SwarmGuard: Serverless IDS for Byzantine UAV Swarms Under Dual-Layer Attack

Abstract

UAV swarms rely on AODV, a cooperative routing protocol where every node forwards packets for its neighbors, which a compromised node can exploit by advertising optimal routes and then silently dropping traffic. As UAV swarms adopt federated learning (FL) for intrusion detection, a second attack surface opens: a compromised node can mislabel its local training data to corrupt the shared detector while simultaneously launching the routing attack it should be helping to catch. Existing approaches aggregate models through a central server, assume honest participants, and measure success by detection accuracy alone, without measuring network recovery. We present SwarmGuard, a serverless federated IDS for Flying Ad-hoc Networks (FANETs) that operates over existing AODV peer links: each node scores every neighbour’s model by its effect on local detection accuracy, accumulates this as an EMA reputation score, and excludes neighbours below a principled threshold with no coordination required. In a controlled evaluation ($N=12$ nodes, 3 seeds), SwarmGuard outperforms FedAvg, Gossip, and Krum on macro-F1 and matches FLTrust without a trusted root dataset, with attacker-identification AUROC reaching 0.97–1.000 under coordinated poisoning. Beyond detection, SwarmGuard excludes confirmed malicious nodes from routing, restoring packet delivery to $H_{t}=1.000 \pm 0.000$ at both 17% and 33% Byzantine ratio, compared to baselines that lack any isolation mechanism and remain at the no-isolation floor. We introduce Network Health (NH) to formalise this recovery. An ablation confirms isolation is governed by the EMA threshold alone; an optional PPO policy further reduces the mean false-exclusion rate from 37.5% to 10.4%.