Covert communication theory promises that a transmitter can hide a positive-rate signal from an adversarial observer (the "warden") by exploiting what the warden does not know: the exact level of background noise, or the power of interference that fluctuates from moment to moment. Under that uncertainty, the warden cannot tell a transmission from a noise drift, and the usual square-root limit on covert bits collapses. But this promise rests on a specific picture of the warden: a single antenna measuring total received energy against a fixed threshold. Meanwhile, experimental physics spent the last several decades in exactly the opposite situation — drowning in data, with statistical errors gone and persistent systematic uncertainties left as the binding constraint — and built a toolbox specifically for cancelling nuisance fluctuations: cross-baseline spectral ratios that cancel common-mode drift, orthogonal detection channels with uncorrelated systematics, and adversarially trained classifiers whose outputs are forced to be independent of nuisance parameters.
Those two communities have never met on this question. This proposal builds the meeting: a warden that uses physics-style systematic-cancellation to see through artificial-noise and cover-traffic tricks, a re-derivation of covert limits under such a warden, and — the closing move — a transmitter designed adversarially against it.
Core Claim
The claim that fluctuating artificial noise or cover traffic yields a non-vanishing covert rate is, in large part, an artifact of assuming the warden is a radiometer. Two imported mechanisms should reverse it. First, a warden observing two baselines or two sub-bands can form a power ratio, which cancels any fluctuation that is common-mode across its sensors — precisely how reactor-neutrino analyses at DANSS and NEOS removed unknown reactor flux and detector efficiency to expose a 3.7$\sigma$ oscillation preference model-independently (Model-independent $\bar{\nu}_e$ short-baseline oscillations from reactor spectral ratios). Second, a warden trained with the adversarial pivoting of Learning to Pivot with Adversarial Networks learns a detection statistic whose distribution is invariant to continuous nuisance parameters — noise variance, jammer power, cover-traffic level — so its false-alarm calibration does not drift, which is exactly the failure mode covert designs prey upon. Under either hardened warden, existing transmitter designs should fall back to the square-root law ($\mathcal{O}(\sqrt{n})$ bits over $n$ channel uses), and restoring a positive rate requires transmitter-side adversarial waveform synthesis, which no work has formulated.
The closest existing work is Deep Learning for Spectrum Sensing, whose DetectNet normalizes energy away to detect signals roughly 5 dB below the energy detector's breakdown; but it hard-codes one invariance and, as the file itself notes, never faces a transmitter that adapts to evade it. The pivoting papers (Learning to Pivot with Adversarial Networks, Decorrelated jet substructure tagging using adversarial neural networks) supply the training technique but have only ever been applied to LHC jet data, never to radio detection. The covert designs we attack — dual-functional artificial noise (Dual-Functional Artificial Noise (DFAN) Aided Robust Covert Communications in Integrated Sensing and Communications) and cover traffic with robust beamforming (Robust Beamforming Design for Covert Communications) — both model the warden as a scalar energy detector, and both leave open exactly the multi-antenna, ratio-forming, or structurally-aware warden we build.
Research Program
The program has three components, each answering one question.
First, a theory component on the ratio warden. We analyze a warden with two receive points (antennas or sub-bands) that tests the ratio of received energies rather than their sum, against a transmitter whose artificial noise or cover traffic fluctuates in power. The derivation mirrors the spectral-ratio logic of Model-independent $\bar{\nu}_e$ short-baseline oscillations from reactor spectral ratios: any component of the interference fluctuation that is common-mode across the warden's two observation points cancels in the ratio, so the warden's effective nuisance uncertainty shrinks to the differential component. We will state the condition — a bound on the cross-sensor correlation of the jammer fluctuation — under which the warden's detection error decays as if the noise variance were known, restoring the square-root law that Computationally Efficient Deniable Communication and Undetectable Radios: Covert Communication under Spectral Mask Constraints establish for the known-noise case. This also tells Alice what to do: make the artificial-noise fluctuation statistically independent across the warden's likely baselines, at a cost in jamming power that we quantify.
Second, a learned pivotal warden. We train a detector $f_\phi$ on raw normalized I/Q samples with the adversarial pivoting objective, in which an auxiliary network $r_\psi$ tries to recover the nuisance vector $Z$ (receiver noise variance, jammer power, cover-traffic level) from the detector's output, and the detector is rewarded for making that recovery fail:
$$\min_{\phi}\; \mathcal{L}_{\mathrm{det}}(f_\phi) \;-\; \lambda\, \mathcal{L}_{\mathrm{rec}}(r_\psi \!\circ\! f_\phi), \qquad \max_{\psi}\; \mathcal{L}_{\mathrm{rec}}(r_\psi \!\circ\! f_\phi)$$
Here $\mathcal{L}_{\mathrm{det}}$ is the ordinary binary cross-entropy for distinguishing the no-transmission hypothesis from the transmission hypothesis, $\mathcal{L}_{\mathrm{rec}}$ is the adversary's loss in regressing $Z$, and $\lambda$ trades detection power against nuisance invariance, following Learning to Pivot with Adversarial Networks. Unlike DetectNet's hand-written energy normalization, the invariance here is learned and covers a multi-dimensional, drifting nuisance. We know from the same paper that a fully pivotal and fully discriminative solution need not exist; the interesting empirical question is where that trade-off lands for radio signals, and we will reuse its $\lambda$-sweep methodology.
Third, an adversarial transmitter. Once the hardened warden exists, we train Alice's waveform generator — a neural network emitting baseband samples conditioned on the message bits and the observed cover-traffic signal — to make the pivotal warden's output distribution identical under the two hypotheses, while a fixed radio-frontend check enforces the spectral-mask and pulse-shape constraints of Undetectable Radios. Training alternates between strengthening the warden and adapting Alice, so neither side is fit to a frozen opponent.
Proposed Experiments
All learned components are trained and evaluated on RadioML2016.10a, the only wireless benchmark in the file, which spans SNRs from $-20$ to $+18$ dB with realistic multipath, clock, and frequency offsets, and includes pure-noise samples that serve directly as the no-transmission hypothesis. On top of it we build a simulated covert layer reproducing the DFAN fluctuating-power setting of and the cover-traffic setting of, with nuisance drift injected as multiplicative gain wander.
The decisive measurement is the covert-rate scaling exponent: total bits delivered to the receiver Bob, with the warden's detection error probability held at or above $1-\epsilon$, as a function of the number of channel uses $n$, read off a log-log fit. Slope near $1/2$ means the square-root law holds; slope near $1$ means it is broken. The first experiment runs the ratio warden against the DFAN transmitter and asks whether the slope that the radiometer assumption predicts (near 1) drops to near $1/2$ when the artificial-noise fluctuation is common-mode across the warden's two sensors, and how much differential fluctuation Alice must introduce to climb back. The second experiment compares four wardens — radiometer, DetectNet retrained on our scenario, an un-pivoted detector, and the pivotal warden — under nuisance drift, scoring them by the warden's area under the ROC curve and by false-alarm stability across drift bins; as a sanity check that our pivoting implementation is faithful, we first reproduce the toy Gaussian result of Learning to Pivot with Adversarial Networks, which reports an Approximate Median Significance of about 7.8 versus about 7.0 for an unpivoted classifier at penalty $\lambda = 10$, and the boosted-resonance result of Decorrelated jet substructure tagging, which reports 3.67$\sigma$ versus 2.82$\sigma$ discovery significance under a 50% background systematic. The third experiment trains the adversarial transmitter against the strongest warden from experiment two and reports the scaling exponent and Bob's bit error rate; if the pivot literature's non-existence caveat bites here, it will show up as an inability to push the warden's error toward chance without destroying Bob's rate, and we will report that boundary honestly. For the stochastic-resonance pre-processor of Improvement of Noise Uncertainty and Signal-To-Noise Ratio Wall in Spectrum Sensing Based on Optimal Stochastic Resonance, which reports 9–10 dB output SNR gains at input SNRs of $-30$ and $-25$ dB but only on synthetic channels, no reported result exists on Lacuna for it on RadioML-like faded data, so we treat it as a candidate pre-processing stage for the warden rather than a baseline with a comparable number.
Baselines
The naive reference point is the radiometer with noise uncertainty — a single-antenna energy detector thresholding total received power, which is the implicit warden in the covert designs of Dual-Functional Artificial Noise (DFAN) Aided Robust Covert Communications and Robust Beamforming Design for Covert Communications. It requires no training, and beating it means nothing more exotic than keeping the warden's detection error near chance under conditions where the DFAN and cover-traffic transmitters claim positive rate; against it we expect to fail, which is the point.
The strongest existing competitor is DetectNet from Deep Learning for Spectrum Sensing: a convolutional LSTM network trained on energy-normalized raw I/Q samples from RadioML2016.10a, which maintains reliable detection down to approximately $-8.5$ dB while holding a false alarm rate of 5.92%, roughly a 5 dB improvement over the energy detector that fails around $-6$ dB. It is the strongest competitor because it is the only published detector that already defeats noise uncertainty on our benchmark dataset. To make the comparison fair we retrain it from scratch on our covert-layer data with the same architecture, the same sample budget, and the same tuning search over hyperparameters granted to our pivotal warden, and evaluate both on identical held-out nuisance-drift trajectories. Beating it means one of two measurable things: maintaining a false-alarm rate that stays within a stated band as the nuisance drifts outside the training range while DetectNet's drifts out, or retaining non-chance detection against the adaptive transmitter of experiment three after DetectNet's accuracy collapses. If neither holds, DetectNet wins and the pivotal machinery is not needed for the warden side — we would then state that and concentrate the contribution on the ratio warden and the adversarial transmitter.
On the transmitter side, the baseline is the robust semidefinite-relaxation beamformer with cover traffic from Robust Beamforming Design for Covert Communications, which reports that non-robust designs violate covertness nearly 50% of the time under channel error while the robust design holds the constraint, and that a zero-forcing variant costs less than 1/10th the computation time. We reimplement it in our simulator at equal transmit power and cover-traffic budget. Beating it means delivering a positive covert rate under the hardened wardens where it delivers essentially none. The key ablation isolates the proposal's central ingredient: the same warden network trained with the pivot penalty $\lambda = 0$ versus $\lambda > 0$, and the ratio warden versus a single-sensor energy detector given the same two sensors' data summed rather than divided. If $\lambda = 0$ matches $\lambda > 0$ under drift, the pivoting claim is empty; if summing matches dividing, the ratio claim is empty.
Why not just…?
Why not just use DetectNet as the warden and skip the pivoting? This is the most serious alternative, and we treat it as a live hypothesis rather than a strawman: DetectNet already beats the energy detector by about 5 dB on the exact dataset we use, at a 5.92% false alarm rate (Deep Learning for Spectrum Sensing). Its weakness is that its invariance is one hand-coded normalization fitted to a transmitter that is not trying to hide, and the file explicitly notes it does not model an active adversary. The head-to-head in experiment two — same data, same budget, nuisance trajectories that drift outside training — and the adaptive-transmitter test in experiment three are designed to settle this directly, and if DetectNet holds up under both, we will report that the learned invariance is unnecessary and the contribution narrows accordingly.
Why not just collect more samples? Because the obstacle is systematic, not statistical. The precision-physics literature in the file shows repeatedly that once statistical error shrinks as $1/\sqrt{n}$, residual systematics form a floor that no additional data removes — the calibration-limited mass measurement of Measurement of the $\eta'$-Meson Mass using $J/\psi \rightarrow \gamma \eta'$ and the systematic-limited MAGIC Crab analysis (Performance of the MAGIC stereo system obtained with Crab Nebula data) are the pattern. On the warden's side the mirror image holds: with noise variance unknown, longer observation forces Alice's power down as $1/\sqrt{n}$ and the covert rate vanishes (Computationally Efficient Deniable Communication). More samples solve neither player's systematic problem; cancellation or invariance does.
Why not just set a conservative energy threshold (CFAR-style) with margin for the uncertainty? The SNR wall makes this strictly dominated: under noise uncertainty there is a power level below which no threshold, however conservative, and no sample size suffice, which is why the energy detector fails around $-6$ dB while DetectNet reaches $-8.5$ dB (Deep Learning for Spectrum Sensing), and why the stochastic-resonance work exists at all (Improvement of Noise Uncertainty and SNR Wall in Spectrum Sensing). A conservative threshold also cannot address the ratio-cancellation question, which is about the structure of the statistic, not the threshold placement.
Why not just use worst-case robust optimization on the transmitter (S-lemma and semidefinite relaxation)? That machinery is genuinely strong for bounded, static channel error — it is what keeps covertness intact where non-robust designs fail about half the time (Robust Beamforming Design for Covert Communications) — but it optimizes against a fixed norm-bounded set and has no mechanism for continuous, non-stationary nuisance drift or for decorrelating a learned statistic from a systematic shape. The jet-tagging result is the cleanest evidence of that gap in the file: under a 50% background systematic, a conventional network degrades to 2.82$\sigma$ while adversarial decorrelation reaches 3.67$\sigma$ (Decorrelated jet substructure tagging). Robust optimization is also an Alice-side tool, complementary to our warden-side contribution; our experiments keep it as the transmitter baseline rather than discarding it.
Decisive Checks
Three results would decide the proposal. The scaling exponent under the ratio warden against DFAN transmission dropping from near 1 to near $1/2$ when the artificial-noise fluctuation is common-mode, matching the derived condition, establishes the theory claim. The pivotal warden holding a stable false-alarm band under out-of-range nuisance drift where the equally-budgeted DetectNet and the $\lambda = 0$ ablation do not establishes the learned-invariance claim. And the adversarial transmitter restoring a positive scaling exponent against the best hardened warden, at a Bob error rate a communication engineer would accept, establishes that the arms race — not the radiometer assumption — is the right model of covert communication.
References
- 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
- Learning to Pivot with Adversarial Networks
- Deep Learning for Spectrum Sensing
- Decorrelated jet substructure tagging using adversarial neural networks
- Dual-Functional Artificial Noise (DFAN) Aided Robust Covert Communications in Integrated Sensing and Communications
- Robust Beamforming Design for Covert Communications
- Computationally efficient deniable communication
- Undetectable Radios: Covert Communication under Spectral Mask Constraints
- Improvement of Noise Uncertainty and Signal-To-Noise Ratio Wall in Spectrum Sensing Based on Optimal Stochastic Resonance
- Measurement of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>η</mml:mi><mml:mo>′</mml:mo></mml:msup></mml:math>-Meson Mass using<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>J</mml:mi><mml:mo>/</mml:mo><mml:mi>ψ</mml:mi><mml:mo>→</mml:mo><mml:mi>γ</mml:mi><mml:msup><mml:mi>η</mml:mi><mml:mo>′</mml:mo></mml:msup></mml:math>
- Performance of the MAGIC stereo system obtained with Crab Nebula data