The rapid proliferation of digital payment infrastructure has established credit card transactions as the backbone of the modern global economy, concurrently exposing financial networks to sophisticated fraudulent activities. The automated detection of such anomalies presents a significant algorithmic challenge due to extreme class imbalance, as fraudulent instances typically represent less than 0.5% of the overall transaction volume. This research proposes a robust, machine learning-based classification architecture utilizing a highly imbalanced dataset of 284,807 transactions, where the minority fraud class constitutes merely 0.17% of the data. To neutralize the statistical bias introduced by this skew, rigorous data preprocessing techniques including Z-score standardization and stratified splitting were implemented. The Synthetic Minority Over-sampling Technique (SMOTE) was deployed strictly within the training environment to synthetically balance the class distributions and prevent algorithmic convergence toward majority-class predictions. A comparative analysis was conducted evaluating a linear Logistic Regression classifier against a non-linear Random Forest ensemble. Empirical analysis demonstrates that while the linear model achieved high theoretical class separation, the Random Forest ensemble delivered superior operational performance. By optimizing the precision-recall trade-off, achieving a precision of 0.84 and a recall of 0.83, the ensemble model successfully minimized false negative rates without inflating false positive rates, proving its viability for real-world deployment in institutional financial security systems.