ThreatSense: Neuro-Symbolic Reasoning for Autonomous Threat Assessment in Unmanned Systems

Abstract

Autonomous unmanned systems increasingly operate in adversarial environments where real-time threat recognition and response are safety-critical imperatives. This paper presents ThreatSense, a novel neuro-symbolic AI framework that integrates deep neural perception with symbolic logic reasoning for autonomous threat assessment in unmanned aerial vehicle (UAV) networks. Unlike conventional intrusion detection approaches that rely solely on either pattern-based learning or handcrafted rule systems, ThreatSense unifies both paradigms through a layered inference architecture: a neural perception module for multi-modal sensor feature extraction, a symbolic knowledge engine encoding domain-specific threat ontologies in first-order logic, and a joint inference layer mediating probabilistic symbol grounding. ThreatSense is evaluated against four representative threat classes including GPS spoofing, denial-of-service, adversarial sensor perturbation, and black hole routing attacks. Experimental results demonstrate a detection accuracy of 97.3%, a false positive rate (FPR) of 1.8%, and an average inference latency of 38 ms, outperforming deep learning baselines by 4.8 percentage points in accuracy while maintaining full symbolic interpretability. The framework advances explainable and computationally efficient threat intelligence for safety-critical autonomous platforms.