Credit cards, mobile wallets, and other electronic payment methods are gaining popularity. Online transactions are increasingly the norm - global fraud increases as electronic payments increase. As credit cards and online shopping become increasingly popular, fraud has skyrocketed. Fraud detection and prevention are being prioritized due to the global economy. The trillion-dollar fraud business threatens financial loss and financial institution trust. Financial fraud detection could avert trillions in losses. Thus, detecting fraud is one of the most challenging real-world problems. Unbalanced datasets with more "normal" samples than fraud cases impair fraud detection. Rapid fraud changes complicate training cutting-edge machine learning classifiers. If there were more labeled datasets in real-world settings, fraud detection solutions could learn from the events in the training dataset to identify fraudulent patterns. Businesses need a fraud detection solution that can be trained on unlabeled financial transaction datasets widely available in financial transaction systems to detect fraudulent occurrences accurately. This paper proposes a fraud detection approach based on a memory compression methodology (FDMCM) machine learning approach to enhance detection. We suggest using a machine learning network to identify fraudulent transactions and a novel nonlinear embedded machine learning base autoencoding layered technique to correct dataset imbalances. The proposed model has 93% success with an 80:20 training-validation dataset accuracy ratio.