Information Reliability Service - a novel approach to fraud detection by Giovanni Degiorgi

Over the past years, criminal activities on the Swiss Post portal have amounted to approximately 5 million Swiss Francs per year . Criminal organizations and their activities have become increasingly sophisticated and organized over time, with cybercrime remaining a significant concern for online retailers and services. In this presentation we address the challenges of fraud detection and imbalanced datasets, outlining the procedures implemented to manage data access at an enterprise level within Swiss Post. We present a machine learning approach to detect potential fraudulent addresses, preventing the activation of an account when risk is identified. Our innovative strategy employs a contextual Multi-Armed Bandit (MAB) model, trained on shipment data and expert fraud evidence, to take actions that maximize rewards by minimizing potential fraud cases. This novel approach, leveraging artificial intelligence and modern reinforcement learning techniques, is fully hosted on a cloud platform, allowing for on-demand scaling based on specific time requirements. We explore innovative machine learning approaches, including MLOps and our contextual MAB-based strategy, discussing their implementation and benefits. The empirical results, based on real-world data from Swiss Post logistic services, demonstrate that we can successfully identify a portion of the fraudulent addresses with the ability to improve our algorithm over time through expert feedback. These promising outcomes pave the way for future deployment in Swiss Post's production environment. We delve into the integration of our solution within a large-scale enterprise setting, emphasizing the role of Continuous Integration and Continuous Deployment (CI/CD) pipelines in maintaining and improving the system.