An Edge Computing Framework for Deep Learning-Based Anomaly Detection in Satellite-Linked Autonomous Vehicles

MD Mehedi Hasan, Maumita Bhattacharya, Rafiqul Islam

Keywords: Adversarial Machine Learning, Anomaly Detection System, Connected Autonomous Vehicles, V2X Communications, 6G Networks

Abstract:

Modern satellite-linked autonomous vehicles require real-time anomaly detection to address multi-layer cyber threats across CAN Bus,
V2X communication, and sensor systems. Existing cloud-based approaches fail due to 500-600ms satellite latencies incompatible with
safety-critical requirements. We present MARIS-ADS, a constraint-aware multi-modal anomaly detection framework for edge deployment. The system employs attention-based fusion of 41 CAN, 37 V2X, and 28 sensor features, applying entropy-based selection (ARS)
to reduce dimensionality 83% while achieving 99.3% accuracy. Constraint validation enforcing SAE J2735, ISO 11898, and vehicle dynamics constraints filters 82% of adversarial attacks as protocol-invalid or physically impossible. Evaluation demonstrates 96.8% FGSM, 94.2% PGD, and 91.7% C&W robust accuracy with 0.31ms inference on resource-constrained TCUs. Cross-modal attack detection achieves 97.3% accuracy versus 12-34% for single-modality approaches, with comprehensive validation across CICIoV2024, CAN-Intrusion, and BurST-ADMA datasets confirming practical deployment viability for 6G vehicular networks.

Submission Number: 14