Missing-Energy Tagging for Silent Transmitters in Mobile Wireless Meshes

Predictive routing for drones and vehicles works by having every node broadcast its position and velocity, so that neighbors can compute when a link will stretch past radio range or when an interferer will drift out of earshot. These predictors share one blind spot: they assume every transmitter that matters is announcing itself. A node whose positioning has failed, a radio that transmits but never beacons, or a jammer is invisible to them. Its interference then shows up as unexplained link failures, and the predictor has no way to reason about the cause. Particle physicists at electron-positron colliders solve the same kind of problem when they detect neutrinos, which leave no trace in the detector. They fully reconstruct one of the two B mesons produced in each collision (the "tag"), subtract everything they can account for from a precisely known total, and treat the residual, including leftover calorimeter energy that no reconstructed particle explains, as evidence of something unseen.

We propose to apply this procedure to radio energy. Each node already decodes neighbors' frames and can measure each decoded frame's received power individually. Subtracting these accounted-for contributions from the total channel energy it measures in each time slot leaves an "unexplained energy" residual. Pooling residuals across nodes whose positions are known lets us detect, locate and track silent emitters. Their predicted trajectories then enter the interference-duration and link-lifetime estimates that predictive routing already uses. The idea borrowed from tagging is specific: interpret the residual only in windows where the visible side is cleanly accounted for, and accept lower coverage in exchange for purity.

Core Claim

Link-lifetime predictors fail in a systematic way when non-beaconing transmitters are present. A residual-energy accounting step, gated by a per-window "accounting purity" score, can recover enough information about those transmitters to restore most of the lost routing performance. It should do so better than naive energy thresholding and better than a learned predictor that sees the same raw measurements. What is new is not measuring energy, which any radio does. It is (i) the subtraction of decoded-frame energy from measured energy, rather than of path-loss-predicted energy, to form the residual, (ii) the purity gating that decides which residuals may be trusted, and (iii) feeding a tracked hidden-emitter trajectory into kinematic interference-duration prediction.

Where the Directions Meet, and What Prior Work Leaves Open

The networking side supplies the machinery we extend. OPAR and LB-OPAR take three consecutive GPS fixes, fit 3D velocity and acceleration, and solve for the link-failure time by bisection. They then pick routes that balance path lifetime against hop count. A Cross-Layer Routing Protocol Based On Minimum Interference Duration (MIR) goes further by estimating how long each interferer stays inside the carrier-sensing range, $t_{\text{leave\_away}}$, and routing to minimize interference duration relative to link duration. MIR is the protocol our method most directly extends. Every one of these protocols can only reason about nodes that beacon or produce decodable signals. We keep their prediction and routing machinery unchanged and add hidden emitters to the set of nodes they reason about.

Two lines of networking work motivate the gap. Geographic routing in the presence of location errors and Connectivity vs. control show that position- and heading-based predictors degrade quickly when their inputs are wrong. The limiting case of a wrong input is a node that reports nothing at all. Security-Sensitive Task Offloading in Integrated Satellite-Terrestrial Networks treats adversarial risk at the level of encryption and offloading policy, but not the physical interference an adversary causes. One line of work constrains the design. Investigation of routing reliability of vehicular ad hoc networks (AODV-R) shows that probabilistic link-survival models absorb unmodeled disturbances well, raising delivery from about 37% to about 56% at 140 km/h. Its success is a serious threat: a hidden emitter might be treated just as well as extra noise. We address this directly in the experiments below.

The physics side supplies the method, not an application. In hadronic-tag analyses such as Search for B0 meson decays into Lambda and missing energy with a hadronic tagging method at Belle, one B meson is fully reconstructed by Full Event Interpretation, a staged cascade of boosted-decision-tree classifiers. Invisible particles are then inferred from the missing four-momentum, together with checks on leftover calorimeter energy that no reconstructed particle accounts for. BABAR used the same logic in Measurement of the semileptonic decays B- -> D0 tau- nu_tau... and Search for the rare decay B -> K nu nubar to isolate multi-neutrino decays. The lesson we reuse from these analyses is that the residual is informative only when the visible side is reconstructed with high purity. Tag efficiency is typically only 0.1–0.3%, and analysts accept that loss because an impure tag smears the missing-mass signal into background. B-flavor tagging at Belle II adds a cautionary result: a structured, category-based tagger (effective efficiency 30.0%) held up on real data better than an end-to-end deep network (28.8%), whose gap between simulation and data was larger. This informs our choice to make the subtraction explicit rather than learned end to end. No paper in the file connects the two fields.

Research Program

The first component is the per-node energy ledger. In each slot of a few milliseconds, a node records total in-band received energy $E_{\text{tot}}$ and the energy of every frame it decoded in that slot. The unexplained residual is

$$r = E_{\text{tot}} - \sum_{j \in \mathcal{D}} \hat{E}_j - N_0,$$

where $\mathcal{D}$ is the set of decoded frames, $\hat{E}_j$ is the measured received energy of frame $j$, and $N_0$ is the calibrated noise floor, all in linear units. Using measured rather than path-loss-predicted energies removes shadowing uncertainty from the visible side. This mirrors how a tag is built from measured tracks rather than from expected ones.

The second component is purity gating. Collisions, partial decodes, and frames from beaconing nodes at the edge of range all leave residual energy that no hidden emitter produced. This is the wireless counterpart of combinatorial background from mis-reconstructed tags. We will train a small boosted-tree classifier on slot-level features, including decode failures, CRC errors, carrier-sense state, and the number of beaconing neighbors near the edge of range. It outputs a purity score, and only slots above a threshold contribute residuals. Sweeping the threshold traces a coverage-versus-purity curve, the analogue of the efficiency-purity trade-off in tag selection.

The third component is joint inference and tracking. Purity-gated residuals from nodes at known positions define a likelihood over emitter position and transmit power under a log-distance path-loss model with log-normal shadowing. A particle filter carrying a Gauss-Markov motion model tracks each detected emitter. The track's predicted trajectory is inserted into MIR-style $t_{\text{leave\_away}}$ computations and into OPAR's path-lifetime objective as an additional interfering node, weighted by track confidence. The routing and offloading optimizers are left unchanged, so any gain can be attributed to the new input.

Baselines

The naive reference is AODV and OLSR with no prediction, run in the same scenarios, which shows how much a silent emitter hurts the unprotected case. The strongest existing methods are OPAR and LB-OPAR, which are predictive but blind to hidden emitters. They are strongest because they report the largest gains in exactly the NS-3 UAV setting we use. LB-OPAR reports up to a 400% throughput increase and a 30% higher flow success rate over AODV/OLSR. OPAR reports more than 30% higher file delivery, 25% higher throughput and 35% lower flow completion time than AODV, with about 0.83 MB of control overhead for 50 UAVs. These are relative figures. No reported result was found on Lacuna for any of these protocols in the presence of non-beaconing emitters.

MIR is the closest interference-aware competitor. It reports 28.5% lower delay, 32.3% lower packet loss and 53.2% lower routing overhead than AODV. AODV-R represents the approach of absorbing the disturbance statistically. All baselines will be reimplemented in the same NS-3 build, with identical mobility seeds, radio parameters and traffic, and with an equal hyperparameter search budget for any tunable threshold.

We add two further competitors. The first is an energy-detection baseline: flag any slot whose total energy exceeds the floor plus the sum of decoded energy by a fixed margin, with no purity gating and no tracking. The second is a learned predictor, a graph neural network given the same raw per-slot energy logs and neighbor tables and trained to predict link breakage directly. It receives the same simulated training volume.

Three ablations isolate the key ingredients: removing purity gating, replacing measured decoded energies with path-loss-predicted ones, and dropping the tracker in favor of per-slot inference. Beating the strongest baseline means a statistically significant improvement over OPAR/LB-OPAR in packet delivery and flow completion time when hidden emitters are present, while staying within 5% of them when none are, so that the added machinery does not cost performance in clean conditions.

Why not just…?

Why not just threshold excess energy and avoid busy links? This is the most likely simple winner when hidden emitters are sparse and strong. The energy-detection baseline is included for exactly that reason. Our expectation is that without purity gating, collisions among beaconing nodes will produce false alarms that grow with density across the 50–100-node range, causing needless route switches. The ablation without gating, compared against the threshold baseline, settles this. If the two perform equally, the tagging contribution reduces to tracking alone, and we will report it that way.

Why not just let a probabilistic reliability model like AODV-R absorb the effect? AODV-R shows that treating unmodeled disturbances as noise pays off. It does, however, spread the penalty across all links instead of the specific links a moving emitter will cross, and it cannot anticipate where the emitter is heading. If AODV-R matches our method against stationary jammers but falls behind against moving ones, the tracking step is justified. If it matches in both cases, the idea is not worth its overhead.

Why not just train an end-to-end learned predictor on the raw measurements? The Belle II result in B-flavor tagging at Belle II suggests that structured pipelines transfer better when simulation mismatches reality. In our setting the analogous mismatch is shadowing. We test this by training the graph neural network under one shadowing standard deviation and one mobility model, and evaluating under others. If the learned model is robust to these shifts and matches our method, the explicit subtraction is unnecessary, and we would say so.

Proposed Experiments

The primary testbed is NS-3 with 50–100 UAVs under 3D Random Waypoint and 3D Gauss-Markov mobility, matching the OPAR/LB-OPAR setup. We inject 1–5 non-beaconing emitters whose duty cycles range from sparse to near-continuous, which are either stationary or moving. The secondary testbed uses SUMO Luxembourg bus traces with the same injection, following the trace choice of Energy-Latency Tradeoff for Dynamic Computation Offloading in Vehicular Fog Computing. There we also measure offloading failures caused by interference in the middle of a task.

Component metrics are hidden-emitter detection ROC, localization error in meters, and error in interference-duration prediction. End-to-end metrics are packet delivery ratio, throughput, flow completion time, routing overhead, and, for the vehicular runs, offloading failure rate and energy-latency cost.

Three checks decide whether the idea holds. First, with zero hidden emitters, our method must be indistinguishable from OPAR. Second, the coverage-versus-purity curve must show that residual-based localization improves as the threshold tightens even while coverage falls. If it does not, the tagging analogy contributes nothing. Third, performance under shadowing that differs from training must separate the explicit subtraction from the learned baseline. Everything here is simulation. A small software-radio replay would come after a positive result, and we make no claim beyond simulated channels.

References

  1. OPAR: Optimized Predictive and Adaptive Routing for Cooperative UAV Networks
  2. LB-OPAR: Load balanced optimized predictive and adaptive routing for cooperative UAV networks
  3. A Cross-Layer Routing Protocol for Mobile Ad Hoc Networks Based On Minimum Interference Duration
  4. Geographic routing in the presence of location errors
  5. Connectivity vs. control: Using directional and positional cues to stabilize routing in robot networks
  6. Security-Sensitive Task Offloading in Integrated Satellite-Terrestrial Networks
  7. Investigation of routing reliability of vehicular ad hoc networks
  8. Search for <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math> meson decays into <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="normal">Λ</mml:mi></mml:math> and missing energy with a hadronic tagging method at Belle
  9. Measurement of the semileptonic decays<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mover accent="true"><mml:mi>B</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mo>→</mml:mo><mml:mi>D</mml:mi><mml:msup><mml:mi>τ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>τ</mml:mi></mml:msub></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mover accent="true"><mml:mi>B</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mo>→</mml:mo><mml:msup><mml:mi>D</mml:mi><mml:mo>*</mml:mo></mml:msup><mml:msup><mml:mi>τ</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>τ</mml:mi></mml:msub></mml:math>
  10. Search for the rare decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:mi>K</mml:mi><mml:mi>ν</mml:mi><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover></mml:math>
  11. B-flavor tagging at Belle II
  12. Energy-Latency Tradeoff for Dynamic Computation Offloading in Vehicular Fog Computing