Adaptive Defense Optimization Under Intelligent Data Poisoning Attacks in Industrial Control System

Abstract

As industrial control systems (ICSs) 4.0 evolve into open and networked environments, they face a critical challenge in maintaining robust closed-loop operation under feedback streams that are asynchronous, unreliable, and subject to adversarial data poisoning. This paper proposes a detection-aware stochastic optimization framework to address adaptive decision-making under these adversarial conditions. Unlike evaluations relying solely on synthetic threat models, we utilize the SWaT dataset to characterize realistic and temporally irregular poisoning behaviors. To mitigate the risks of timing-based attacks, we introduce an adaptive acknowledgment (ACK) bundling mechanism that regulates feedback timing and reduces communication overhead. Central to our approach is a multi-objective stochastic gradient descent (SGD) algorithm incorporating a residual-based anomaly indicator. This algorithm dynamically adjusts the ACK bundling window online, balancing control performance, communication efficiency, and anomaly awareness. The proposed framework enables adaptive communication adjustment when anomalous behavior is detected, while Lyapunov–Krasovskii (LK)-based conditions are established to ensure closed-loop stability. Numerical experiments on an autonomous surface vehicle (ASV) benchmark demonstrate adaptive ACK-window regulation, closed-loop recovery under AI-driven data poisoning attacks, and favorable delay tolerance over the tested parameter settings.