Authors: Abhishek Ramprasad Kharde, Professor Bangar Abhishek Popat, Professor Bhosale Sachin Bajirao, Dr. A. A. Khatri
Abstract: Digital payments have made everyday life easier, but they have also given attackers a larger and more profitable surface to work on. Banks and payment processors now face two problems at the same time: stopping fraudulent transactions quickly, and protecting the personal data that fraud models need in order to learn. These two goals often pull in opposite directions, because stronger detection usually means collecting and sharing more sensitive data. This paper proposes CPDF, a five-layer framework that treats data protection and fraud detection as one design problem instead of two separate ones. The layers cover secure ingestion, privacy preservation through pseudonymisation, minimisation and differential privacy, hybrid detection that pairs an unsupervised Isolation Forest outlier score with a class-weighted gradient-boosted classifier, explainable alerts using SHAP, and continuous monitoring aligned with the NIST Cybersecurity Framework 2.0. We describe the architecture, the data flow, the evaluation protocol and the threats the design is meant to withstand. The work is a proposed design: it reports no experimental results, and the expected outcomes are stated as testable hypotheses for future implementation.
