Low-Latency Edge-AI Intrusion Detection Framework for Secure Microfluidic Manufacturing Systems

Abstract

 No existing intrusion detection system simultaneously satisfies hard real-time latency bounds for legacy Modbus RTU industrial serial buses, edge-only deployment without cloud connectivity, and deterministic process-level anomaly detection validated on physical PLC-controlled manufacturing hardware. This paper closes this gap with a low-latency edge AI-driven IDS/IPS for precision manufacturing systems relying on Modbus RTU protocols that lack native authentication, encryption, or integrity verification. This exposes deterministic cyber-physical manufacturing environments to unauthorized command injection, Man-in-the-Middle (MitM) interception, replay desynchronisation, Denial-of-Service (DoS) flooding, and register reconnaissance—each capable of catastrophically disrupting precision chemical synthesis, actuator timing synchronisation, and sub-millisecond flow-rate control, This paper presents a low-latency edge AI-driven Intrusion Detection and Prevention System (IDS/IPS) integrating a lightweight Random Forest anomaly detection engine ( trees, 8-dimensional feature vector), a deterministic eleven-rule access control firewall, and a Deterministic Finite Automaton (DFA) process integrity model within a unified hardware-software co-design deployable on existing edge hardware. Experimental validation on a Velocio ACE11 PLC-controlled chemical micro-mixing testbed used a curated 5,700-sample labelled Modbus RTU dataset spanning seven attack categories with systematic sub-variant exploration. The framework achieved: 95.2% detection accuracy (F1: 95.1%), 3.2% false positive rate, 87.4% unknown attack detection, 98.7% firewall block rate, and 17.4 msdeterministic end-to-end latency (65.2% margin below the 50 msPLC threshold). McNemar statistical significance tests confirm RF superiority over all baselines (), except LSTM ( 0.043, not significant with paired testing). Zero-shot cross-dataset transfer on the Mississippi State Gas Pipeline (MSU) Modbus RTU dataset achieves AUC  0.705, confirming protocol-level generalisation. Learning-curve analysis confirms accuracy saturation at  samples, scientifically justifying the curated dataset approach. Mean CPU utilisation was 23.7%, and a 95% cost reduction versus commercial industrial security appliances is demonstrated.