Predictive Proactive Eavesdropping over Arbitrary Wireless Graphs

When a radio monitor wants to intercept a suspicious transmission but sits farther away than the intended receiver, physics is against it: the target transmits at a rate the monitor's channel simply cannot support, so a passive listener decodes nothing. Proactive eavesdropping turns this around by jamming the receiver, not the transmitter — the degraded receiver forces the transmitter to slow down (through the rate adaptation built into modern wireless protocols), and because the monitor knows its own jamming waveform, it can cancel it and decode the now-slowed signal. This idea is well worked out for one or two links: closed-form jamming power policies exist for single Rayleigh-faded links, HARQ retransmission schemes, UAV relays, and two co-channel links. But every one of these results is a hand-derived formula for a specific, tiny topology. Meanwhile, two adjacent literatures have solved problems that this one needs solved: graph neural networks that generalize wireless resource allocation to arbitrary, never-seen network sizes and shapes, and hybrid statistical-neural forecasters that combine a simple per-node linear model with a neural residual learner and beat both pure alternatives. This proposal connects the three: a proactive eavesdropping system whose jamming policy is a graph neural network that runs on any wireless topology, and whose decisions are driven by forecasts of future channel and traffic conditions rather than instantaneous knowledge the monitor does not have.

Core Claim

The claim is that proactive eavesdropping can be lifted from fixed two- and three-node geometries to arbitrary, time-varying multi-hop wireless graphs, and that this requires exactly the two ingredients the other directions supply. First, the jamming policy must be parameterized by a permutation-equivariant graph neural network operating on the wireless interference graph, because per-topology convex or semidefinite optimization — the approach of the dual-link monitors — cannot be re-derived and re-solved for every mesh, cluster, or scale-free ad-hoc network, and trained graph policies are the only method shown to transfer across graph sizes in wireless resource allocation. Second, the policy should act on forecast channel states rather than instantaneous ones, because a real monitor's channel observations are delayed, partial, and expensive to gather; and the right forecaster for a set of per-link RF time series is a hybrid in which each link gets its own simple statistical model of the predictable component and a shared neural graph model captures the cross-link nonlinear residual. The hybrid structure is not decoration — it is what protects the system against link heterogeneity, the documented failure mode of pure graph models when different links obey different dynamics.

Research Program

The system to be built has two stacked components, both learning over the same dynamic graph $G_t$, whose nodes are transceiver pairs and whose weighted edges carry measured channel gains, interference coupling, and recent rate-activity history.

The forecasting component predicts, for each link $i$, the state of its channel and traffic horizon $h$ steps ahead. It follows the residual decomposition that achieved the best published results on spatiotemporal benchmarks: fit a per-link statistical model (an ARIMA/SARIMA-style local linear model) to each link's observed series to produce a base forecast $\hat{L}_{i,t+h}$; train a spatiotemporal graph neural network only on the residuals $e_{i,t} = Y_{i,t} - \hat{L}_{i,t}$ to produce $\hat{N}_{i,t+h}$; and fuse additively,

$$\hat{Y}_{i,t+h} = \hat{L}_{i,t+h} + \hat{N}_{i,t+h}.$$

This reuses the recipe validated in the critical audit of spatiotemporal benchmarks, where a graph convolutional recurrent network trained on per-node SARIMA residuals set a state-of-the-art MAE of 3.32 on the METR-LA traffic benchmark, and where a plain well-tuned SARIMA alone beat almost every published graph model on WikiMaths — evidence both that the hybrid is the strongest known forecaster of this kind and that pure neural graph models break under node heterogeneity, which is precisely the regime of a wireless network where heavily shadowed links and clear links coexist. Two design points must be settled empirically rather than by convention, because the literature itself is in tension here: the CPU-utilization hybrid paper prescribes first-order differencing for stationarity, while the audit shows differencing acts as an aggressive high-pass filter that destroys low-frequency predictability. On RF series with diurnal traffic cycles and slow shadowing drift, we expect the audit's warning to dominate, and will test both preprocessing paths rather than assume either.

The policy component consumes the forecast graph $\hat{G}_{t+h}$ and outputs a jamming power (and, with multiple antennas, a beam covariance) for each suspicious receiver, trained unsupervised to maximize expected intercepted rate under an average-power budget:

$$\max_\theta \; \mathbb{E}\!\left[\sum_{i} r^{\text{int}}_i\!\big(P_J^{\,\theta}(\hat{G}_{t+h})\big)\right] \quad \text{s.t.} \quad \mathbb{E}\!\left[\sum_{i} P_{J,i}\right] \le \bar{P},$$

where $r^{\text{int}}_i$ is the rate the monitor can decode on link $i$ after the transmitter throttles in response to the jamming. Training uses the model-free primal-dual policy gradient scheme proven for GNN resource allocation in fronthaul networks, replacing the cooperative capacity objective with the surveillance objective and, critically, with the nonlinear coupling that makes eavesdropping hard: jamming too much suppresses the suspicious link entirely and, as the UAV relay study showed for decode-and-forward relaying, can even spill into the monitor's own reception. The graph parameterization is what buys topology generalization: the theory-to-practice GNN paper proves that message-passing architectures align with iterative wireless optimization and achieve near-optimal behavior on a 20-user network with 40 training samples where an unstructured network needed 64,000, which is the sample-efficiency argument for why the policy should not be a flat multilayer perceptron over concatenated channel states.

A third thread runs through both components: topology churn. Mobile and ad-hoc networks change adjacency over time, and both the eavesdropping literature and the hybrid forecasting literature assume static or near-static structure. The forecaster handles this by design — the per-link statistical models survive edge rewiring because they are local, and the residual graph model is re-fed the current adjacency each step — and we will evaluate explicitly under node mobility, where a purely graph-structured forecaster trained on a static adjacency has no mechanism to adapt.

Baselines

The naive reference point is passive eavesdropping: the monitor only listens and never jams. It is the correct null because it needs no full-duplex hardware and no self-interference cancellation, and it is quantitatively characterized: the Rayleigh-fading proactive eavesdropping paper reports that when the monitor's channel gain is below $-15$ dB relative to the suspicious link, passive interception falls to near zero, while proactive jamming recovers substantial throughput; the UAV relay study likewise finds optimized jamming strictly dominates passive listening at every point on the flight trajectory. Any configuration of our system that fails to beat passive listening in the weak-channel regime has failed outright.

The strongest existing surveillance method is the optimal cognitive jamming policy of the cognitive-jamming paper, which uses Lagrange duality to derive threshold and water-filling power allocations for both delay-sensitive (non-outage probability) and delay-tolerant (relative intercepted rate) objectives. It is the strongest competitor because it is the best published optimality result in this problem family; but it is defined only for an isolated link, so on multi-link graphs we run it greedily per suspicious receiver with the same average-power budget our method gets, and on a single-link validation setup we check that our learned policy recovers its qualitative behavior (non-monotonic rate-versus-power, threshold structure). The file reports no numerical value for this policy's intercepted rate, so we will establish its numbers ourselves in the shared simulator; no reported result found on Lacuna for its multi-link extension. The second strong competitor is our own architecture with instantaneous, perfect channel state information instead of forecasts — an ablation that isolates the forecasting ingredient entirely, run with the same GNN, data, tuning budget, and compute, so the only difference is whether the policy sees $G_t$ or $\hat{G}_{t+h}$. Feasibility of the GNN side is supported by the link-scheduling comparison study, where supervised GNN schedulers on synthetic device-to-device interference graphs achieved over 80% of the exhaustive-search optimal sum-rate at $K=10$ links; that paper reports no result on adversarial or surveillance objectives, which is exactly the gap we fill. Finally, the forecasting stack is baselined internally by pure per-link SARIMA (no graph residual) and pure spatiotemporal GNN (no statistical base), with the published hybrid — which reported MAE 3.32 on METR-LA — serving as both the recipe and, in a traffic-graph sanity check, the reference number. Beating the baselines means: higher average intercepted rate and eavesdropping non-outage probability than passive, greedy-cognitive, and constant-power jamming at equal power budget; a gap to the perfect-CSI ablation that stays small as observation delay grows; and forecast MAE below both pure forecasters.

Why not just…?

Why not just deploy more passive listening nodes? Adding monitors raises diversity, but the limit is physical, not statistical: the Rayleigh-fading paper shows interception collapsing to near zero whenever the monitor channel is $-15$ dB or weaker, because the suspicious transmitter's rate exceeds the monitor's Shannon capacity outright. More receivers help only if at least one already has a comparable channel; jamming-induced rate throttling is the only mechanism in this literature that changes the capacity relationship itself, and the dual-link study confirms regions where passive eavesdropping yields exactly zero while jamming-assisted monitoring yields positive rate.

Why not just jam at maximum power, or with simple on/off pulses? The single-link result is explicit that average intercepted rate is non-monotonic in jamming power — it peaks and then decays — so maximum power is provably suboptimal even on one link, and the cognitive-jamming paper shows its optimized allocation beats both constant-power and on-off jamming. On multi-hop graphs the problem worsens: uncalibrated jamming spills into neighboring links and into the monitor's own reception, and the UAV study shows constant-power jamming under decode-and-forward actually decreases intercepted rate as transmit power grows. Constant power stays in the baseline set precisely because it is the cheap heuristic a practitioner would try first, and we expect the learned policy's advantage to grow with graph size.

Why not just train one end-to-end deep network (GNN or LSTM) to do forecasting and policy jointly? This is the most serious alternative. The sample-efficiency result cuts in its favor at small scale (a GNN reached near-optimal on a 20-user network with 40 samples), so with abundant simulator data a monolithic policy might match us. Three pieces of evidence argue it will not: pure MLP-style unstructured policies needed 64,000 samples for the same performance; pure graph forecasters lost to a plain SARIMA on WikiMaths and struggled specifically under node heterogeneity, which is severe in wireless; and the hybrid beat its own pure-neural component on server telemetry (mean relative error 0.0186 for ARIMA–BP versus 0.0265 for BP alone), roughly a 30% improvement over the pure neural model. Still, we treat this as a genuine threat, not a straw man: the joint end-to-end policy with no statistical decomposition is an explicit ablation, given the same total parameter count and tuning budget, and if it matches the hybrid under full-information training, the honest conclusion would be that the hybrid's value lies only in the delayed-CSI and low-data regimes, which we will report as such.

Why not just solve each topology exactly with convex optimization or game theory? The dual-link eavesdropping papers already do semidefinite programming over jamming covariances, and the Stackelberg surveillance work derives optimal patrol policies — but the latter only for complete, complete-bipartite, and star graphs, explicitly leaving arbitrary graphs open, and the SDP approach must be re-solved from scratch for every topology and every channel realization, which the fronthaul GNN paper identifies as the core reason real-time model-based optimization is intractable in wireless settings. Exact optimization remains in the proposal as ground truth: we use per-instance SDP or exhaustive search on small graphs to score what percentage of optimal the learned policy achieves, exactly as the link-scheduling study scored its schedulers against exhaustive search.

Proposed Experiments

Three simulation environments, all grounded in the file. First, a single-link Rayleigh-fading setup replicating the foundational study, used to validate that the learned policy reproduces the known optimum and its non-monotone power curve. Second, synthetic device-to-device interference graphs with $K \in \{4, 6, 8, 10\}$ link pairs, dual-slope path-loss with log-normal shadowing, following the link-scheduling comparison study's protocol of 256 training and 256 test samples per configuration, extended to cross-size transfer (train on $K=4$, test on $K=10$ and beyond) and to multi-hop mesh, clustered, and scale-free topologies that no eavesdropping paper has touched. Third, a churn setting with mobile nodes and time-varying adjacency, with channel series generated to include diurnal traffic cycles and heterogeneous shadowing so the hybrid forecaster's strengths and the differencing question are both exercised.

Metrics are the ones the eavesdropping literature already reports: average intercepted rate, eavesdropping non-outage probability, and relative intercepted rate, plus the forecasting MAE of predicted channel states and the fraction of the per-instance-optimal (SDP or exhaustive-search) surveillance value achieved, mirroring the normalized sum-rate protocol of the scheduling study. Reported anchors from the file: passive interception near zero at gains below $-15$ dB; GNN schedulers exceeding 80% of optimal sum-rate at $K=10$; hybrid SARIMA-plus-GNN MAE 3.32 on METR-LA; ARIMA–BP mean relative error 0.0186 against 0.0265 for pure BP and 0.0370 for pure ARIMA on server telemetry. No reported result found on Lacuna for intercepted rate of any proactive eavesdropping method on graphs of more than two links, which is the quantitative void this proposal aims to fill.

The decisive checks are four. The learned policy must transfer across graph sizes and structures with performance degrading gracefully rather than collapsing, where per-topology SDP provides the ceiling. It must degrade slowly as CSI delay grows from zero to several slots, while the perfect-CSI ablation and an instantaneous-CSI classical policy degrade faster — this is the test that predictive jamming under delayed observation is a real capability, not a convenience. The hybrid forecaster must beat both pure components on link-heterogeneous RF series, with and without default differencing, resolving the preprocessing tension in favor of whichever the data supports. And the whole pipeline must survive topology churn — nodes joining, leaving, and moving — a regime in which the entire prior proactive-eavesdropping literature, locked to fixed geometries with perfect channel knowledge, has no entry. One scope boundary is worth stating plainly: like the foundational papers, we assume the suspicious transmitter reacts naively to jamming rather than mounting counter-surveillance of the kind studied in the covert-communication literature; making the policy robust to an aware target is the natural next step, not part of this proposal.