Today's image generation technology can generate high-quality face images, and it isn't easy to recognize the authenticity of the generated images through human eyes. This study aims to improve deepfake detection, a face swapping forgery, by absorbing the advantages of deep learning technologies. This study generates a unified and enhanced data set from multiple sources using spatial enhancement technology to solve the problem of poor detection performance on cross-data sets. Taking the advantages of Inception and ResNet networks, new deepfake detection architecture composed of 20 network layers is proposed as the deepfake detection model. To further improve the proposed model, hyperparameter values are optimized. The experiment result shows that the proposed network significantly enhanced over the mainstream methods, such as ResNeXt50, ResNet101, XceptionNet, and VGG19, in terms of accuracy, loss value, AUC, numbers of parameters, and FLOPs. Overall, the methods introduced in this study can help to expand the data set, better detect deepfake contents, and effectively optimize network models.