Adaptive Few-Shot Fraud Detection: A Meta-Learning Approach

Credit card fraud detection remains challenging due to constantly evolving fraud patterns and severe class imbalance, where fraudulent transactions typically constitute less than 1% of total transactions. This paper introduces an adaptive few-shot fraud detection model based on meta-learning to rapidly adapt to emerging fraud tactics using minimal labeled data. Leveraging a prototypical network integrated within a model-agnostic meta-learning (MAML) framework, our approach achieves superior performance compared to traditional methods. Experimental results using a realistic dataset of 100,000 simulated credit card transactions demonstrate that our model attains an accuracy of 97.75%, precision of 89.30%, recall of 85.50%, and an F1-score of 87.36%, outperforming baseline methods such as Random Forest (accuracy 94.25%, F1-score 76.72%) and standard deep learning architectures. Despite slightly increased computational demands during training, the inference time of 2.2 ms per transaction confirms the realtime applicability of the model. This research provides valuable insights into effectively handling dynamic fraud detection scenarios and highlights promising future directions in metalearning and continual learning for enhanced adaptive fraud management systems.