Improving Financial Fraud Detection with Quantum Reinforcement Learning

The growing use of online banking and digital payment systems has increased the risk of fraudulent activities. For this reason, identifying suspicious transactions is an essential task that should continuously evolve with technological advances and new forms of criminal activity. In recent years, fraud detection has been commonly implemented by exploiting Machine Learning (ML) models, which are able to detect irregular patterns and identify fraudulent transactions. However, classical ML performance is not always satisfactory, especially in strongly unbalanced classification tasks like fraud identification. To overcome these limits, Quantum Machine Learning (QML) has recently gained researchers’ interest. Indeed, these systems, working intrinsically in a higher-dimensional computation space thanks to superposition and entanglement, can implement more complex models than the classical counterparts, potentially providing a significant advantage in complex pattern recognition. Among the various QML paradigms, Quantum Reinforcement Learning (QRL) is an emerging field that combines the ability of classical Reinforcement Learning (RL) to solve sequential decision-making problems with principles of Quantum Computing (QC) to improve speed and efficiency. This thesis specifically focuses on the application of QRL to a fraud detection task based on bank transaction data. The aim of this work is to provide a performance comparison between classical and quantum-based reinforcement learning. The study first shows how RL achieves higher classification performance than traditional supervised learning methods, thanks to its sequential decision-making nature and the use of a reward function. Additionally, by implementing two distinct algorithms, Q-learning and REINFORCE, this work compares two different approaches to RL. Then, the thesis studies how QC principles can further improve these capabilities. For each RL algorithm, a quantum version was created by substituting the classical neural network with its quantum counterpart implemented through a Variational Quantum Circuit (VQC). The reinforcement learning environment was implemented using Gymnasium library, which allows to create the agent–environment interaction loop and the reward function. Quantum components were implemented using PennyLane library, which enables integrating VQCs within the PyTorch models to create hybrid quantum–classical architectures. Experimental results show that QRL achieves comparable or better performance than classical RL approaches with a limited number of trainable parameters (84 in Q-learning and 150 in REINFORCE, instead of the 5938 required by both classical architectures), with F1 score improvements of 0.39% in Q-learning and 36.7% in REINFORCE. Due to the limitations of employing ideal quantum simulators, which are both computationally expensive and time-consuming, it has not been feasible to perform large-scale experiments. Future work should focus on optimizing training efficiency and exploring more scalable quantum-classical integration strategies.