An Intellectual Zero Trust Security Framework Using Deep Reinforcement Learning for Predictive Threat Mitigation in AI-Based Fraud Detection Systems
Author:
Open Access
- 13 February 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 24602-24617
Abstract
The rapid growth of Artificial Intelligence (AI) driven financial platforms has led to an increase in sophisticated cyberattacks and fraudulent activities that easily bypass traditional security controls. Current fraud detection mechanisms mostly utilize fixed rule-based engines, monitored classifiers, based on signature, or high-dimensional decision space and all of them find it difficult to cope with changing threat patterns, zero-day attacks, and high-dimensional decision spaces. Even though the recent research has covered machine learning and deep learning architectures, the solutions remain to have significant limitations, such as slow adaptation to novel threats, lack of proactive defense, high rates of false-positives, and incapacity to implement dynamic trust assessment in real-time AI-based fraud detection systems. To counter such limitations, this research proposes an Intellectual Zero Trust Security Framework that is based on Deep Reinforcement Learning (DRL) to mitigate threats in predictive mode in an attempt to detect fraud in AI-based systems. The suggested framework would incorporate a policy-optimization DRL agent with an active trust-evaluation mechanism that would monitor dynamically user behaviors, system interaction, and new anomaly patterns. Incorporating state-state modeling, reward-shaping, and incremental threat-scoring, the DRL-Zero Trust agent is able to learn the best security actions which include access restriction, adaptive authentication, and anomaly-suppression without any predefined rules. Evaluation of the proposed DRL-Zero Trust model on a benchmark dataset of fraud detection shows that it significantly outperforms classical classifiers, including K Nearest Neighbor (KNN), Random Forest, Logistic Regression (LR) and Support Vector Machine (SVM). The proposed model achieved 98.7% accuracy, 98.4% precision, 98.9% recall and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.995, and is better resistant to zero-day attacks and adversarial instances. These findings verify that the suggested framework presents a very flexible, intelligent, and proactive defense system that can protect the modern AI-driven financial systems.Keywords
This publication has 27 references indexed in Scilit:
- Clone node detection in static wireless sensor networks: A hybrid approachJournal of Network and Computer Applications, 2024
- Performance Analysis of an OpenFlow-Enabled Network with POX, Ryu, and ODL ControllersIETE Journal of Research, 2024
- Enhancing Zero Trust Security in Edge Computing Environments: Challenges and Solutions2024
- DDoS Attack Classification Leveraging Data Balancing and Hyperparameter Tuning Approach Using Ensemble Machine Learning with XAI2024
- AE-MLP: A Hybrid Deep Learning Approach for DDoS Detection and ClassificationIEEE Access, 2021
- DoS and DDoS attacks in Software Defined Networks: A survey of existing solutions and research challengesFuture Generation Computer Systems, 2021
- Sieve: A flow scheduling framework in SDN based data center networksComputer Communications, 2021
- An SDN-Assisted Defense Mechanism for the Shrew DDoS Attack in a Cloud Computing EnvironmentJournal of Network and Systems Management, 2021
- A Systematic Review of Quality of Services (QoS) in Software Defined Networking (SDN)Wireless Personal Communications, 2020
- Machine learning‐based IDS for software‐defined 5G networkIET Networks, 2018