Sideband-Calibrated Uncertainty Sets for Secure Beamforming Against Silent Eavesdroppers

A base station that wants to keep a transmission secret from a passive eavesdropper has a basic problem: the eavesdropper never transmits, so the base station does not know its channel. Recent work in integrated sensing and communication (ISAC), where the same waveform serves as both radar and data signal, solves this by pointing a radar at the eavesdropper. It estimates the eavesdropper's direction and distance from the echo. It then bounds the estimation error with the Cramér-Rao bound, which is the smallest error variance any unbiased estimator can achieve, and designs beams that stay secure for every channel inside that bound. This works well on paper. It rests on one fragile assumption: the only thing in the echo is the eavesdropper plus white noise. Real scenes also contain walls, ground bounces and parked cars. Those reflections produce structured "clutter", and a bound computed from noise alone ignores it. When clutter biases the estimate, the true eavesdropper channel falls outside the assumed uncertainty set. The beam's null then lands in the wrong place, and the security guarantee silently fails.

Particle and neutrino physicists face the same situation in their own language. A rare signal sits on top of a background that simulations predict imperfectly, and the error left after all corrections is what they call a residual systematic. Their standard answer has two parts. They simulate the background shape as a template. They then fix the template's normalization and distortions using "sidebands", which are nearby regions of the measurement known to contain no signal, and they carry the leftover uncertainty into the final fit as extra free parameters. This note proposes applying that recipe to eavesdropper sensing. Clutter would be simulated by ray tracing, calibrated on range–angle cells next to the suspected eavesdropper, and propagated into a confidence region for the eavesdropper's channel. That region would replace the Cramér-Rao ellipse in secure beam design. The test is whether the resulting beams meet their leakage and covertness targets at the promised rate when clutter and simulator error are present, and how much secrecy rate that costs.

Core Claim

The claim is that uncertainty sets for a silent adversary's channel should be built from a clutter-aware profile likelihood fitted over spatial sidebands, not from a noise-only Cramér-Rao bound. Such sets keep their nominal coverage when the scene contains multipath clutter and the clutter simulator is itself biased. They are also much smaller than the inflated worst-case sets an engineer would otherwise need.

Concretely, let $y_c$ be the radar return in range–angle cell $c$. We model it as

$$ y_c = \alpha\, a_c(\theta_E, d_E) + \sum_{k} \mu_k\, T_{k,c}(\nu) + n_c , $$

where $a_c(\theta_E,d_E)$ is the response of an eavesdropper at angle $\theta_E$ and distance $d_E$ with unknown reflection gain $\alpha$, $T_{k,c}(\nu)$ is the ray-traced clutter template for scatterer group $k$, $\mu_k$ is a free normalization, $\nu$ collects shape nuisance parameters (material permittivity, diffuse-scattering fraction, small geometry offsets), and $n_c$ is receiver noise. Sideband cells, meaning cells adjacent to the candidate eavesdropper cell but excluded from it, have $\alpha a_c \approx 0$. They therefore constrain $\mu$ and $\nu$ directly. The uncertainty set handed to the beamformer is the profile-likelihood region

$$ \mathcal{U}_{1-\beta} = \Big\{ (\theta_E, d_E, \alpha) : -2\log \frac{\max_{\mu,\nu} L(\theta_E,d_E,\alpha,\mu,\nu)}{\max_{\text{all}} L} \le \chi^2_{3,1-\beta} \Big\}, $$

where $L$ is the joint likelihood over signal and sideband cells, $1-\beta$ is the target coverage, and the chi-square threshold follows from Wilks' theorem for three parameters of interest. We would check that approximation by simulation rather than trust it. Unlike a Cramér-Rao ellipse, this set grows when clutter is poorly constrained and shifts when clutter biases the estimate.

The new element is not ISAC-based eavesdropper sensing, which already exists, nor template-plus-sideband fitting, which is routine in physics. It is using a sideband-calibrated, simulator-aware likelihood region as the uncertainty set inside robust secure and covert beamforming, and judging that set by its empirical coverage under simulator mismatch.

Grounding in Prior Work

Three ISAC security papers motivate the gap and supply the optimization machinery. Sensing-Enhanced Secure Communication: Joint Time Allocation and Beamforming Design splits a slot into a sensing phase that locates a silent eavesdropper and a joint phase that uses artificial noise both to jam and to track. Joint Sensing and Covert Communications in RIS-NOMA Systems ties the warden's channel error directly to the sensing Cramér-Rao bound. Adaptive Sensing Performance Design for Enhancing Secure Communication in Networked ISAC Systems embeds that bound inside the leakage constraints of multi-base-station designs. All three assume line-of-sight point targets with no structured clutter, and the file flags this as their open limitation. We keep their beamforming solvers and change only the uncertainty set they are handed. MIMO Integrated Sensing and Communication: CRB-Rate Tradeoff constrains the design: unified signaling beats time switching across the Pareto boundary. We therefore take sideband measurements from the same unified waveform instead of adding dedicated calibration slots.

The physics side supplies the method and a warning. NOvA's coherent π⁰ measurement renormalizes GENIE/GEANT4 background templates on a control sample. MINERvA's antineutrino single-pion measurement uses sideband fits and reports that competing generators miss normalization by +8% (GENIE) and −10% (GiBUU). That is direct evidence that even detailed simulators carry residual bias, and it shows that correcting normalization alone leaves shape errors behind. This is why our model has shape nuisances $\nu$ as well as normalizations $\mu$. Belle and BESIII analyses (tau decays, φχc1(3872)) show the sideband-adjacent-to-peak construction we reuse in angle and range. Reactor spectral ratios show that ratios between measurements can cancel common unknowns such as absolute flux. We borrow this in one ablation that normalizes each cell by its sideband neighbours to cancel unknown array gain calibration.

One wireless paper crosses the boundary. Martian dust storm detection with THz opportunistic ISAC propagates physical-model uncertainty into sensing error and finds that an unconstrained input, dust particle size, dominates. That is the failure we aim to prevent: a simulator input left uncalibrated. Their setting is attenuation-only and non-adversarial, whereas ours involves phase-sensitive nulling against an adversary.

Research Program

The work has three stages. The first builds the clutter simulator and its deliberately wrong twin. We use the DeepMIMO ray-tracing scenario O1 at 140 GHz, which is the scenario used by Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback. The file lists no resource identifier for DeepMIMO. We place a base station, a legitimate user and an eavesdropper at sampled positions, including positions behind partial blockage. From this we generate "truth" echoes and communication channels. The file contains only one ray-traced scene. We therefore create simulator mismatch in a controlled way: the template simulator uses perturbed material reflectivities, omits or rescales diffuse scattering, and shifts geometry by sub-wavelength-to-decimetre offsets relative to the truth generator. Mismatch magnitude is a swept experimental variable, and each perturbation is tied to one nuisance parameter in $\nu$. This is weaker than two independent ray tracers, and we say so. It does let us measure coverage as a function of known bias, which a single-simulator study cannot do.

The second stage builds the estimator. For each sensing frame we fit the likelihood above over the candidate cell and a ring of sideband cells, compute the region $\mathcal{U}_{1-\beta}$ by profiling, and map it to a set of eavesdropper channel vectors. The sideband width trades off two things. Wider rings constrain the clutter better, but the clutter is less likely to be the same as in the target cell. We study this trade-off explicitly, because transferring a background estimate from control region to signal region is the classic source of residual systematics.

The third stage plugs the set into existing robust designs: the leakage-constrained beamforming with artificial noise from the two-phase design above, and the covert power constraint from the RIS-NOMA formulation. In both, the robust constraint is imposed over sampled or convex-hulled points of $\mathcal{U}$ instead of the Cramér-Rao ellipse. We evaluate on held-out truth channels.

Baselines

The naive reference is perfect-knowledge nulling applied to the raw point estimate. Beams are designed as if the estimated eavesdropper direction were exact. This shows how badly clutter bias breaks nulling, and our method must beat it by a wide margin in leakage violations.

The strongest existing method is Cramér-Rao-bounded robust design as in Joint Sensing and Covert Communications in RIS-NOMA Systems and Adaptive Sensing Performance Design. That paper reports a covert rate of roughly 98.7% of the perfect-CSI rate under its line-of-sight, clutter-free model. The DFRC paper operates at a leakage limit of 0.5 bits/s/Hz with an optimal sensing duration of about 0.2 ms in a 1 ms slot. We rerun this baseline on exactly our scenes, with the same waveform, sensing time, power budget and solver, and give it the true noise variance. The only difference is the uncertainty set. Beating it means one of two outcomes. Either the violation rate of the 0.5 bits/s/Hz leakage target (or of the covertness target) is brought to the nominal $\beta$ in cluttered scenes where the Cramér-Rao design exceeds it. Or, at matched violation rate, our design achieves higher secrecy or covert rate. In clutter-free scenes our method should approach the Cramér-Rao baseline's rate; that is a sanity check, not a win.

A norm-bounded worst-case baseline uses a spherical error set whose radius is tuned on a validation split, with the same tuning budget as our sideband width and coverage level, until it reaches nominal coverage. This isolates whether shaped sets matter or simply larger ones.

Two ablations isolate the key ingredient. The first, simulation-only, fits templates with $\mu,\nu$ fixed at simulator defaults and uses no sidebands. The second, sidebands-only, estimates clutter in the target cell by interpolating the sideband cells without any template, in the spirit of classical neighbouring-cell radar background estimation. The full method must beat both for the claim to hold. A third ablation adds the sideband-ratio normalization to test robustness to unknown array gain errors.

Why not just…?

Why not just inflate the norm-bounded error set until it is safe? Joint Sensing and Covert Communications in RIS-NOMA Systems reports that norm-bounded models are overly conservative relative to sensing-grounded ones. Adaptive Sensing Performance Design finds that static margins become infeasible at low transmit power. Clutter bias is directional: it pushes the estimate toward specific scatterers. A sphere large enough to cover that shift wastes most of its volume, while a likelihood region shifts in the right direction. The tuned-radius baseline tests this directly. If it reaches our coverage at similar rate, the shaped set adds nothing, and we would report that.

Why not just trust a high-fidelity ray tracer and skip calibration? The physics record argues against it. MINERvA's generators disagree with data by +8% and −10% in normalization, and absolute reactor flux predictions were off by about 5% until ratio methods bypassed them. The Mars THz study found an uncalibrated simulator input to be its dominant error. The simulation-only ablation measures this in our setting as mismatch magnitude grows. We expect its coverage to collapse at moderate mismatch.

Why not just estimate clutter from neighbouring cells, as radar receivers already do? This is the most serious alternative. The file contains no wireless paper evaluating it, so we have no reported number for it. Neighbouring-cell averaging assumes the clutter in the target cell resembles its surroundings. That holds in homogeneous scenes and fails near edges, corners and strong discrete reflectors, where the ray-traced template supplies the missing shape. The sidebands-only ablation, broken down by scene type, settles the question. If it matches the full method everywhere, including heterogeneous positions, the contribution shrinks to the uncertainty-set construction alone. If it matches only in homogeneous scenes, the template's value is confined to cluttered geometry, and we would state that scope.

Decisive Checks

The primary check is a coverage curve. Across thousands of eavesdropper placements, we plot the fraction of trials in which the true eavesdropper channel lies inside the uncertainty set against injected simulator mismatch, for the Cramér-Rao ellipse, the tuned sphere, both ablations and the full method. The method succeeds only if its coverage stays near the nominal $1-\beta$ across the mismatch sweep while the Cramér-Rao ellipse undercovers once clutter is present.

The second check is operational. It measures the empirical probability that the eavesdropper's actual rate exceeds 0.5 bits/s/Hz, and the warden's detection error in covert mode, both against achieved secrecy or covert rate. Results are split into line-of-sight, partially blocked and strongly cluttered positions.

The third check is a stress test of the sideband assumption. We place a second scatterer inside the sideband ring and measure how coverage degrades. This tells us whether profiling the shape nuisances absorbs contaminated control regions or whether an explicit contamination term is needed.

A negative outcome would still be informative. If Cramér-Rao sets keep nominal coverage in cluttered DeepMIMO scenes, the line-of-sight assumption in current ISAC security work would matter less than its authors suggest, and that finding would be worth reporting.

References

  1. Sensing-Enhanced Secure Communication: Joint Time Allocation and Beamforming Design
  2. Joint Sensing and Covert Communications in RIS-NOMA Systems
  3. Adaptive Sensing Performance Design for Enhancing Secure Communication in Networked ISAC Systems
  4. MIMO Integrated Sensing and Communication: CRB-Rate Tradeoff
  5. Measurement of neutrino-induced neutral-current coherent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>π</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math> production in the NOvA near detector
  6. Measurement of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo stretchy="false">¯</mml:mo></mml:mover><mml:mi>μ</mml:mi></mml:msub></mml:math> charged-current single <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math> production on hydrocarbon in the few-GeV region using MINERvA
  7. Search for lepton-number- and baryon-number-violating tau decays at Belle
  8. Search for the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:mi>ϕ</mml:mi><mml:msub><mml:mi>χ</mml:mi><mml:mrow><mml:mi>c</mml:mi><mml:mn>1</mml:mn></mml:mrow></mml:msub><mml:mo stretchy="false">(</mml:mo><mml:mn>3872</mml:mn><mml:mo stretchy="false">)</mml:mo></mml:math> process at BESIII
  9. Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
  10. Martian Dust Storm Detection With THz Opportunistic Integrated Sensing and Communication in the Internet of Space (IoS)
  11. Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO