Adversarial Machine Learning in Smart Grid Intrusion Detection: A Review of Threat Models, Defences, and Scada-Specific Vulnerabilities in Developing Economies

Smart grids in developing economies are surprisingly but rapidly adopting Machine Learning (ML) for Intrusion Detection Systems (IDS) to secure Supervisory Control and Data Acquisition (SCADA) networks. However, the same Machine Learning (ML) models are detrimentally vulnerable to Adversarial Machine Learning (AML) attacks that manipulate input data in order to evade detection. This review examines AML threat models, defense mechanisms, and SCADA-specific vulnerabilities with emphasis on resource-constrained environments in developing economies. A total of 52 studies from 2018-2025 were analyzed and categorized attacks into evasion, poisoning, and model extraction. Key findings indicate that developing economies face compounded risk due to legacy SCADA, limited cybersecurity expertise, and poor data quality. We propose a tiered defense framework combining adversarial training, anomaly detection, and SCADA protocol-aware filtering. Future research should prioritize lightweight defenses, domain adaptation, and threat intelligence sharing tailored to low-bandwidth grids.