Some next-generation wireless systems propose to run part of a neural network inside the radio channel itself. A stacked intelligent metasurface (a stack of thin programmable panels whose tiny elements each shift the phase of a passing radio wave) can act as a trained matrix multiplication at the speed of light. Receivers built on these metasurfaces then read out signal power with cheap nonlinear parts such as diodes. Language models are an obvious workload for such over-the-air edge inference, and a fairness intervention such as removing gender from an embedding is an obvious thing to ask of them. This note argues that the standard fairness tool does not survive the trip. A digital eraser applied before transmission removes gender in the sense of equal group averages. A linear channel preserves equal averages, but a power detector does not: it squares the field, so differences in group variance reappear as differences in group mean at the receiver.
The proposal is to derive and enforce a "detector-aware" guardedness condition. This is a constraint on the end-to-end transfer of the wave hardware under which the power readings, not the transmitted field, carry no linearly recoverable gender signal. We would train the metasurface under that constraint on text classification benchmarks with known gender and dialect bias, including 1- and 2-bit phase quantization, noise, and channel-estimation error. The contribution is a precise answer to a question neither field has posed: where in a physical inference pipeline fairness must be enforced, and what enforcing it there costs in accuracy.
Core Claim
Linear concept erasure, as formalized in LEACE: Perfect linear concept erasure in closed form, guarantees that no linear probe can recover a protected attribute exactly when the class-conditional means are equal. Write $x$ for a text embedding, $g \in \{0,1\}$ for the protected group, $\mu_g$ for the group mean of $x$ and $S_g = \mathbb{E}[x x^H \mid g]$ for its group second-moment matrix. A metasurface plus propagation channel acts as a complex matrix $H$ whose rows $h_i$ feed receive antenna $i$. Two cases follow.
With a coherent receiver, the received field $y = Hx + n$ has group means $H\mu_g$. Equal means before transmission therefore stay equal regardless of quantization or noise. This is an important and somewhat deflating observation: for linear readout, digital pre-erasure already suffices.
A square-law detector instead outputs $z_i = |h_i^H x + n_i|^2$. Its group mean is
$$ \mathbb{E}[z_i \mid g] = h_i^H S_g h_i + \sigma^2, $$
where $\sigma^2$ is the noise power, which is the same for both groups. The detector outputs are linearly guarded if and only if
$$ h_i^H D\, h_i = 0 \quad \text{for every antenna } i, \qquad D = S_1 - S_0 . $$
Equal means ($\mu_1 = \mu_0$) do not imply $D = 0$, so pre-transmission LEACE leaks through the detector whenever the groups differ in spread along task-relevant directions.
The condition is weaker than equalizing the full covariance, which no label-free affine map can do in general. It asks only that each row of $H$ lie on the "null cone" of the indefinite matrix $D$. That leaves most of the design freedom for the task. We claim this per-row quadratic constraint is the right fairness target for power-detecting wave receivers. We further claim that imposing it during metasurface training yields lower gender leakage at a given accuracy than any transmitter-side linear eraser followed by an unconstrained metasurface.
Positioning Against Prior Work
On the wave side, Stacked Intelligent Metasurfaces for Task-Oriented Semantic Communications trains metasurface phase shifts with Adam against cross-entropy so that diffracted energy lands on class-specific antennas. It reports over 90% MNIST accuracy with picosecond latency. Metasurfaces-Integrated Wireless Neural Networks for Lightweight Over-The-Air Edge Inference, the MINN framework, treats the programmable channel as a differentiable linear layer between a digital encoder and a decoder. It gets its nonlinearity from saturating amplifiers or diodes and reaches roughly 95% on MNIST at 10 dB SNR with a 4-layer surface. These two papers supply our method: we reuse their end-to-end differentiable simulator and training procedure. Neither touches text or demographic attributes.
Stacked Intelligent Metasurfaces for Multi-Modal Semantic Communications sends text as an ordinary digitally modulated bitstream and never computes on it in the wave domain. We instead feed dense sentence embeddings into the first metasurface layer. Stacked Intelligent Metasurface Assisted Multiuser Communications: From a Rate Fairness Perspective uses "fairness" to mean balancing data rates across radio users. That is mathematically unrelated to group fairness over people described in text, and we note it only to avoid confusion.
On the fairness side, three erasure methods motivate the proposal and supply its baselines. Linear Adversarial Concept Erasure finds a rank-1 projection through a minimax game. LEACE gives the closed-form, minimal-distortion affine eraser. Preserving Task-Relevant Information Under Linear Concept Removal, called SPLINCE, additionally preserves the covariance between embeddings and task labels. All three guarantee only first-order (mean) guardedness, and the first two papers themselves report that nonlinear probes can recover erased attributes. That limitation is exactly what bites here: the nonlinearity that recovers the attribute is not an adversary's MLP but a diode built into the receiver. Bias in Bios constrains the design by showing that removing words does not remove gender. Its gender classifier exceeds 65% accuracy on scrubbed text, so the intervention must operate on representations.
Research Program
The pipeline has a device side and a receiver side. The device encodes a biography or tweet with a frozen sentence encoder, reduces it to $d$ dimensions with a fixed linear map, and loads it onto the transmission coefficients of the first metasurface layer. Here $d$ ranges from 64 to 441 to match the meta-atom counts studied in the wave papers. Diffraction through $L$ programmable layers, following the Rayleigh–Sommerfeld model used in those papers, then free-space propagation produce the field at $M$ receive antennas. Each antenna applies square-law detection, and a small digital decoder predicts the occupation or sentiment label.
Training minimizes task cross-entropy plus a penalty $\lambda \sum_i (h_i^H \hat D h_i)^2$. The matrix $\hat D$ is estimated once from labeled training embeddings, so no group labels are needed at inference. The penalty is applied under straight-through estimation for 1- and 2-bit phases.
The analysis has three parts. First, we characterize feasibility: for a given $D$, what fraction of the row space satisfies the constraint, and how quantization inflates the residual $|h_i^H D h_i|$. This gives a leakage bound in terms of phase resolution and meta-atom count. Second, we prove and test the coherent-versus-square-law dichotomy stated above. Third, we measure how channel-estimation error perturbs $H$ and therefore the residual, which determines how often the surface must be re-trained as the channel changes.
Baselines
The naive reference point is an unconstrained MINN-style metasurface trained only for task accuracy on raw embeddings. It fixes the accuracy ceiling and the unmitigated TPR gap, the difference between groups in the rate at which each group's members are correctly classified. The same architecture and training budget as every other variant is used.
The strongest existing method is digital LEACE applied at the transmitter, followed by the same unconstrained surface. On a purely digital Bias in Bios pipeline, LEACE reduced TPR-GAP from 0.198 to 0.084 with about a 2% accuracy loss (LEACE). We will first reproduce that number digitally on our embeddings to calibrate, then pass the erased embeddings through the simulated surface and detector. SPLINCE before transmission is the second strong competitor, because it should lose less task accuracy under the strong gender–occupation correlation in Bias in Bios. For SPLINCE on Bias in Bios the file gives only a qualitative claim of outperforming baselines; no reported numeric result was found on Lacuna, so we will measure it ourselves. R-LACE rank-1 projection serves as a cheaper linear reference.
All methods share encoder, dimensionality, surface geometry, SNR, quantization level, and a fixed hyperparameter search of equal size on a held-out validation split. For every method we select the operating point with the best accuracy at matched gender-probe accuracy.
Two ablations isolate the key ingredient. One trains the surface with a first-order constraint only, equal means of the field. The other applies our quadratic constraint through a digital transmitter-side linear map instead of the surface, which separates "detector-aware target" from "enforced in hardware."
Beating the baselines means two things at matched task accuracy. The linear probe on detector outputs must sit closer to chance than for transmitter LEACE or SPLINCE, and the TPR-GAP must be lower. Both must hold across three seeds and at 2-bit quantization.
Why not just…?
Why not just run LEACE or SPLINCE digitally before transmission? For coherent receivers this is the right answer, and our analysis says so: equal means pass unchanged through any linear channel. The threat is real and the experiment is designed to expose it. If detector-output probe accuracy after transmitter LEACE is already at chance on Bias in Bios and Moji, the variance differences are negligible in practice and the constrained surface is unnecessary. We would then report that as the main finding.
Why not just use a coherent receiver so linear erasure suffices? Coherent receivers need phase-locked RF chains per antenna. Those chains are precisely the power and cost that MINN-style systems (Metasurfaces-Integrated Wireless Neural Networks for Lightweight Over-The-Air Edge Inference) remove by relying on analog nonlinearities. We will include a coherent-receiver arm to quantify the accuracy and leakage one would buy back, but the low-power regime is the one worth studying.
Why not just distill the encoder into something smaller? Can Model Compression Improve NLP Fairness finds that distillation lowers WinoBias bias and toxicity. The same paper reports rising perplexity, inconclusive StereoSet results, and no fairness gain from small models trained without a teacher. More fundamentally, compression acts on the encoder and says nothing about leakage created downstream by the detector. We include a distilled-encoder arm to check whether it changes $D$ enough to matter.
Why not just scrub gendered words from the text? Bias in Bios shows that gender remains predictable above 65% after scrubbing, through proxy words. We run scrubbing as a weak baseline mainly to confirm that result survives our pipeline.
Proposed Experiments
Bias in Bios is the primary benchmark: occupation classification with gender as the protected attribute, reported as accuracy, TPR-GAP, and its correlation with occupational gender imbalance. Moji tests a second attribute, dialect, in sentiment classification. For each method we report linear and MLP probe accuracy on detector outputs, TPR-GAP, and accuracy. We sweep phase resolution (1-bit, 2-bit, continuous), meta-atom count, layer depth up to the $L=7$ plateau reported for the task-oriented surface, SNR, and channel-estimation error.
Three checks settle the proposal. The first tests whether transmitter-side LEACE leaks measurably at the square-law detector while staying guarded under a coherent receiver; that result validates or kills the core premise. The second tests whether the quadratic constraint closes that leak at 2-bit quantization with accuracy loss no worse than LEACE's reported 2%. The third tests whether hardware enforcement beats the digital transmitter-side version of the same constraint; if it does not, the contribution reduces to the detector-aware erasure target itself, which remains useful for any power-detecting receiver. MLP probes will be reported throughout. We make no claim of nonlinear guardedness, and a large gap between linear and MLP leakage would mark the next problem rather than a success.
References
- LEACE: Perfect linear concept erasure in closed form
- Stacked Intelligent Metasurfaces for Task-Oriented Semantic Communications
- Metasurfaces-Integrated Wireless Neural Networks for Lightweight Over-The-Air Edge Inference
- Stacked Intelligent Metasurfaces for Multi-Modal Semantic Communications
- Stacked Intelligent Metasurface Assisted Multiuser Communications: From a Rate Fairness Perspective
- Linear Adversarial Concept Erasure
- Preserving Task-Relevant Information Under Linear Concept Removal
- Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Setting.
- Can Model Compression Improve NLP Fairness