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

Abstract

Infrastructure-free mesh networks must route overlinks that appear, degrade and vanish faster than any controlplane can advertise them. A common proposal is to let a graphneural network (GNN) predict which links are about to fail androute around them pre-emptively. We build that system endto end — a seeded mobility-and-radio simulator, a deterministicmulti-metric baseline, a 5,089-parameter GraphSAGE linksurvivalpredictor, and a behavioural trust graph — and measurewhat each component is worth, reporting the components thatdo not pay as carefully as the ones that do. We first show thatthe choice of delivery metric decides the answer: in a correctlybuffered store-and-forward mesh, eventual delivery saturatesnear 0.95 for every strategy we test and ranks hop-count routingabove loss-aware routing, while delivery within a 2 s deadlineranks them the other way round, the two curves crossing atabout 9 s. Measured on the deadline an application actuallyexperiences, three findings stand out. First, the deterministicmulti-metric cost is a failure: tuned on disjoint seeds it doesnot beat the single classical ETX metric (0.736 against 0.746),and an ablation shows why — removing its energy, battery orlatency terms improves on-time delivery by up to 1.5 percentagepoints each. Only convexity in packet loss earns its place, worth3.7 points, and only behavioural trust earns more. Second, linksurvivalprediction does pay, and an oracle bounds how much:perfect foreknowledge of the next 5 s is worth 2.6 points, of whichour 19.9 kB model captures 83% (+2.9%, p = 0.030, with routebreaks cut 43%). At a 20 s horizon that headroom persists but ourpredictor cannot reach it, gaining nothing while the oracle gains3.1 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 to10%, 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, afalse-positive rate below 0.03, and no significant cost when noattacker is present (p = 0.701), and it still pays against adaptiveattackers that evade detection. We also derive and confirm theclosed-form ultra-low-bandwidth frontier imposed by the 158-byte header-plus-cryptography floor. All code, seeds and resultsare released.