Wireless transmitters often protect their traffic by broadcasting deliberate interference alongside the real signal — so-called artificial noise, shaped so that only the intended receiver can cancel it while every eavesdropper is drowned in junk. The entire design rests on one unexamined assumption: that the noise is meaningless. This proposal argues that the noise is in fact a cryptographic component in disguise, and that it currently fails the same test that has broken hybrid encryption schemes. Specifically, the noise is generated from randomness that is completely independent of any public key or authenticated binding — pre-stored Gaussian sequences picked by a secret seed schedule, or safeguarding signals reproduced from a shared random seed. That is exactly the "public-key-independent ciphertext generation" weakness that lets algorithm-substitution attacks exfiltrate session keys from subverted KEMs Subvert KEM to Break DEM: Practical Algorithm-Substitution Attacks on Public-Key Encryption.
The consequence is a clean inversion of physical-layer security. Null-space beamforming deliberately delivers maximum noise power to eavesdroppers Multiple-Input Multiple-Output Gaussian Broadcast Channels with Confidential Messages, so a transmitter whose noise generator has been subverted possesses a high-power, dedicated channel to precisely the adversary the noise was meant to blind — a channel that no ciphertext auditor inspects, because everyone believes it carries nothing. We propose to formalize this attack, measure it against the benchmarks used by the physical-layer-security community, and close it with a constructive fix: derive the artificial noise deterministically from the session key that the KEM encapsulates under the receiver's public key, so that the last "free" randomness in the system is bound to the key it is meant to protect.
Core Claim
Artificial noise generation, as deployed in OFDMA secrecy transmission Secrecy Wireless Information and Power Transfer in OFDMA Systems), shared-seed safeguarding Generative AI based Secure Wireless Sensing for ISAC Networks, and cooperative jamming Wireless Covert Communications Aided by Distributed Cooperative Jamming over Slow Fading Channels, is public-key-independent randomness, and is therefore subvertible by the same stateful-pseudorandom-generator blueprint that Chen, Huang, and Yung used against KEMs. Our claim has two halves: first, that this "noise-channel algorithm-substitution attack" succeeds with exact distributional covertness and even helped by the secrecy beamforming, at an exfiltration bandwidth so small it suffices to leak a single ephemeral scalar — because in the fragile certificateless schemes of the target deployments, one signature already yields universal forgery Cryptanalysis on Two Certificateless Signature Schemes; second, that anchoring noise generation to the KEM session key eliminates this channel as an independent break, restoring the chaining that hybrid composition was always supposed to provide.
Research Program
The attack. In the OFDMA design of Secrecy Wireless Information and Power Transfer in OFDMA Systems, the base station injects per-subcarrier noise drawn from pre-stored Gaussian sequences indexed by secret pseudorandom keys shared with legitimate receivers. The sequence selection, not the Gaussian generation, is the subvertible path. A subverted transmitter replaces the seed schedule with a stateful generator keyed by an attacker-held subversion key, and embeds leakage bits in the sign of each noise sample:
$$z_i = (-1)^{b_i}\,|g_i|, \qquad b = \mathrm{Enc}_{sk_{\mathrm{sub}}}(\mathrm{leak}_t)$$
where $g_i$ is the nominal Gaussian noise sample on subcarrier $i$, $b_i$ is a bit of the subkey-encrypted leak, and $\mathrm{leak}_t$ is the $t$-th fragment of a session key or a signer's ephemeral scalar. The marginal distribution of $z_i$ is exactly standard Gaussian regardless of $b$, so any statistical test that only examines the noise distribution is blind by construction. Meanwhile, the intended receiver cancels the noise as usual and notices nothing, while the eavesdropper — who, by the design of null-space projection, receives the noise at full power — reads the sign bits at high signal-to-noise ratio. The physics that was supposed to blind Eve becomes a bulk data pipe to her.
The bandwidth requirement is remarkably small, and this is where the certificateless literature does heavy lifting. In the schemes broken by Cryptanalysis on Two Certificateless Signature Schemes, a single observed signature exposes the ephemeral randomness and the user's secret value, yielding the full private key; in Su's short-signature scheme broken by On the security of an enhanced signature scheme, one signature enables universal forgery by linear ratio scaling. So the subverted device need not exfiltrate keys at high rate: a few hundred bits per epoch — one ephemeral scalar — collapses the entire authentication layer. Short signatures exist precisely because these deployments are bandwidth-starved Practical certificateless short signature scheme, which makes the tiny leak channel fit the threat model perfectly.
A unified adversary model. We formalize a third adversary class alongside the standard Type I (public key replacement) and Type II (malicious key generation center) adversaries of certificateless cryptography. The Type I literature — from the degrees-of-freedom forgeries of An Attack on a Certificateless Signature Scheme to the Type I breaks and flawed-proof exposés of On the Security of an Efficient and Robust Certificateless Signature Scheme for IIoT Environments — shows what happens when verification algebra cannot distinguish authentic from attacker-chosen public parameters. Our attack is the physical-layer dual: the receiver cannot distinguish the noise sequence the transmitter should have sent from the one a subverted implementation chose, because nothing binds the choice to any key. Both failures share one root: an unauthenticated degree of freedom. The lesson of non-additive security — that an uncoupled component sinks the composite system — is documented across both the procedural layer (KEM subversion) and the algebraic layer (knapsack-over-pairing independence in Tian et al.'s cryptanalysis); we extend it to the waveform layer, where no prior cross-layer work reaches: the one existing cross-layer metric effort, Cross-Layer Security for Semantic Communications: Metrics and Optimization, bridges beamforming to application-layer task fidelity but treats payloads as abstract bitstreams and assumes honest endpoints.
The defense: KEM-anchored noise. Replace the pre-stored, seed-scheduled noise with a deterministic function of the session key established under the receiver's public key:
$$z_{i,t} = \mathrm{PRF}_{K_{\mathrm{sess}}}(i \,\|\, t), \qquad (c, K_{\mathrm{sess}}) = \mathrm{KEM.Encaps}(pk_B)$$
The receiver still holds a secret (the session key) unknown to the eavesdropper, so the secrecy property is preserved; but the noise is now transitively bound to $pk_B$, destroying the public-key independence that the subversion needs. Detection becomes possible where it was impossible before: the receiver decodes the payload, regenerates the entire expected waveform including the noise, and flags any discrepancy beyond thermal noise. A subverted transmitter can no longer choose the sign pattern freely without triggering this re-derivation check. The honest limitation is that if the KEM itself is subverted à la Chen et al., the session key leaks and the anchoring is moot — but then existing subversion-resilience measures for KEMs apply, and crucially the noise path is no longer an independent second break. The repair converts an unchecked parallel channel into a chained one, which is what composition was supposed to mean all along.
Baselines
The "do nothing" reference point is the honest transmitter running legitimate artificial noise, exactly as deployed in Secrecy Wireless Information and Power Transfer in OFDMA Systems and Generative AI based Secure Wireless Sensing for ISAC Networks; by construction its leakage rate is zero, and its defensive effect is the number to beat or invert: on a USRP N321 testbed with four indoor activities at 100 to 600 packets per second, the diffusion-based safeguarding of Wang et al. degraded unauthorized activity recognition accuracy by approximately 70%, with an Accuracy Degradation Rate reaching as high as 0.82. Our attack should move unauthorized capability back toward the honest-transmitter regime once the seed or key is exfiltrated — if unauthorized accuracy recovers from roughly 30% toward unhindered levels, the attack is decisive. The strongest existing attack baseline is Chen et al.'s KEM subversion Subvert KEM to Break DEM, which recovers session keys from just two successive ciphertexts (no reported runtime result found on Lacuna for this attack). We will run it in the identical channel realizations, eavesdropper geometry, and subversion-key cryptography as our attack: it requires the eavesdropper to read ciphertexts, so under calibrated artificial noise we expect its success rate to collapse while ours holds — that contrast is the central claim. On the defense side, the strongest existing method is the standard-model certificateless scheme of Fully Secure Lightweight Certificateless Scheme for IIoT, proven secure against Type I and Type II adversaries, reporting 5.45 ms for signing plus verification on an Intel Core i5-8500 in C with the PBC library (versus 9.26 ms for the Karati et al. comparison scheme) and a two-element signature. We run it unmodified and show it is bypassed — its proof assumes faithful execution, so a subverted noise channel leaks its users' secrets regardless of the algebra; beating it means demonstrating the bypass and then showing our anchored-noise defense restores the guarantee at derivation overhead that is a small fraction of that 5.45 ms budget on the same hardware class, with the same tuning effort. Finally, the key ablation isolates our one ingredient: the same PRF, the same power allocation, and the same detector, with only the seed source varying between an independent pre-stored schedule (Zhang et al.'s design) and the KEM-anchored derivation. If the defense works only under anchoring, the causal claim is established.
Why not just…?
Why not just fix the KEM subversion and audit the ciphertexts? This is the most natural objection: Chen et al.'s attack is on KEMs, so apply subversion-resilient KEM construction or ciphertext auditing and be done. It falls short because the noise waveform is not a ciphertext — no auditor regenerates or inspects it, since it is defined as non-information-bearing. Our first experiment shows the leak succeeding with ciphertexts fully intact and, more strikingly, with ciphertexts unreadable to the eavesdropper: the noise channel works precisely in the regime where the KEM-subversion channel dies. The two channels are disjoint, so fixing one leaves the other open.
Why not just run anomaly detection on the transmitted waveform? A statistical test on the noise samples cannot catch our sign-embedding construction: the marginal distribution is exactly Gaussian by design, so moment-based, kurtosis, and Kolmogorov–Smirnov tests operate at chance level, which our experiment quantifies. We will also train the strongest plausible detector — a learned classifier on spectrograms and cross-frame statistics, given the same compute as our defense — because a stateful leak might leave higher-order fingerprints. We are honest that this is a real threat: if a learned detector reliably separates subverted from honest transmitters, it is a cheaper countermeasure than KEM anchoring, and we would report it as such. Our expectation, grounded in the exact-distribution construction, is that it will not; the experiment settles it either way.
Why not just use a hardware random number generator or root of trust? A true random source fixes the quality of randomness but not the choice of noise: in Zhang et al.'s design the transmitter selects from pre-stored sequences by a seed index, and in Wang et al.'s design the receiver reproduces the safeguarding signal from a shared seed — in both cases the subvertible component is the scheduling and sharing path, which is software regardless of how good the underlying entropy is. Only making the choice predictable to the legitimate receiver and unpredictable to the attacker — which is what key anchoring does — closes it. The root-of-trust alternative also does nothing for the uncoordinated helper jammers of Wireless Covert Communications Aided by Distributed Cooperative Jamming or the idle-device jammers of Accuracy and Security-Guaranteed Participant Selection and Beamforming Design for RIS-Assisted Federated Learning, which are exactly the low-cost nodes most likely to run subvertible firmware.
Why not just measure whether artificial noise blocks the one-signature attack? A skeptic might prefer the simpler defensive study: put a certificateless signature on the air, add noise, and measure the eavesdropper's bit error rate on the signature tuple — if she cannot receive one pristine signature, the ratio-scaling forgeries of On the security of an enhanced signature scheme never get their single observation. We will run this as a sub-experiment, and we expect it to partially succeed: noise does raise the eavesdropper's error rate on group-element encodings, since a corrupted bit invalidates a curve point. But concluding there would be exactly the wrong lesson, and our attack is the proof: the transmitter that beams noise at the eavesdropper can hand her the key instead of the signature, at better signal-to-noise ratio than the legitimate receiver enjoys. The defensive reading of the intersection is incomplete in a way that actively misleads, which is why the offensive reading deserves the study.
Decisive Experiments and Checks
The core simulation is an OFDMA downlink with null-space artificial noise, using channel matrices from the DeepMIMO ray-tracing environment in the configuration reported by Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO: the O1 outdoor scenario at 140 GHz, an 8×8 uniform planar array at the base station, 4×4 at the receiver, 128 subcarriers, and a 6:2:2 train/validation/test split. We sweep noise power allocation and eavesdropper position, and report the eavesdropper's bit error rate on the sign channel, her success rate in recovering a full session key within a fixed number of frames, and the legitimate receiver's detection rate for a battery of statistical tests — the decisive numbers are a near-zero sign-channel error rate under heavy noise allocation together with a chance-level statistical detection area under the curve. The same run replays Chen et al.'s two-ciphertext KEM attack under identical conditions, and its failure there is the head-to-head result against the strongest existing attack. A hardware demonstration follows the USRP N321 methodology of Generative AI based Secure Wireless Sensing for ISAC Networks: we subvert the shared-seed safeguarding path and measure whether unauthorized activity recognition climbs back from the roughly 30% accuracy their noise enforces (a 0.82 degradation rate) toward the unhindered level — partial recovery, even to 60%, would already establish that the safeguarding channel is reversible by its own subversion. The defense evaluation has two parts: detection, where the receiver's re-derivation check is scored against the learned-detector baseline at matched false-alarm rates, and cost, where per-receiver noise derivation is timed on an Intel Core i5-8500-class machine against the 5.45 ms signing-and-verification budget reported by Fully Secure Lightweight Certificateless Scheme for IIoT, and where the secrecy-rate penalty of receiver-specific versus broadcast noise is measured in the three-cell, four-user-per-cell Rician-fading setting of Secrecy Rate Beamforming for Multicell Networks With Information and Energy Harvesting, whose successive convex approximation converges within 12 to 18 iterations — our variant should converge in the same range with a quantified worst-case-secrecy-rate delta. A final application study places a subverted helper jammer in the 60 m × 60 m, five-device RIS-assisted federated learning setting of Accuracy and Security-Guaranteed Participant Selection and Beamforming Design, testing whether the noise that protects model updates can instead exfiltrate them. Together these experiments either establish the noise channel as a real, closed class of vulnerability — or show that the learned detector or the physical channel itself already blocks it, which we would report with equal interest.
References
- Subvert KEM to Break DEM: Practical Algorithm-Substitution Attacks on Public-Key Encryption
- Multiple-input multiple-output Gaussian broadcast channels with confidential messages
- Secrecy Wireless Information and Power Transfer in OFDMA Systems
- Generative AI Based Secure Wireless Sensing for ISAC Networks
- Wireless Covert Communications Aided by Distributed Cooperative Jamming Over Slow Fading Channels
- Cryptanalysis on Two Certificateless Signature Schemes
- On the security of an enhanced signature scheme
- Practical certificateless short signature scheme
- An Attack on a Certificateless Signature Scheme
- On the Security of an Efficient and Robust Certificateless Signature Scheme for IIoT Environments
- Cross-Layer Security for Semantic Communications: Metrics and Optimization
- Fully Secure Lightweight Certificateless Signature Scheme for IIoT
- Accuracy and Security-Guaranteed Participant Selection and Beamforming Design for RIS-Assisted Federated Learning
- Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO
- Secrecy Rate Beamforming for Multicell Networks With Information and Energy Harvesting