement.
The key ablation keeps everything fixed but chooses the per-phase subsets at random among size-$K$ sets. This separates "which channels" from "having any phase switching at all."
Why not just…?
Why not just use one good static subset? This is the real threat. Pooled selection already exploits redundancy, and the U-shaped curve may be largely phase-independent. If static and phase-switched fixed-$K$ subsets tie on terminal-phase $R^2$ and on RLBench success, the adaptive component should be dropped. The finding would then become that static constant-rate pruning is sufficient and leak-free, which is still a useful negative result for the masking proposed by the Networked Neuroprosthesis work.
Why not just pad variable-rate packets to the maximum length? Padding removes the leak but sends the maximum-size packet at every moment, so the radio pays the worst-case power that masking was meant to avoid. The power estimates make this cost explicit. The comparison is legitimate only if padding is cheaper than we assume, for example if front-end acquisition rather than transmission dominates. Our fixed-overhead sensitivity analysis reports where that crossover lies.
Why not just encrypt and stop worrying? Content encryption is assumed in every scheme here, and we expect it to defeat content-inversion attacks of the kind the model inversion paper describes. It does nothing about packet length and timing. The traffic-only attacker on the variable-count baseline directly measures what encryption leaves exposed.
Decisive Checks
Three results decide the proposal. First, the terminal-phase $R^2$ gap between phase-switched and static fixed-$K$ subsets, across $K$ from 8 to 48 on both macaque datasets, tells us whether adaptation earns its place. Second, the traffic attacker's phase and target accuracy under variable-count versus fixed-$K$ schemes, reported against chance with confidence intervals across sessions, tells us whether the leak is real and whether it closes. Third, RLBench shared-control success with occluded final approach, compared across schemes at matched estimated power, tells us whether decoding differences matter physically. As a secondary generality check, we repeat the static-versus-phase comparison on BCICIV 2a and PhysioNet motor-imagery EEG. There, the trial structure defines phases and EEG-ARNN's reported 92.2% and 82.0% anchor the static-subset level. This checks whether the conclusion holds beyond intracortical recordings.
References
- Methods Towards Invasive Human Brain Computer Interfaces
- EEG channel selection based on sequential backward floating search for motor imagery classification
- Graph Convolution Neural Network Based End-to-End Channel Selection and Classification for Motor Imagery Brain–Computer Interfaces
- Design and testing of a 96-channel neural interface module for the Networked Neuroprosthesis system
- Decoding Finger Velocity from Cortical Spike Trains with Recurrent Spiking Neural Networks
- The Model Inversion Eavesdropping Attack in Semantic Communication Systems
- Secure Deep-JSCC Against Multiple Eavesdroppers
- Diffusion-Enabled Secure Semantic Communication Against Eavesdropping
- ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
- DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter
- Cross-Layer Encrypted Semantic Communication Framework for Panoramic Video Transmission