Calibrating Eavesdropper Uncertainty Without Ever Seeing the Eavesdropper

A transmitter that wants to keep a message secret from a silent listener usually aims a "null", a direction of near-zero radiated energy, at where it believes the listener is. That belief usually comes from radar-style sensing of reflected echoes. The null only protects the message if the belief is right. Current designs size their margin of error with textbook formulas that assume clean line-of-sight geometry, or calibrate it inside a simulator. Neither tells you how often the margin fails in the building, street, or hardware where the system actually runs. The obstacle is basic. Statistical calibration methods need labeled examples of "where the target really was", and an eavesdropper never supplies one.

We propose to get those labels from stand-ins. Cooperative devices in the same environment, such as legitimate users or deployed test nodes with known positions, are sensed by the same echo pipeline. Each one is then treated as a pretend eavesdropper: the transmitter steers a null at its estimated position, and the stand-in reports how much signal actually leaked to it. These reports are real, in-environment measurements of leakage, and they include null-depth limits caused by hardware flaws that no angle-error model captures. A reweighting step corrects for the fact that stand-ins are not distributed like eavesdroppers. It uses only the echo features of the unknown target, which the transmitter does observe. The result is a beamformer whose rate of secrecy failures is certified against a user-chosen target on data from the deployment environment, not from the design simulator.

Core Claim

Secrecy guarantees for silent adversaries should be calibrated directly on measured leakage, not on angle error, and the labels for that calibration can come from cooperative surrogates in the deployment environment. A covariate-shift correction driven by the unknown target's own echo features carries the guarantee from the surrogates to the eavesdropper. We expect this to keep secrecy outage near its target in conditions where simulator-calibrated and formula-based margins under-cover. We also expect it to give legitimate users more throughput than designs that need no knowledge of the eavesdropper at all.

Where This Sits

Three lines of prior work motivate the proposal and supply its tools. On the security side, Sensing-Enhanced Secure Communication uses a dedicated radar phase to estimate a silent eavesdropper's angle and range, then injects artificial noise. Joint Sensing and Covert Communications in RIS-NOMA Systems sizes the angle uncertainty of an uncooperative warden by the Cramér-Rao bound, the theoretical minimum variance of an unbiased estimator. Both assume line-of-sight sensing and are validated only in synthetic Monte Carlo runs. They motivate the proposal: they show sensing-based security works when the error model is right, and they never test what happens when it is not.

On the calibration side, Conformal Robust Beamforming Via Generative Channel Models builds conformal uncertainty sets for a cooperative user's channel and optimizes the worst case inside them. Conformal prediction is a wrapper that turns any predictor into sets guaranteed to contain the truth at a chosen rate, given exchangeable calibration data. The labels in that work come from pilots that a cooperative user transmits, which an eavesdropper will not do. Conformal Risk-Averse Decision Making with Optimized Certainty Equivalent Risk Control supplies our calibration machinery. It calibrates a beam-steering decision rule through Learn-Then-Test, a procedure that picks a tuning parameter with a high-probability bound on risk, and it targets tail risk via conditional value-at-risk. Its "reality" is a Sionna ray-traced Munich scene whose material properties are perturbed by 20%, and it addresses only throughput to a single legitimate user. We reuse its decision-level calibration but change two things: the risk is leakage to an adversary, and the calibration labels come from surrogates in the target environment rather than from a simulator.

Two further papers constrain the design. PyJama shows that adversaries can exploit the timing of estimation slots, staying silent while the receiver estimates their spatial signature, so that nulls built from those estimates go stale. This warns that a passive-eavesdropper guarantee says nothing about adaptive adversaries, and we scope our claims accordingly. Calibrating AI Models for Wireless Communications via Conformal Prediction shows that conformal calibration holds up on real radio traces, but only for scalar received-strength prediction. That is encouraging, but it is not evidence for spatial nulling.

Research Program

The system has three parts. The first is an echo-based estimator. A dual-function radar-communication transmitter, one that sends signals serving both as radar probes and as data, outputs for each detected target a point estimate of its angle together with a feature vector $z$. The features include echo SNR, range, angular distance to Bob and to strong clutter, and residual multipath energy.

The second part is a family of protective beamformers indexed by a single knob $\lambda \ge 0$. For a target with estimate $\hat\theta(z)$ and model-predicted spread $\hat\sigma(z)$, the beamformer maximizes Bob's rate subject to the eavesdropper's SINR staying below a threshold $\Gamma$ for every angle in $[\hat\theta(z)-\lambda\hat\sigma(z),\, \hat\theta(z)+\lambda\hat\sigma(z)]$. It may optionally add artificial noise outside Bob's subspace. Larger $\lambda$ means wider protection and less rate for Bob.

The third part is surrogate calibration. Each of $n$ cooperative surrogates $i$ is sensed, assigned a beamformer $w_\lambda(z_i)$ that treats it as the eavesdropper, and reports a measured leakage outage indicator $\ell_i(\lambda) = \mathbf{1}[\mathrm{SINR}_i(w_\lambda) > \Gamma]$. A probabilistic classifier is trained to tell surrogate echo features apart from the unlabeled echo features of non-cooperative targets. It yields importance weights $\omega(z) \propto p_{\text{target}}(z)/p_{\text{surrogate}}(z)$, which measure how much more likely a feature vector is under eavesdroppers than under surrogates. We then choose

$$\lambda^\star = \min\Big\{\lambda : \mathrm{UCB}_\delta\Big(\tfrac{\sum_i \omega(z_i)\,\ell_i(\lambda)}{\sum_i \omega(z_i)}\Big) \le \alpha\Big\},$$

where $\alpha$ is the target secrecy outage rate and $\mathrm{UCB}_\delta$ is a weighted upper confidence bound that holds with probability $1-\delta$. Because $\ell_i$ is measured leakage, the calibration absorbs every source of error at once: angle error, multipath, and limited null depth from phase noise or mutual coupling between antennas. An angle-only conformal set would miss the null-depth part.

Two questions are analyzed rather than assumed. The first is when weighting is valid. It requires that eavesdroppers differ from surrogates only through observable echo features. We test this deliberately by giving eavesdroppers different radar cross-sections and antenna gains. The second is how many surrogates are needed. The weighted bound loosens as the effective sample size $(\sum\omega)^2/\sum\omega^2$ shrinks, which predicts a measurable trade-off between how unlike the surrogates are to eavesdroppers and how conservative $\lambda^\star$ becomes.

Baselines

The naive reference point is a null designed at the point estimate with no margin, which is what a perfect-CSI design does under imperfect sensing. It shows the size of the problem. A second reference is artificial noise spread isotropically outside Bob's channel, which needs no information about the eavesdropper. This is the most serious practical competitor because it cannot be fooled by bad sensing. It is run with the same power budget and a power split tuned on the same calibration budget, and it must be beaten on Bob's rate at matched measured outage.

The formula-based competitor follows Lei et al.: the protection set is sized by the Cramér-Rao bound under a line-of-sight model, with identical beamformer optimization. For this pipeline the paper reports a covert rate of about 98.7% of the perfect-knowledge rate, but only within its own synthetic model. No reported result was found on Lacuna for its outage under ray-traced or hardware-impaired channels. We also include worst-case norm-bounded robust beamforming in the style of Tuan et al. and Hua et al., with the bound radius tuned in the design simulator.

The strongest existing method is simulator-calibrated decision risk control in the style of Farzaneh and Simeone. It uses the same Learn-Then-Test rule and the same $\lambda$ family, but the $\ell_i$ are computed in the design simulator. Their reported number is that at risk target 0.25 their calibration met the target in 100% of 500 Monte Carlo trials while a value-at-risk baseline failed in 53.4%, with an 8-antenna array on the Munich scene. That figure concerns throughput, not leakage. Using the same trial count and the same calibration-set size makes the comparison fair. Beating this baseline means keeping secrecy outage at or below $\alpha$ in at least 95% of 500 trials (with $\delta=0.05$) on the held-out reality model in cases where it does not, or matching its coverage with higher rate for Bob.

Two ablations isolate the key ingredients. Unweighted surrogate calibration ($\omega\equiv1$) isolates the shift correction. A version that calibrates angle coverage instead of measured leakage isolates the decision-level label.

Why not just…?

Why not just use isotropic artificial noise and skip sensing? It is robust by construction, and if eavesdroppers sit close in angle to Bob it may be near-optimal. Its cost is power spent everywhere, so Bob's rate falls sharply as the antenna count grows relative to the number of users. Our experiment plots Bob's rate at matched measured outage against angular separation between Bob and the eavesdropper. If isotropic noise matches our method across separations, we will report that sensing-based security is not worth its overhead in that regime.

Why not just inflate the Cramér-Rao bound by a constant tuned on surrogates? This is a real threat, because it is essentially our method with $\hat\sigma$ fixed to the formula's value and no weighting. The scalar-inflation baseline is exactly this. Our method wins only if outage failures concentrate in feature regions, such as multipath-heavy or low-echo-SNR targets, that a single constant must over-protect everywhere to cover. We will report conditional outage per feature bin to show whether that structure exists.

Why not just calibrate in a high-fidelity simulator, as prior work did? The design simulator controls only what its designer parameterized. PyJama and the Munich-perturbation study both stay inside simulation. Our protocol uses different simulators for design and for reality and adds hardware impairments only to the reality side, which tests directly whether simulator calibration transfers. If it transfers, surrogates add cost without benefit, and we will say so.

Why not just widen to a fixed protected sector? Fixed sectors in the style of Topcu carry no guarantee and do not adapt to echo quality. They are included as a $\lambda$-swept curve so that their best operating point can be read off.

Proposed Experiments

Design runs use a 3GPP Urban Micro geometric model with an OFDM pipeline, built on PyJama's Sionna extension. "Reality" is the Sionna Munich ray-traced scene with materials perturbed by 20%, following Farzaneh and Simeone, plus injected phase noise, I/Q imbalance, mutual coupling, and low-resolution converters. Arrays range from 8 to 64 antennas. Surrogates are placed at random street positions. Eavesdroppers are drawn from a shifted distribution: more non-line-of-sight placements, different radar cross-section, and closer angular proximity to Bob.

The primary metric is the fraction of 500 trials in which measured secrecy outage stays at or below $\alpha \in \{0.05, 0.1\}$. Secondary metrics are Bob's rate and conditional outage by echo-SNR and angle-separation bins.

The decisive check is whether simulator-calibrated risk control under-covers on reality while surrogate calibration does not. A second check deliberately breaks the weighting assumption by making eavesdroppers differ in unobservable gain. Here we expect, and will report, that the guarantee degrades, and we measure by how much. A third check is the surrogate-count sweep against the effective-sample-size prediction. A small over-the-air replication on a multi-antenna software radio is a stretch goal. No testbed resource appears in the file, so the main conclusions will rest on the two-simulator protocol, and we will state that limitation plainly.