Algorithmic Bias in Machine Learning-Based Cyber Defence: Taxonomy, Mathematical Frameworks, and Ethical Governance

As Artificial Intelligence (AI) becomes the operational backbone of Security Operations Centres (SOCs), algorithmic bias poses a dual threat: technical failure and ethical violation. This study expands the taxonomy of bias within AI-driven cybersecurity systems, focusing on Intrusion Detection Systems (IDS) and Automated Threat Hunting pipelines. Drawing on peer-reviewed literature from 2015 to 2026, it analyses the mathematical foundations of fairness constraints-including Equalised Odds [14] and Predictive Parity-and their application to real-time network anomaly detection. Mitigation strategies across the entire machine learning pipeline (pre-processing, in-processing, and post-processing) [7] are surveyed. The role of Explainable AI (XAI) methods, specifically SHAP [18] and LIME [27], as safeguards against bias is evaluated. Governance implications of Regulation (EU) 2024/1689 (EU AI Act) [25] and the NIST AI Risk Management Framework (AI RMF) [24] are examined. Additionally, adversarial bias injection-including data poisoning attacks in Federated Learning [29] is explored as an emerging offensive vector. The synthesis demonstrates that fairness is not only a social imperative but also a fundamental component of system robustness, and that biased models are exploitable.