TY - EJOU
AU - Şahin, Canan Batur
AU - Razak, Siti Fatimah Abdul
AU - Ullah, Arif
AU - Gündüz, Ali Fatih
AU - Nawi, Nazri Mohd
TI - HADAR-UAV: Risk-Calibrated One-Class Learning Framework for Zero-Day Intrusion Detection in Unmanned Aerial Vehicle Networks
T2 - Computers, Materials \& Continua
PY -
VL -
IS -
SN - 1546-2226
AB - Unmanned Aerial Vehicle (UAV) networks face escalating cybersecurity threats, especially from zero-day attacks that exploit previously unknown vulnerabilities. To address this, we present HADAR-UAV (Hybrid Anomaly Detection with Adaptive Risk-calibration for UAV). This novel intrusion detection framework integrates masked autoencoder representation learning with Deep Support Vector Data Description (Deep SVDD) under conformal prediction guarantees to calibrate risk. Our method overcomes three critical limitations of existing approaches: (i) over-reliance on attack signatures, (ii) lack of statistical guarantees on false alarm rates, and (iii) insufficient robustness in feature extraction under partial observation. Using a rigorous Leave-Two-Attack-Families-Out (L2AFO) evaluation protocol on the UAVIDS-2025 benchmark, HADAR-UAV achieves strong zero-day detection—0.997 ± 0.001 ROC-AUC and 0.992 ± 0.002 F1-Score—while empirically maintaining a target false alarm rate through conformal calibration applied to deterministic scores. All results are reported as mean ± standard deviation across 20 independent runs (5 seeds × 4 folds) and show statistically significant improvement (paired t-test, p < 0.01) over current one-class methods. Ablation studies confirm that every architectural component adds measurable value to the framework. Additional cross-dataset validation on NSL-KDD under a one-class zero-day-inspired setting further indicates that the proposed framework generalizes beyond MAVLink-specific traffic patterns.
KW - UAV cybersecurity; intrusion detection; zero-day attacks; one-class classification; masked auto encoder
DO - 10.32604/cmc.2026.080874