Adaptive Multi-Agent Deep Reinforcement Learning for Distributed Zero Trust Architectures in Global Financial Ecosystems

<title>Abstract</title> <p>Digitalization of global financial systems has been one of the main exposures of financial infrastructures to highly sophisticated cyber attacks. Based on the industry reports released recently, it is stated that close to 25% of global internet attacks target financial organizations and the average data breach cost per attack is about 5.9 million USD. However, although Zero Trust Architecture (ZTA) emerges as a promising security approach that guarantees continuous authentication and stringent control of access for all users and devices, conventional designs of ZTA rely on centralized policy enforcement architectures. This research proposes Adaptive Multi-Agent Deep Reinforcement Learning-Zero Trust Architecture (AMADRL-ZTA). In this regard, AMADRL-ZTA combines multi-agent deep reinforcement learning and distributed Zero Trust policy enforcement, where autonomous agents continuously scan network traffic, evaluate the level of trust, and change access control policies. The experimentally derived results show that AMADRL-ZTA has a 96.8% threat detection capability, 38% reduction in response time, and a near 31% improvement in terms of policy changes in comparison to the existing centralized Zero Trust architecture and the rule-based intrusion detection system. It is evident that the proposed framework has a high chance of improving the cyber resilience of global financial infrastructures and enabling secure digital transactions.</p>