Answers are grounded in this paper and cite their sources.
Industrial Control Systems (ICS) underpin critical national infrastructure, from electrical grids and water treatment facilities to oil and gas pipelines, yet remain vulnerable to highly stealthy cyberattacks that exploit weaknesses in legacy protocols such as Modbus/TCP. Among these threats, Least Significant Bit (LSB) manipulation enables adversaries to embed covert command-and-control or data exfiltration channels within protocol-compliant traffic, evading conventional intrusion detection systems entirely. The covert nature of these channels allows persistent adversaries to maintain undetected footholds in operational networks for extended periods, posing a direct risk to the continuity and integrity of critical infrastructure operations. This paper proposes an end-to-end lightweight and explainable intrusion detection framework to identify LSB-based covert channels in Modbus/TCP communications. Three tightly coupled design decisions distinguish the framework from prior ICS intrusion detection systems. First, a protocol-aware bit-level feature representation decomposes Modbus Protocol Data Units (PDUs) into their constituent LSBs, yielding a compact 146-dimensional vector that explicitly aligns the feature space with the sub-byte manipulation surface that flow-level and packet-level representations cannot expose. Second, a purpose-built Separable Convolutional Neural Network (SCNN) exploits the spatial structure of this representation to capture short-range temporal dependencies and cross-bit correlations with 61% fewer parameters than a standard 1D CNN, enabling deployment on resource-constrained edge hardware without sacrificing detection fidelity. Third, Integrated Gradients (IG) attribution is applied directly over the bit-indexed feature space, producing explanations that map back to named Modbus register fields and specific LSB positions, making detections interpretable to ICS operators. We evaluated the proposed framework on a public Modbus covert-channel dataset using stratified four-fold cross-validation. The SCNN achieved a mean accuracy of 99.86% and an F1-score of 99.90%. On Raspberry Pi 4B hardware, the model achieved a median inference latency of 0.0558 ms per preconstructed 10-packet window, corresponding to approximately 17,922 window predictions per second. These results demonstrate the computational feasibility of lightweight edge inference and motivate further end-to-end validation in industrial monitoring environments.
No takes yet. Share an insight, caveat, or question.
Iliyasu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: