An air-gapped computer — one with no network connection to the outside world — is supposed to keep its secrets by topology. But every computation a processor performs costs power, and power leaves traces: fluctuations in the current drawn through the power cord, heat, faint radio noise. Twenty years of covert-channel research has shown how malware can deliberately toggle CPU load, fan speed, or screen pixels to carry bits across the air gap through these physical side effects, as in PowerHammer, BitWhisper, and TEMPEST-LoRa. The recurring weakness is that this modulation requires obviously artificial workloads — tight loops slamming cores between 0% and 100%, or synthetic "attack videos" streamed through the display pipeline — exactly the execution patterns host monitoring is built to catch.
Meanwhile, air-gapped and isolated facilities increasingly run neural networks around the clock: surveillance cameras, industrial inspection, perimeter monitoring, all doing inference on streaming sensor data. That inference is legitimate work, invisible to anomaly detectors by construction. And its cost is not constant. Multi-exit networks — models with early "shortcut" exits so that easy inputs finish after a few layers while hard inputs run the full depth — make inference duration depend on the input itself. Prior work proved this dependency is exploitable, but only by redesigning the model: an electromagnetic covert channel built on a deliberately pruned, hostile architecture achieves a clean timing signal on edge devices, and a weight-level neural Trojan leaks classifications through power traces — both requiring the attacker to control the model's weights or training pipeline. This proposal asks the harder, more realistic question: can an attacker who controls only the inputs — never the weights, never the code — turn an unmodified, cryptographically signed, deployed network into a transmitter of arbitrary data? The answer would combine the four directions as follows: the physical covert channel is the core mechanism, air-gap exfiltration is the application regime, adversarial stealth is the constraint on what inputs are allowed, and black-box attack methodology is the search procedure that finds them.
Core Claim
There exist small, natural-looking perturbations to camera or sensor frames that steer an unmodified multi-exit network to an attacker-chosen exit depth while leaving both the predicted label and the frame's visual appearance intact. By injecting such frames into the normal inference queue, malware encodes arbitrary bits — each frame carries $\log_2 E$ bits, where $E$ is the number of usable exits — into per-frame inference duration, which a nearby receiver reads as conducted power or electromagnetic fluctuations. The perturbations are found without gradients or confidence scores, using only measured wall-clock inference time as feedback, in a few thousand queries per symbol.
Formally, for a benign frame $x$, a perturbation $\delta$, an exit index $e(\cdot) \in \{1,\dots,E\}$, a predicted label $f(\cdot)$, and a perceptual distance $d$, the attacker solves
$$\min_{\delta}\; d(x,\,x+\delta)\quad\text{subject to}\quad e(x+\delta)=k,\;\; f(x+\delta)=f(x),$$
where $k$ is the exit index that encodes the next symbol of the payload, and $\delta$ is parameterized in a low-frequency or generative-latent basis so that $d$ stays small implicitly. The label constraint matters as much as the perceptual one: a surveillance system whose outputs suddenly drift is itself an anomaly detector, so the attack must move computation without moving decisions.
The concrete target: exfiltrate a 256-bit key in a few seconds of a 30-frame-per-second stream at a bit error rate at or below 1% after error correction, while the frame stream, the model's outputs, and host telemetry remain statistically indistinguishable from their benign baselines. That throughput class — tens of bits per second — is exactly the regime where low-bandwidth covert channels already cause total security failure, since a 256-bit AES key suffices to break everything else.
Research Program
Threat model and hardware. The victim is an air-gapped edge computer in the class of the NVIDIA Jetson devices used in the PMU electromagnetic channel or the standard desktop used in PowerHammer, running a signed multi-exit model (MSDNet- or Shallow-Deep-style, following DeepSloth) on a live camera stream. The model may sit behind a secure-boot boundary or a trusted execution environment, a separation demonstrated to be permeable by thermal side channels in Through Fabric — so the malware can enqueue frames and observe timing but cannot touch weights. The receiver is either a split-core current transformer on the power cord, PowerHammer's non-invasive rig, or a software-defined radio near the device's power management circuitry, the PMU channel's rig, which operated within 70 cm and saw a maximum timing estimation error of 11 ms — a number that directly sets our symbol spacing and guard intervals.
Symbol design. We build an M-ary exit-shift keying scheme: an exit-index alphabet, a measured map from exit depth to observed current or emission waveform, and error-correcting coding sized to the confusion matrix between adjacent exits. The PMU paper's 11 ms timing noise gives us the physical noise floor to design against; its finding that class-dependent timing was decodable with 100% accuracy (albeit with an architecture built for separability) suggests the channel itself is clean when symbols are well separated.
Black-box steering search. The only feedback the attacker gets is a noisy scalar: how long inference took. We adapt greedy random-basis search in the spirit of SimBA, which needed an average of 1,283 queries (median 582) to fool ImageNet classifiers using only confidence feedback, but over directions chosen in a low-frequency cosine-transform basis or a generative latent space rather than raw pixels. Two known traps shape the design. First, restricted frequency bases have dead ends: SimBA's low-frequency variant failed on 2.5% of ImageNet images even after 60,000 queries, so our search must adaptively widen its basis when steering stalls. Second, the search queries are themselves frames entering the inference queue, so probe traffic must look like normal frames at normal cadence — which the latent parameterization provides, and which blunts stateful query-correlation detection of the kind studied in Stateful Detection of Black-Box Adversarial Attacks, since successive probes are natural images rather than near-duplicates.
Universal perturbations. To drive the online query count toward zero, we precompute per-exit universal perturbations offline against a surrogate model and transfer them, motivated by Universal DeepSloth in the slowdown-attack paper, which achieved 5–45% early-exit efficacy reduction single-shot across architectures. The transfer gap between surrogate and deployed model is a central measured quantity, not an assumption.
Stealth layer. The perceptual constraint draws on the camouflage literature: AdvRM disguises adversarial road patches as lane markings, and Adversarial Wear and Tear disguises them as weathering; our perturbations should read as ordinary sensor noise, compression artifacts, or scene variation. The situational-stealth lesson of SOPE — that exfiltration blended into normal tool activity evaded nine defenses while maintaining 77% attack success — transfers directly: our "tool activity" is the inference queue itself.
Defense analysis. PowerHammer proposes software jammers that inject random CPU load, and Sponge Examples proposes hard latency cutoffs. We measure both: a jammer must mask legitimate inference to defeat us, which throttles the facility's real work, and our modulation is designed to stay below cutoff thresholds while remaining decodable. This is offensive-security research reported with its countermeasures, in the evaluative tradition of the cited channel papers.
Baselines
The do-nothing reference point is the unmodulated benign stream: the model processing ordinary camera frames while a receiver watches the power line. This measures natural timing leakage — inference duration already correlates with scene content, which is the phenomenon the PMU channel exploited, though only after redesigning the architecture to make it decisive. If natural leakage alone carries a decodable payload, the adversarial ingredient is unnecessary; this baseline defines the capacity floor the attack must beat.
The strongest existing method on the same physical medium is PowerHammer's stress-loop modulation, PowerHammer, which toggles CPU cores between 0% and 100% load and reports 1000 bit/sec with 0% error at line level (transformer clamped on the power cord) and 10 bit/sec with 4.2% error at phase level (breaker panel), using frequency-shift keying between 15 kHz and 24 kHz. We will re-implement its modulation on our hardware with the same receiver, the same 256-bit payload, the same error-correction target, and the same detection budget. Beating it means delivering the payload at comparable or better effective bit error rate while the detection suite — host telemetry monitors, frame-stream statistics, output-drift checks, and a stateful query detector — flags it near certainty and flags us near chance. PowerHammer itself concedes it must hide in idle cores when host monitoring is present, which throttles it; our comparison makes that trade-off quantitative rather than anecdotal. No reported result found on Lacuna for PowerHammer's detection rate under a specific host-monitor configuration, so we measure it ourselves on shared hardware.
The strongest existing neural channel is the hostile-architecture PMU electromagnetic channel, which reports 100% eavesdropping accuracy on Jetson devices at up to 70 cm. It cannot be run against an unmodified model at all — it requires pruning, retraining, and installing a replacement network — so the comparison is over the assumption set rather than a head-to-head run; we reuse its receiver front-end and decoding approach for signal-processing parity, and its 11 ms timing noise figure as our design constraint.
Two ablations isolate the proposal's key ingredients. A naive-perturbation ablation uses unconstrained pixel noise for steering with no semantic or latent constraint, isolating the stealth ingredient; and a white-box oracle assumes the weights are somehow leaked and steers exits by direct gradient, isolating the cost of the black-box search. Every variant gets the same hardware, receiver, payload, error-correction coding, and detection suite, and the black-box variants get a matched query budget.
Why not just…?
Why not just toggle the CPU like PowerHammer? It is simpler and faster on raw throughput — 1000 bit/sec at line level versus our projected tens of bits per second — and if a facility has no host monitoring, it strictly wins. But its own paper acknowledges that its 0%-to-100% core saturation is exactly what host-based monitoring catches, forcing it into idle cores and throttled rates; and the PMU channel paper makes the same point about "injecting useless computation" generally. The head-to-head experiment settles it: both attacks run under the same detector suite, and we report each one's throughput–detectability frontier. Our claim is only that in the stealth-constrained regime the frontier shifts decisively in favor of riding legitimate inference.
Why not just trojan the model during training? The neural Trojan side-channel work achieves 56–89% output recovery from power traces with only a 2–4% accuracy drop, and if the attacker controls the training pipeline or weight distribution, that is indeed the more efficient attack — a real threat we do not dispute. But it cannot be launched post-deployment against signed, read-only, or enclave-resident weights, and it can only leak the model's own classification of the current input, not arbitrary host data such as keys or files. Our threat model — unprivileged code execution plus input control, against a fixed benign model — is precisely the gap left open, and it is the gap that matters when model provenance is verified.
Why not just use plain norm-bounded noise instead of latent-space stealth? Unconstrained $L_p$ noise is faster to search and well understood, and if no human ever reviews the footage and no one monitors the model's outputs, it may work as well — we say so plainly. The evidence against it in monitored settings is that high-frequency pixel noise is conspicuous to human observers, the failure mode that pushed the stealth literature toward semantic camouflage in AdvRM and Adversarial Wear and Tear; and our label-preservation constraint means noise that flips predictions trips an output-drift monitor regardless of perceptual quality. The naive-perturbation ablation measures the exact boundary: we report detection cost curves for plain noise versus latent-constrained steering under human review, output monitoring, both, and neither, so the reader can see in which regimes the extra machinery pays for itself.
Proposed Experiments
The feasibility study (E1) deploys multi-exit models trained on CIFAR-10 and Tiny ImageNet — the benchmarks used in DeepSloth, which reduced early-exit efficacy by 90–100% and amplified latency 1.5–5× with white-box perturbations — on a Jetson-class device and a desktop, and measures the full chain from exit depth to inference duration to transformer-measured current and SDR-measured emission. The output is an exit-alphabet confusion matrix under the 11 ms-scale timing noise reported by the PMU channel, and a choice of exit spacing and coding that holds the raw bit error rate near 1%.
The search study (E2) measures steering success rate and query count per symbol for the low-frequency greedy search, the latent-space search, and the white-box oracle. The anchor for expectations is SimBA's 1,283 average queries against ImageNet classifiers with confidence feedback; ours is a harder setting — a noisy scalar rather than class scores — so we treat that figure as an optimistic floor and define failure as needing more than roughly 50,000 queries per symbol. The transfer study (E3) measures how per-exit universal perturbations built on surrogates degrade on the deployed model, anchored to Universal DeepSloth's 5–45% single-shot efficacy range.
The end-to-end exfiltration run (E4) delivers a 256-bit key over the power-line channel and reports delivered bit rate, bit error rate after coding, and total exfiltration window, side by side with the re-implemented PowerHammer baseline under identical conditions; we expect to lose on raw rate and win on the joint metric. The stealth study (E5) runs the full detection suite — host telemetry anomaly detection, frame-stream statistical tests, output-drift monitoring, a stateful query detector in the style of Stateful Detection of Black-Box Adversarial Attacks, and a forced-choice human study on perturbed versus benign frame pairs — for our attack, the PowerHammer baseline, and the naive-noise ablation, reporting flag rates and detector AUC for each. The defense study (E6) activates a software jammer as proposed in PowerHammer and a latency cutoff threshold as proposed in Sponge Examples, and measures both our post-jamming bit error rate and the degradation the jammer inflicts on the facility's legitimate inference throughput.
The decisive checks are three. If black-box steering cannot reach reliable symbol control within a practical query budget, or if natural-looking perturbations cannot hold the bit error rate below a decodable threshold, the core claim fails. If the do-nothing baseline already carries the payload, the adversarial mechanism is unnecessary. And if PowerHammer-style stress loops achieve equal stealth under the same detector suite, the entire motivation for using the model as the transmitter collapses — the experiments are built to expose that outcome rather than argue around it.
References
- PowerHammer: Exfiltrating Data From Air-Gapped Computers Through Power Lines
- BitWhisper: Covert Signaling Channel between Air-Gapped Computers Using Thermal Manipulations
- TEMPEST-LoRa: Cross-Technology Covert Communication
- An Electromagnetic Covert Channel based on Neural Network Architecture
- Leveraging Neural Trojan Side-Channels for Output Exfiltration
- A Panda? No, It's a Sloth: Slowdown Attacks on Adaptive Multi-Exit Neural Network Inference
- Through Fabric: A Cross-world Thermal Covert Channel on TEE-enhanced FPGA-MPSoC Systems
- Simple Black-box Adversarial Attacks
- Stateful Detection of Black-Box Adversarial Attacks
- Beware of Road Markings: A New Adversarial Patch Attack to Monocular Depth Estimation
- Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples
- SOPE: Situation-Aware and Statistically Indistinguishable Privacy Exfiltration for MCP-enabled Agents
- Sponge Examples: Energy-Latency Attacks on Neural Networks