War Machine Learning - AI in Defence

Abstract

This dissertation examines the development and integration of machine learning within the military domain, arguing that the primary constraint and greatest opportunity for advancing military Artificial Intelligence (AI) is the data ecosystem. Across research in computer vision (CV), reinforcement learning (RL), federated learning (FL), and generative AI (GenAI), the analyses consistently show that progress is limited by systemic issues related to data availability, quality, and infrastructure.The work synthesizes findings from six original publications to demonstrate that practical military AI requires a shift from an algorithm-centric view to a holistic, system-focused perspective that treats data as a first-class operational capability. To bridge the gap between high-level strategy and granular technical research, this dissertation adapts the Cross-Industry Standard Process for Data Mining (CRISP-DM) as a framework for assessing military AI applications.Key findings from the studies validate this thesis. A CV study on sonar imagery highlighted model failure due to poor-quality sensor data, underscoring the need for integrated data pipelines. RL research revealed that a lack of high-fidelity simulators and operational data hampers real-world transfer. The investigation into GenAI identified a dependency on proprietary models misaligned with military needs, proposing FL as a secure, collaborative paradigm for developing military-specific foundation models. Finally, an ethical analysis addresses the "reliability-oversight paradox" in autonomous systems, proposing a new human-machine teaming model of human support rather than simple oversight.In conclusion, this dissertation claims that the effective integration of AI into military forces depends on building a robust data ecosystem that includes expertise and understanding on doctrinal and policy-making levels, data and algorithm understanding on the technical level as well as governance, operator-in-the-loop feedback and annotation mechanisms, and an interoperable infrastructure.