Due to the growing usage of DeepFake technology to produce fake photos and videos, the problem of DeepFake identification has become a serious concern in recent years. With the use of this technology, fake material that is difficult to tell apart from real content may be produced. Therefore, the development of accurate DeepFake detection algorithms is essential to identify and prevent the spread of manipulated content. However, current publicly available DeepFake detection datasets suffer from a lack of diversity, with only a few actors appearing in multiple videos. This results in an oversampled training dataset, leading to model overfitting and inadequate performance when tested on new data. The overfitting problem can cause deep neural networks to focus more on facial features than on specific traits of DeepFake content, thereby reducing the model's ability to generalize. To address this issue and broaden the training dataset's variety, we suggest applying data augmentation methods like Face Cutout and Random Erase. The Face Cutout technique randomly removes a rectangular area of an image containing the face, while the Random Erase technique randomly removes a rectangular area of the image. These techniques introduce variations in the images, making it more challenging for the model to overfit and focus on specific traits of DeepFake content. We evaluated our approach using several models and achieved significant accuracy improvements using the EfficientnetV2B0 model and Random Erase augmentation.