Autodetect: an Actor-critic Reinforcement Learning Framework for Financial Fraud Detection
Year: 2025
Type: article
Abstract: Financial fraud detection systems face significant challenges in adapting to evolving fraudulent behaviors while maintaining optimal balance between detection accuracy and operational efficiency in dynamic financial environments. Traditional supervised learning approaches struggle with the sequential decision-making nature of fraud detection, limited labeled fraud data availability, and the need for real-time adaptation to emerging fraud patterns that continuously evolve to circumvent existing detection mechanisms. The challenge lies in developing intelligent systems that can learn optimal detection strategies through interaction with financial transaction environments while balancing exploration of new fraud patterns with exploitation of known fraud indicators. This study proposes AutoDetect, a novel Actor-Critic Reinforcement Learning (RL) framework that formulates fraud detection as a sequential decision-making problem where an intelligent agent learns optimal detection policies through continuous interaction with transaction data streams. The framework employs actor-critic architecture where the actor network generates detection decisions and the critic network evaluates the quality of these decisions based on fraud detection rewards and penalty structures. The RL approach enables autonomous learning of detection strategies that maximize long-term fraud detection effectiveness while minimizing false positive rates through dynamic policy optimization based on environmental feedback. Experimental evaluation using large-scale financial transaction datasets demonstrates that AutoDetect achieves 53% improvement in fraud detection accuracy compared to traditional supervised learning approaches. The framework results in 46% better adaptation to novel fraud patterns and 42% reduction in false positive rates while maintaining real-time processing capabilities suitable for high-throughput financial transaction environments. AutoDetect successfully combines reinforcement learning with fraud detection domain knowledge to provide 38% better interpretability of detection decisions while supporting autonomous improvement through continuous learning from transaction feedback.
Source: Social science and management.
DOI: 10.61784/ssm3059
Authors: Emily Yip, Thomas Lau, Patrick Ho
Institution: Hong Kong Polytechnic University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7077061462
FWCI:
Cites:
Cited by:
Related to: