Mitigating Bias and Data Poisoning in Large Language Model–Based Fraud Detection Pipelines

The rapid expansion of online transactions and changing adversarial tactics has complicated financial fraud detection due to the rapid growth of the digital transactions. Conventional machine learning and deep learning methods have disadvantages of being data-imbalanced, biased, and susceptible to data poisoning attacks. This paper presents a new set of guidelines on how to reduce bias and data poisoning in Large Language Model (LLM)-based fraud detection pipelines. The suggested method combines cost-sensitive learning and resampling strategies to cope with the issue of class imbalance, and anomaly-based filtering systems to identify and remove poisoned data. Moreover, semantic and contextual relationships in financial information are committed and implemented through the help of LLM-based feature extraction to detect a larger range of data. It includes a reinforcement learning component to allow adaptive learning and enhance the model robustness with time. It has been shown through experiment that the proposed model is better than the existing methods in accuracy, recall, bias reduction, and robust to adversarial manipulation. The framework is a credible and scalable means of detecting fraud in the contemporary financial platform.