Hiding in the Handover: Churn-Cloaked Attacks on Time-Varying Networks

Satellite constellations, drone swarms, and other mobile networks rewrite their own wiring every few seconds: at orbital speeds of roughly 7.5 km/s, links appear and disappear faster than routing updates can propagate, so every few seconds a satellite hands its traffic to a new neighbor (Wang et al.; Saeed et al. in their work). Defenders in these networks must therefore tolerate a constant background of "normal" disruption — dropped packets, sudden reroutes, capacity swings — and any anomaly detector worth deploying learns to shrug it off. That tolerance is an attack surface. An adversary who can forecast the churn schedule (from public orbital data, which is exactly what makes predictive constellation management possible) can schedule its own malicious traffic — exfiltration, routing manipulation, congestion signaling — to land inside the natural noise of handover events, so that the perturbations a detector would normally flag get dismissed as ordinary mobility.

The proposal in one sentence: we will build and analyze an attacker that solves the classic stealth-utility tradeoff by exploiting predictable topology churn as cover, quantify how much this shifts the utility-visibility frontier compared to an attacker blind to churn, and develop the matching churn-conditioned defense that closes most of the gap. This generalizes a proven idea from a stylized setting — Bahrani and Weinberg showed in Undetectable Selfish Mining that a blockchain attacker holding at least 38.2% of hashrate can match the honest network's orphan-block distribution and become mathematically undetectable — to spatially mobile networks where the "natural noise" is rich, scheduled, and publicly predictable. Nobody has done this: the closest works either study adaptive attacks on static software topologies, or dynamic-graph anomaly detection against non-adaptive faults.

Core Claim

Let $c_t$ denote the churn state at time $t$ — which links are entering or leaving visibility windows, and when. A defender runs a detector that assigns an anomaly score $a_t$ and locks out any node whose score exceeds a threshold $\beta$. An attacker chooses actions $x_t$ (injection rate, perturbation magnitude, timing) to maximize utility while staying below the threshold:

$$\max_\pi \; \sum_t U(x_t) \quad \text{s.t.} \quad \Pr(a_t > \beta \mid c_t) \le \delta \;\; \text{for all } t.$$

The key difference from prior stealth formulations (Waniek et al.'s Stackelberg hiding games, Kazari et al.'s constrained-RL Dynamic Adversary) is that the constraint is conditioned on churn state: the detector's noise floor, and hence the effective stealth budget, varies over time. Our claim is threefold. First, a churn-aware attacker achieves strictly higher utility at fixed detection probability $\delta$ than a stationary one, because churn windows raise the noise floor the attacker hides under — the cloaking advantage. Second, the advantage is governed by churn-timing predictability: since orbital mechanics are deterministic enough to build routing plans from knowledge graphs ahead of time (Wang et al.), a defender who conditions detection on the same schedule removes most of the cover; the irreducible advantage scales with residual timing uncertainty (packet jitter, bursty traffic, unannounced maneuvers — precisely the gaps the dynamic-topology literature itself flags). Third, detectors validated only against non-adaptive anomalies — GADY and GeneralDyG — will degrade sharply against a churn-cloaked attacker, exposing a false sense of security of exactly the kind documented across adaptive-attack research (Diaa et al.).

What Exists and What Does Not

Four lines of work bracket this idea without covering it. ATLAS performs adaptive attacks over multi-agent delegation topology, reaching a 90.3% attack success rate with only 6.0 queries per objective on AgentBreaker — but on static three-agent software pipelines with fixed communication edges; it never touches a graph whose links physically vanish mid-attack. GADY and GeneralDyG detect anomalies on continuously evolving graphs — GADY reaching an AUC of roughly 0.98 on Bitcoin-OTC, GeneralDyG an Average Precision of 24.00 on Bitcoin-Alpha at the 1% anomaly setting — but their anomalies are historical or synthetic faults, never a strategic adversary optimizing evasion against the running detector, and never churn driven by physical mobility. Kazari, Shereen, and Dán built the closest attacker: a Dynamic Adversary using constrained RL to maximize sabotage while keeping a normality score above threshold, against which their peer-prediction detector achieves AUCs of 0.96–1.0 and detects attacks in under 10 steps on SMAC-MMM. But their agents sit on a static communication graph, and they explicitly note the detector fails when topology churn separates agents into disjoint partitions — a failure mode that motivates but does not encompass our setting, because our attacker does not need to break monitoring; it hides inside what the monitor already tolerates. Finally, the constellation managers themselves — SkyLink with its decentralized bandit link selection, and SatFlow with hierarchical multi-agent planning — assume cooperative nodes and benign traffic, and their own authors flag unaddressed vulnerability to agents manipulating congestion signals. Our proposal occupies exactly that junction: a strategic, stealth-constrained adversary operating with the defender on a time-varying physical graph.

Research Program

The first workstream builds the attacker. We extend the constrained-RL Dynamic Adversary of Kazari et al. to a time-varying graph, adding churn-phase features to the observation space: time-to-handover per link, predicted capacity transitions, and the residual jitter distribution around scheduled events. The attacker's actions are traffic injection and link-metric perturbation; its utility is exfiltrated volume or induced network cost; its constraint is the output of an actual deployed detector, not a proxy score. The design question is whether the attacker learns to match the conditional distribution of traffic statistics given a churn event (the analogue of the Labeling Strategy in Undetectable Selfish Mining, which matched the honest orphan distribution subject to latency) rather than merely timing bursts to the handover calendar. We will also implement a cheap open-loop variant — gate activity on a fixed handover schedule, no learning — because if churn is fully deterministic, that heuristic may capture most of the gain, and we want to know.

The second workstream supplies the theory. We formalize indistinguishability in the churn-conditioned sense: an attacker is undetectable by any detector if its observable statistics conditioned on $c_t$ equal the natural conditional statistics. This immediately yields the predictability result — when the defender knows $c_t$ exactly, conditioning its detector on $c_t$ collapses the cloaking advantage to zero for attackers that only exploit timing — and locates the attacker's real opportunity in the residual uncertainty the defender cannot forecast. We will state and prove this in a simplified model (independent link events, known churn calendar, Gaussian residuals), then measure how far the simulation departs from it.

The third workstream builds the defense and the evaluation harness. The defense is a churn-conditioned detector: instead of comparing current behavior to a time-averaged baseline, it compares behavior to a per-event baseline indexed by the predicted churn schedule, flagging deviations from the expected disruption rather than disruption itself. The harness couples a LEO constellation simulator following SkyLink's setup — real two-line-element orbital propagation with gridded population traffic, at which scale SkyLink serves 25.4 million users — with a dynamic-graph anomaly monitor, so an attacker can be evaluated against both a learning link manager and a learning detector simultaneously. A smaller space-air-ground testbed modeled on Jia et al.'s SAGIN scheduling environment (2 LEO satellites, 30 UAVs, service-chain traffic) provides a second regime with UAV mobility, where churn is less perfectly predictable than orbital mechanics and the cloaking advantage should grow.

Baselines

The do-nothing reference is the unmodified network: SkyLink running its bandit link management with a uniform-rate malicious node and no detector response, which establishes how much damage an unconstrained, immediately visible attacker does before lockout, and the benign operating point it distorts (SkyLink reports a 95% drop-rate reduction versus k-shortest paths and 99% versus Dijkstra, and holds cost increases under 10% when 3% of ground stations fail — the headroom an attacker attacks). The strongest existing detector baselines are GADY, which reports an AUC of approximately 0.98 on Bitcoin-OTC, and GeneralDyG, which reports an Average Precision of 24.00 on Bitcoin-Alpha at the 1% anomaly setting; both will be run with their published architectures and tuning procedures on identical injected-attack streams, with anomaly volume matched between churn-cloaked and random-timing attackers so only scheduling differs. The strongest attacker baseline is the Kazari et al. Dynamic Adversary ported to the time-varying graph with churn features masked from its observations — identical RL algorithm, training steps, and observation budget, so the comparison isolates churn awareness as the single varying ingredient. This port doubles as the key ablation. We will also run an adversarial-bandit variant of SkyLink's UCB (treating the attacker as a non-stochastic adversary in the bandit sense) and the open-loop handover-calendar attacker described above. Beating the detector baselines means, on the attack side, equal attack utility at a fixed false-positive-matched detection probability while driving the detectors' AUC or Average Precision far below their reported non-adaptive values; on the defense side, it means the churn-conditioned detector restoring detection against the churn-cloaked attacker to within a small margin of what GADY and GeneralDyG achieve against random-timing attacks. For context only, ATLAS's 90.3% success rate at 6.0 queries per objective calibrates how much closed-loop adaptivity buys over its 11.1-query baseline; it is not a head-to-head competitor because its benchmark is a static software pipeline, not a time-varying physical network.

Why not just…?

Why not just condition the detector on the handover schedule? This is a real threat to the proposal's defense side, and we do not argue around it — the churn-conditioned detector is that simple idea, and it is our primary defense baseline rather than a foil. The open question is quantitative: how much of the cloaking advantage does plain per-event conditioning recover, and what remains when the attacker matches the conditional jitter distribution rather than just event timing? Our decisive experiment varies the defender's churn-forecast error (adding jitter and unannounced maneuvers to the simulator) and traces detection performance against it; if the curve stays flat, the simple defense suffices and the paper's contribution collapses to the attack-side result, which we will report honestly.

Why not just use non-adaptive, fixed-pattern attacks? The adaptive-attack literature consistently shows static attacks underperform: ATLAS beats its static-style baseline's query efficiency by 46% (6.0 versus 11.1 queries per objective), and the path-inversion attack of When Efficiency Becomes Fragility induces an 80.9% precision drop on DTB70 with a smaller perturbation budget ($\epsilon = 2.5\%$ versus 3.1%) than FGSM-style baselines. Our own masked-observation ablation provides the direct test: if a churn-blind constrained-RL attacker matches the churn-aware one, the core claim is wrong.

Why not just make the link manager robust with adversarial bandit or robust RL methods? Partially effective and worth running, which is why it is a baseline. But note the mismatch: adversarial bandit algorithms protect the reward signal against arbitrary corruption, whereas our attacker corrupts the observable statistics the detector consumes while leaving rewards consistent with a plausible churn event. Robust methods also pay a regret penalty under the constant benign churn that dominates these networks — exactly the regime where SkyLink earns its gains — so we expect them to trade benign performance for protection against an attack that never touches the channel they guard. The experiment will show whether that expectation holds.

Why not just have the attacker go dark during dangerous windows? Naive disconnection is known to fail on the utility side: Waniek et al. showed that simply severing edges destroys the evader's influence, and their ROAM heuristic beats naive deletion precisely because hiding while retaining function requires adding structure elsewhere. The open-loop handover-calendar attacker is the strongest simple version of this alternative, and we treat it as a first-class baseline: if it matches the closed-loop policy under deterministic churn, that is itself a clean finding — the attack is cheaper than feared — and the closed-loop policy's value will be measured specifically under defender adaptation and timing uncertainty, where a fixed calendar cannot respond.

Decisive Checks

Three results decide the proposal. First, on the LEO simulator, at matched detection probability, the churn-cloaked attacker must deliver materially higher attack utility than the churn-blind Dynamic Adversary port, and the open-loop calendar attacker must sit between them — establishing that churn awareness, not just RL capacity, drives the gain. Second, GADY and GeneralDyG, running at their reported operating points (AUC ≈ 0.98 on Bitcoin-OTC for GADY; we graft a churn process parameterized to the simulator's handover statistics onto these dynamic-graph datasets, since they lack physical mobility, and state this as a controlled proxy), must degrade substantially against the churn-cloaked attacker at anomaly volumes the random-timing attacker cannot hide — if they do not, the false-security claim fails. Third, the churn-conditioned defense must recover most of the lost detection, and the recovery must decay gracefully, not abruptly, as the defender's forecast error grows — mapping, for the first time, how much predictability a time-varying network must lose before its natural motion stops being an attacker's ally.