Distributed H−/H∞ Attack Detection for Intelligent Connected Vehicle Platoons: A Coupling-Aware Observer Approach

Cooperative driving stacks for intelligent connected vehicles (ICVs) depend on a continuous exchange of position and velocity data over vehicle-to-vehicle links. That dependence is also their weakness: an adversary who tampers with exchanged measurements can push a platoon toward unsafe spacing before any physical symptom appears. This paper develops a distributed attack-detection scheme for ICV platoons that is sensitive to such data manipulations yet remains robust to road disturbances. Each vehicle runs a local observer whose correction term mixes its own sensor residual with the state estimates broadcast by its immediate neighbours, so the wind-resistance coupling between vehicles is exploited rather than ignored. Using a modal decomposition of the platoon dynamics over the communication graph, sufficient conditions for an H∞ disturbance-attenuation level and an H− attack-sensitivity level are derived as a common set of linear matrix inequalities; a bisection algorithm then computes a single observer gain meeting both. Detection rests on a windowed residual-energy test with per-vehicle thresholds calibrated as empirical quantiles over an ensemble of attack-free runs. On a five-vehicle platoon, the detector identifies transient FDI and replay attacks with detection rates of 100% and 100% at a validated false-alarm rate of 0.46%, reaches a mean detection rate of 97.7% over the transient-attack windows of a 300-run Monte Carlo study, and exposes a denial-of-service attack on a single link through the open-loop prediction mismatch it induces (48.8% of the attacked window); stealthy zero-dynamics attacks, which by construction excite the least observable directions of the plant, are flagged on 0.0% of attacked samples in the documented low-energy configuration, while an amplitude–rate–location sweep shows detection climbing to 98% once the attack energy grows, confirming attack energy rather than placement as the operative limit. Relative to an H∞-only observer, it trades empirical detection aggressiveness for a structural guarantee: a certified worst-case attack-sensitivity floor that the baseline does not provide and that survives coupling levels at which uncertified designs lose feasibility. A closing experiment embeds the detector in a cooperative adaptive cruise control loop: mean peak spacing error during a sensor attack averages 7.93 m, against 10.79 m without detection, while forced false-alarm rates of 5–15% latch the conservative control schedule and degrade spacing beyond the no-detection case; a calibrated low false-alarm rate is a safety property, not a convenience.