To safeguard the financial security of both credit card holders and issuing institutions, this paper proposes an anomaly detection method for credit card transactions based on a multi-dimensionally optimized autoencoder. The data is preprocessed through standardization and stratified sampling to ensure distribution consistency, while isolation forest is employed to filter pure normal samples for model training. To address the limitations of the original autoencoder, optimizations are introduced across several dimensions, including the network architecture, activation functions, and training strategy. Experimental results demonstrate that the proposed model outperforms conventional methods in detection performance, exhibiting enhanced practicality and adaptability for real-world applications.