Real-Time Decision Processing in Artificial Intelligence Systems: A Survey of Scheduling, Inference, and Security Foundations

Abstract

Abstract—Real-time artificial intelligence (AI) decision systems must satisfy three constraints that conventional machine learning pipelines rarely address together: bounded inference latency, deterministic execution under a real-time operating system (RTOS), and resilience against adversarial or unauthorized interference. This survey synthesizes recent literature across these three concerns. We first review classical real-time scheduling theory and RTOS design principles that bound worst-case execution time, together with kernel-level code and data protection mechanisms that secure the execution environment itself. We then examine how comparative evaluation of machine learning approaches and graph-based entity-resolution techniques inform model selection for latency-constrained classification and matching tasks. We survey three deployed case studies — electronic toll collection, peer-to-peer energy trading, and pedestrian trajectory prediction — that illustrate real-time AI decision-making under production constraints. We further review security mechanisms specific to real-time decision pipelines, including behavioral bot detection, federated intrusion detection for networked embedded systems, adversarial robustness, and honeypot-based threat intelligence, and close with low-code/no-code governance as a route to faster, auditable deployment of decision logic. We conclude by identifying open challenges in jointly certifying latency, determinism, and security guarantees for real-time AI systems. Index Terms—Real-time systems, artificial intelligence, decision processing, scheduling, machine learning inference, security, low-code automation.