Quantifying the Value of Link Prediction and Behavioural Trust in Infrastructure-Free Mesh Routing

Abstract

Infrastructure-free mesh networks must route over
links that appear, degrade and vanish faster than any control
plane can advertise them. A common proposal is to let a graph
neural network (GNN) predict which links are about to fail and
route around them pre-emptively. We build that system end
to end — a seeded mobility-and-radio simulator, a deterministic
multi-metric baseline, a 5,089-parameter GraphSAGE linksurvival
predictor, and a behavioural trust graph — and measure
what each component is worth, reporting the components that
do not pay as carefully as the ones that do. We first show that
the choice of delivery metric decides the answer: in a correctly
buffered store-and-forward mesh, eventual delivery saturates
near 0.95 for every strategy we test and ranks hop-count routing
above loss-aware routing, while delivery within a 2 s deadline
ranks them the other way round, the two curves crossing at
about 9 s. Measured on the deadline an application actually
experiences, three findings stand out. First, the deterministic
multi-metric cost is a failure: tuned on disjoint seeds it does
not beat the single classical ETX metric (0.736 against 0.746),
and an ablation shows why — removing its energy, battery or
latency terms improves on-time delivery by up to 1.5 percentage
points each. Only convexity in packet loss earns its place, worth
3.7 points, and only behavioural trust earns more. Second, linksurvival
prediction does pay, and an oracle bounds how much:
perfect foreknowledge of the next 5 s is worth 2.6 points, of which
our 19.9 kB model captures 83% (+2.9%, p = 0.030, with route
breaks cut 43%). At a 20 s horizon that headroom persists but our
predictor cannot reach it, gaining nothing while the oracle gains
3.1 points — a limitation of the model, not of the idea. Third,
the behavioural trust graph is the component that pays outright,
recovering 8.8%, 18.1% and 24.3% of on-time delivery lost to
10%, 20% and 30% malicious relays that each silently drop 80%
of transit traffic (all p < 10−4), at a 0.95–0.97 detection rate, a
false-positive rate below 0.03, and no significant cost when no
attacker is present (p = 0.701), and it still pays against adaptive
attackers that evade detection. We also derive and confirm the
closed-form ultra-low-bandwidth frontier imposed by the 158-
byte header-plus-cryptography floor. All code, seeds and results
are released.