International Journal of Science, Engineering and TechnologyAuthors: Dr. Pankaj Malik, Mannat Bhatia, Abhishek Kumar Tiwari, Hrishit Nagar, Atharva Shrivastava
Abstract: Online transaction systems are increasingly exposed to zero-day fraud attacks, where novel and rapidly evolving fraud patterns bypass conventional detection models trained on historical data. Existing machine learning–based fraud detection approaches struggle to adapt due to their reliance on large labeled datasets and static training paradigms. This paper presents a meta-learning–based adaptive fraud defense framework that enables rapid detection of previously unseen fraud patterns using a limited number of labeled samples. The proposed approach leverages Model-Agnostic Meta-Learning (MAML) to learn transferable representations across diverse fraud tasks and supports few-shot adaptation in real-time transaction environments. Experiments conducted on the IEEE-CIS Fraud Detection and PaySim datasets, with zero-day fraud scenarios simulated through task-wise data partitioning and concept drift injection, demonstrate that the proposed model outperforms state-of-the-art baselines. Specifically, the meta-learning framework achieves an average F1-score improvement of 14.6% and an AUC-ROC increase of 11.2% over deep neural network and XGBoost models under zero-day conditions. Furthermore, the adaptation time is reduced by approximately 3.1×, enabling effective fraud detection within a minimal number of gradient updates. These results confirm that meta-learning provides a robust and scalable solution for rapid defense against zero-day fraud attacks, significantly enhancing transaction risk management in dynamic financial systems.
