Authors: Sushil Panda
Abstract: CAN bus in software-defined vehicles is vital to enhancing the vehicle's performance, safety, and cybersecurity. The CAN bus is the digital nervous system of modern cars, handling the communication stream between all the Electronic Control Units (ECUs) of a Software Defined Vehicle. This research aims to provide a comprehensive security-aware framework for CAN bus data prediction using advanced temporal neural networks, which are designed for a cybersecurity-aware framework for SDVs. The paper proposes a new hybrid architecture that combines a Transformer-based attention mechanism with novel Graph Neural Networks (GNNs) to capture both temporal dependencies and network topology patterns in bus communications. The approach aims to address the challenges associated with high-frequency, complex time series data while ensuring compliance with ISO/SAE 21434 cybersecurity standards and ISO 26262 functional safety requirements. This is achieved by preserving privacy while simultaneously involving multiple vehicle training and capabilities for detecting real-time intrusion. This paper aims to implement a hybrid architecture with transformers and GNNs together on data using random functions in Python. The results thus obtained demonstrate a significant improvement in prediction accuracy (96.3%), cybersecurity threat detection (98.1% precision), and energy efficiency (34% reduction in computational overhead). The proposed framework achieved ASIL-C compliance and reduced false alarm rates by 31% compared to existing methods while maintaining sub-millisecond inference latency suitable for safety-critical automotive applications.
