Redesigning Transportation Cyber-Physical Systems: Enhancing Connectivity and Bolstering Security Using TinyML

The rapid evolution of Transportation Cyber-Physical Systems (TCPS) necessitates secure, efficient, and real-time data processing capabilities to support autonomous and intelligent transportation networks. However, existing machine learning-based security frameworks face significant challenges, including limited computational resources, data scarcity for model training, and evolving cybersecurity threats. To address these challenges, we propose a novel TinyML-driven security framework optimized for resource-constrained IoT edge devices within TCPS. Our approach integrates Few-shot Transfer Learning to enhance model adaptability and improve anomaly detection accuracy, even with limited labeled data. The core of our framework is DenseCapsNet, a hybrid deep learning model that combines DenseNet (for efficient feature propagation) and Capsule Networks (for hierarchical spatial feature representation), significantly improving cyber threat detection while preserving computational efficiency. Additionally, we incorporate adversarial training and privacy-preserving techniques to enhance resilience against sophisticated cyber threats, ensuring secure data transmission and model robustness. The proposed framework leverages 6G-enabled ultra-reliable low-latency communication (URLLC) to facilitate secure, real-time data exchange across autonomous vehicle networks, overcoming latency and connectivity constraints. Extensive experiments demonstrate that our TinyML-optimized DenseCapsNet model achieves 99.72% accuracy, 99.7% F1-score, and 99.4% recall, significantly outperforming state-of-the-art models such as GoogleNet, HybridNet, and MobileResNet in TCPS security applications. By bridging the gap between efficiency, security, and real-time adaptability, our work contributes a highly scalable, energy-efficient, and privacy-preserving AI solution tailored for next-generation intelligent transportation systems, addressing critical security and computational constraints in real-world TCPS environments.