Air-gapped computers leak secrets by turning their own electricity into a signal. Malware on an isolated machine can encode bits in the current it draws from the wall, the heat it sheds, or the radio noise its cables radiate, and a nearby receiver reads them out — PowerHammer moves 1000 bits per second through a power cord, BitWhisper moves 1–8 bits per hour as heat, and TEMPEST-LoRa turns a monitor cable into a LoRa radio transmitter reaching 87.5 meters. For a defender, the catch is that these leaks are tiny needles in a huge haystack: a stolen 256-bit key occupies a few seconds of a months-long recording, and nobody has labeled examples of the attack in your particular building's electrical noise. The only public dataset with real obfuscated power traces, Chameleon, shows that a convolutional network can segment such signals well — but only after training on ground-truth timestamps produced by special on-chip hardware that a defender monitoring a facility will never have.
This proposal argues that the defender's problem is, structurally, a cold-start active learning problem: with zero labeled examples, an untrained detector's uncertainty estimates are garbage, and random labeling wastes an analyst's scarce time on the wrong windows. The cold-start literature has solved versions of this for text, images, and graphs — but every method assumes a modality the defender does not have. We propose the first cold-start acquisition method that operates directly on continuous, unsegmented physical telemetry, and we test it where it hurts: against hardware countermeasures that deliberately deform the signal in time and frequency.
Core Claim
A defender can bootstrap detection of low-bandwidth exfiltration from raw power (or magnetic/thermal) telemetry with a small number of analyst-labeled windows, by combining three predictor-independent signals: self-supervised pretext difficulty on waveform windows, weak supervision from a handful of recordings of known attack families, and diversity over a spectral-temporal similarity graph. The claim has a falsifiable core: PT4AL showed that pretext-task loss ranks sample difficulty well on images (Spearman's ρ ≈ 0.8) PT4AL, while Chameleon showed that frequency-scaling countermeasures collapse learned decoders to a 0.7% key-recovery success rate Chameleon. Whether the pretext-difficulty proxy survives physical obfuscation is open, and we measure it directly, obfuscation regime by obfuscation regime.
Research Program
The unit of labeling is a window: a fixed-length slice of the telemetry stream, described by a vector of spectral and temporal features (band-power in known exfiltration bands such as PowerHammer's 15–24 kHz conduction range, transient statistics, cyclostationary features). The defender's pool is tens of thousands of unlabeled windows; the analyst can afford to label perhaps tens to hundreds. Everything below is about choosing which ones.
The first component adapts FeatProp's decoupling trick to streams. FeatProp propagates raw features over a static citation graph and clusters the result, bypassing any model's uncertainty FeatProp; we replace the citation graph with a dynamically built similarity graph over waveform windows, where edges connect windows that resemble each other in the spectral feature space, and run the same propagation plus k-medoids to find representative windows. This gives coverage of the haystack without any labels.
The second component is a pretext task for waveforms. PT4AL ranks images by rotation-prediction loss; augmentations-heavy contrastive methods actually degraded the ranking PT4AL, which is encouraging for us because our pretext — predicting masked segments of a window's spectrum, or detecting time-warping of a window against its neighbors — uses mild, physics-compatible transformations rather than heavy augmentation. We measure the rank correlation between pretext loss and true informativeness (does the window contain a covert carrier?) under each Chameleon obfuscation regime, since Dynamic Frequency Scaling and Chaffing exist precisely to break learned temporal-frequency structure.
The third component adapts STENCIL's weak supervision. STENCIL seeds acquisition with 15–25 exemplar phrases and submodular mutual information functions that balance relevance to the exemplars against batch diversity, lifting cold-start accuracy by 10–18% and rare-class F1 by 17–40% on text with 1:6 to 1:18 imbalance STENCIL. Our exemplars are not phrases but short recordings of known attack waveforms — a captured PowerHammer-style binary frequency-shift pattern, a TEMPEST-LoRa-style chirp — and the mutual information is computed over the window similarity graph. The combined acquisition score for a window $w$ is
$$a(w) \;=\; \alpha \, \hat{L}_{\text{pretext}}(w) \;+\; (1-\alpha)\, \frac{1}{|E|}\sum_{e \in E} \operatorname{sim}\big(z(w), z(e)\big),$$
where $\hat{L}_{\text{pretext}}$ is the window's normalized pretext loss, $E$ is the exemplar set, $z(\cdot)$ is the propagated window embedding, and $\alpha$ is a single mixing weight tuned on held-out exemplars. Batches are then chosen to maximize facility-location coverage of the graph subject to this score, so the method queries windows that are both hard-for-the-pretext and exemplar-adjacent, while not spending the whole budget on one cluster. This is the same "avoid the missed-cluster effect" logic that DEUCE implements with text embeddings and zero-shot language-model predictions DEUCE — but DEUCE needs RoBERTa, and there is no pretrained language model of power traces. Our graph propagation and pretext loss are the label-free substitute for that missing foundation model.
Proposed Experiments
All experiments run on the Chameleon dataset — over 600 GB of physically measured RISC-V power traces across five regimes: unprotected baseline, Dynamic Frequency Scaling, Random Delay, Morphing, and Chaffing Chameleon. Chameleon does not itself contain exfiltration traffic, so we superimpose synthetic carriers modeled on the published attack physics: PowerHammer's binary and multi-frequency shift keying of CPU load, and TEMPEST-LoRa-style chirps, at realistic signal-to-noise ratios and duty cycles corresponding to payloads of 10 bits per second down to a few bits per hour. Ground truth for evaluation comes from the injection timestamps, held out from all acquisition methods; the hardware Pinpointing Unit labels are used only to define Chameleon's own supervised reference results. The headline settings are the obfuscation regimes, because Chaffing injects decoy cryptographic threads — exactly the workload-mimicry condition under which pure anomaly detectors and fixed demodulators should fail.
Metrics are cold-start detection F1 and localization IoU for covert windows after the first analyst batch (the analog of PT4AL's iteration-1 accuracy), the full labels-to-detection curve up to 500 labeled windows, rare-class F1 for the covert class under its extreme imbalance, and analyst-side acquisition latency. Acquisition latency matters at this scale: ALISE's whole contribution was cutting selection from 3,956 seconds to 6.9 seconds on a large pool ALISE, and a method that takes hours to pick windows on 600 GB of telemetry will not be used. We also report the pretext-informativeness rank correlation per obfuscation regime, as a standalone scientific result.
Baselines
The naive reference point is random window selection with the same downstream detector retrained after each batch; this is the standard cold-start floor, and the literature warns it is where untrained-uncertainty methods often land — Liu et al. formalize the optimism bias that makes early uncertainty sampling worse than random, and the introductory survey documents cold-start learners routinely underperforming random sampling Introduction to Active Learning. Uncertainty sampling with the untrained detector is included as a second reference point to reproduce that failure mode on telemetry.
The strongest existing method as a direct competitor is PT4AL's pretext-ranking strategy, transplanted as faithfully as the modality allows: the same rank-by-pretext-loss acquisition, the same batch stratification, with our masked-spectrum pretext in place of image rotation. PT4AL reports 55.83% iteration-1 accuracy on CIFAR-10 versus 46.02% for random sampling, and 95.13% at 10,000 labels versus 89.93% for Learning Loss PT4AL — the strongest evidence anywhere in this literature that pretext difficulty beats random cold starts, which is why it, not DEUCE or STENCIL, is the method to beat. DEUCE and FeatProp and STENCIL are modality-bound to text embeddings, static graphs, and phrase exemplars respectively, so they enter as component-level ablations (graph coverage without exemplars, exemplar scoring without pretext) rather than whole-system baselines. Two further competitors bracket the design space: an unsupervised autoencoder anomaly detector that flags the most anomalous windows for labeling, and matched-filter demodulation at the known carrier frequencies, the approach PowerHammer's own receivers use. The Chameleon supervised CNN, which reaches 84–91% IoU across all four obfuscation types with 0% false positives and false negatives after post-processing Chameleon, serves as the oracle upper bound — it sees labels no defender could have, and the question is how close the label-efficient methods get. All methods share the same window pool, the same downstream detector architecture and training budget, the same analyst batch sizes, and the same compute for acquisition; the only varied factor is which windows get labeled. Beating the baselines means: cold-start F1 strictly above random and uncertainty sampling at every batch size, above the PT4AL transplant on at least the obfuscated regimes, and within roughly ten IoU points of the Chameleon supervised oracle using no more than five percent of its labeling budget. The key ablation removes the exemplar term ($\alpha = 1$) and the pretext term ($\alpha = 0$) separately, to isolate whether each predictor-independent signal earns its place — and a third ablation that drops graph coverage entirely tests whether diversity matters once exemplars and pretext are present.
Why not just…?
Why not just run the supervised Chameleon CNN on a small labeled set? If a defender can label, say, 500 random windows and train the existing CNN, why build an acquisition method at all? The honest answer is that this is exactly the random-sampling baseline, and the cold-start literature's central empirical finding is that it systematically misallocates the budget: with the covert class at perhaps 1:1000 prevalence rather than Chameleon's dense cryptographic operations, 500 random labels may contain zero positive examples. Our decisive check is the labels-to-detection curve: if random-label CNN training reaches our method's F1 at equal budget, the proposal fails on its own terms. We note that no reported result found on Lacuna for active learning or cold-start experiments on the Chameleon dataset — this comparison has never been run.
Why not just use unsupervised anomaly detection? An autoencoder that flags unusual windows requires no labels at all, which is simpler than anything we propose. This is a real threat, and we expect it to win on the unprotected baseline regime. But Chaffing floods the trace with decoy cryptographic threads that are anomalous in exactly the way the true covert signal is Chameleon, so an anomaly detector cannot tell decoy from target and spends the analyst's budget labeling decoys. The experiment that settles it: anomaly-ranked acquisition versus our method specifically on the Chaffing and Morphing regimes, measuring what fraction of labeled windows are true covert carriers. If anomaly ranking matches us there too, the exemplar and pretext components are dead weight.
Why not just demodulate at the known carrier frequencies? PowerHammer's receivers decode fixed bands at 15–24 kHz, achieving 1000 bit/s with 0% error at line level PowerHammer — if you know the carrier, you do not need learning. But the same paper shows the cost of not knowing: at the building phase level, interference drops throughput to 10 bit/s with 4.2% error, and the survey's own open-problem list names automated signal discovery in unknown noise environments as unsolved PowerHammer. Chameleon's Dynamic Frequency Scaling result — CNN key recovery collapsing to 0.7% success Chameleon — is direct evidence that frequency-deforming obfuscation breaks methods that assume a stable carrier, including matched filters. Our method makes no fixed-carrier assumption; it learns which spectral neighborhoods deserve labels. The matched-filter baseline quantifies exactly how much this flexibility costs or earns under each regime.
Decisive Checks and Honest Limits
Three results would make or break the proposal. First, the pretext-informativeness rank correlation under Dynamic Frequency Scaling and Chaffing: if obfuscation breaks the proxy the way it broke Chameleon's decoders, the whole acquisition score needs a different difficulty signal, and we would report that as a negative result with the same care as a positive one. Second, the Chaffing-regime comparison against anomaly ranking, which directly tests whether workload mimicry — the hardest thing about real covert channels — is what the exemplar component is for. Third, extreme imbalance: STENCIL's exemplar mechanism was validated only up to 1:18 class skew STENCIL; our covert-class prevalence will be orders of magnitude lower, and if the missed-cluster effect reappears at that skew despite exemplars, the honest conclusion is that cold-start active learning needs a rare-class discovery mechanism we have not yet built. The scope is deliberately a defender's tool, not a new attack: no result here improves any exfiltration channel, and every experiment is designed so that publishing it helps the monitoring side only.
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
- Chameleon: A Dataset for Segmenting and Attacking Obfuscated Power Traces in Side-Channel Analysis
- PT4AL: Using Self-supervised Pretext Tasks for Active Learning
- Active Learning for Graph Neural Networks via Node Feature Propagation.
- STENCIL: Submodular Mutual Information Based Weak Supervision for Cold-Start Active Learning
- Deuce: Dual-diversity Enhancement and Uncertainty-awareness for Cold-start Active Learning
- Adversarial Active Learning for Sequences Labeling and Generation
- Shift-Pessimistic Active Learning Using Robust Bias-Aware Prediction
- Introduction to Active Learning and Application to Material Engineering