Trustworthy Machine Learning for Cybersecurity:A Decision-Centric Survey of Explainability, Uncertainty, and Human Factors

Trust failures in automated decision-support systems rarely stem from the absence of sophisticated models; instead, they emerge when system outputs are misaligned with the constraints and responsibilities of human decision-makers. In cybersecurity operations, analysts must act under adversarial pressure, incomplete evidence, and asymmetric error costs, conditions known to amplify both over-reliance on automation and premature disuse after failures. Recent research proposes explainability and uncertainty-aware learning as foundations for trustworthy machine learning, yet these mechanisms are often developed and evaluated in isolation from operational decision workflows. This survey synthesizes prior work across cybersecurity, machine learning, and human-automation interaction through a decision-centric lens. Rather than organizing the literature by algorithms or application categories, we structure evidence around the analyst decisions that systems are intended to support and the trust signals those decisions require. We contribute a unified taxonomy linking decisions to trust requirements and implementation strategies, and we identify trust-signal misalignment-in timing, granularity, and actionability-as a central barrier to dependable deployment. The paper concludes with design implications and research directions that frame trustworthy cybersecurity AI as a human-AI coordination problem, emphasizing calibrated decision support over isolated model improvements.