Credit card fraud is one of the issues that many banks and organizations are concerned about and researching. Building systems to monitor and detect fraud is extremely necessary and urgent for banks, to minimize the loss of money without clear reasons leading to losses, and also to affect the reputation and credibility of the bank or credit organization. Automating credit card fraud detection is a perfect way to meet user needs and provide security on a large scale. Based on that, this paper has researched, experimented, and proposed a new model that combines two algorithms, Autoencoder and Isolation Forest, on two sample datasets simulating bank transactions. Using deep learning research methods, the proposed model achieved nearly absolute efficiency of up to 99.77%, significantly improving accuracy compared to previous studies.