Machine Learning (ML) is fundamentally transforming banking fraud detection, offering unprecedented opportunities for high-volume scalability and the identification of complex, multi channel criminal networks, such as Synthetic Identity Fraud. However, this shift introduces significant operational vulnerabilities, notably the economic burden of high false positive rates, which can account for 19% of the total cost of fraud , and the risk of targeted adversarial evasion attacks against predictive models. This paper synthesizes recent findings to explore the dual-edged nature of ML in financial security, examining how deep learning architectures (e.g., CNN-LSTM) demonstrate superior performance while highlighting systemic challenges like data silos and regulatory conflicts (GDPR vs. AML). Drawing on strategic insights, it discusses the imperative for risk-aligned metric optimization, specifically recommending the adoption of the weighted F-beta score to prioritize recall , alongside technological safeguards such as adversarial training to enhance model robustness. Findings underscore the critical need for unified data architectures and collaborative industry benchmarking to build a resilient defense against evolving financial crime.