Resilient Sensor Fusion via Hybrid Trust

Simple fusion methods can be driven far from reality by a single faulty or malicious sensor. This work treats sensor fusion as a distributed trust-management problem, inspired by swarm intelligence and blockchain principles.

An off-chain layer detects inconsistencies in real time with Leave-One-Out (LOO) consistency checks, comparing each sensor with the aggregate of the others. An on-chain layer maintains a reputation per sensor using an exponentially weighted moving average, so that persistent, low-intensity misbehaviour leads to disqualification while reliable behaviour slowly restores trust.

MATLAB simulations with five sensors (20 Hz, 60 s) under spoofing, bias-drift and coalition attacks showed faulty sensors disqualified in 0.45 s (spoofing), 0.40 s (two-sensor coalition, thanks to the honest majority) and 0.95 s (subtle bias drift). Strict settings react within 0.45 s, while tolerant settings take 3.50 s but behave more stably — a tunable sensitivity–stability trade-off. Presented at ISNET 2025, Antalya.