Threat-Oriented Digital Twinning for Autonomous Security Evaluation

A novel approach for assessing cybersecurity in learning-enabled autonomous systems employs threat-oriented digital twinning. This open-source, modular twin simulates a standard autonomy stack, incorporating distinct functions for sensing, autonomy, and supervisory control, alongside confidence-gated multi-modal perception, clear trust boundaries for commands and telemetry, and runtime hold-safe behavior. The methodology converts threat assessments into tests addressing spoofing, replay, malformed input injection, reduced sensing capabilities, and adversarial machine learning stress. Although it has been applied to a ground-based proxy, the framework is centered around the autonomy stack. The research, available on arXiv (2604.25757), seeks to foster open, unclassified studies on secure autonomy through a reproducible design pattern.