14 August 2026 in SSRN Electronic Journal
Aerial vehicular networks formed by cooperative UAV swarms are increasingly used to extend mobile sensing, emergency connectivity, and infrastructure-free communication services. Their high mobility, intermittent multi-hop wireless links, heterogeneous protocol stacks, and constrained onboard storage make intrusion detection difficult under realistic deployment conditions. When new UAV platforms join an operational swarm, their incompatible traffic features and attack semantics must be learned without retaining historical traffic from previous deployments. Existing CNN-based and Transformer-based detectors are designed for fixed protocol stacks and fail under this cross-domain, cross-protocol setting; continual learning methods usually assume homogeneous feature spaces, which aerial vehicular traffic does not provide. We propose ConNeL-IDS, a rehearsal-free security layer for cross-domain intrusion detection in heterogeneous aerial vehicular networks. It operates locally on packet-header and flow-level observations, without changing routing/MAC protocols, injecting control messages, or storing replay traffic. Dataset-Specific Autoencoders project heterogeneous protocol features into a unified embedding, removing the dimensional mismatch between UAV platforms. A Convolutional-Augmented Temporal Sequence encoder captures both packet-level anomalies and multi-stage attack patterns through attention over subsampled wireless traffic flows. Nested Learning with Temporal Scaling and Elastic Weight Consolidation preserve detection knowledge across sequentially arriving domains. On five UAV/FANET intrusion detection domains spanning 37 to 48 feature dimensions and 3 to 12 attack classes, ConNeL-IDS achieves 90.5% average F1 with 2.0% forgetting and 19.3% forward transfer, outperforming thirteen continual learning baselines while requiring no stored historical traffic.
26 October 2026 in Fractals