In this work, we propose an efficient pipeline for deepfake detection, combining deep learning-based feature extraction with dimensionality reduction techniques. The method begins with feature extraction using a pre-trained ResNet50 model, which captures high-level representations from input images. Given the high dimensionality of the extracted features, we apply Uniform Manifold Approximation and Projection (UMAP) to reduce the feature space while preserving the underlying data structure. This reduction allows for improved computational efficiency without sacrificing critical information for classification tasks. The reduced features are then passed to multiple classifiers, including Random Forest and KNN, for detection. Our experiments demonstrate that the proposed method excels particularly in handling low-quality, compressed data, as shown by its state-of-the-art performance on the C23 and C40 versions of the FaceForensics++ dataset. In particular, the Random Forest classifier achieved an accuracy of 97,3 and an Area Under Curve (AUC) of 99,1, while KNN achieved an accuracy of 96,5 and an AUC of 99,1. The combination of ResNet50 for feature extraction and UMAP for dimensionality reduction proves to be an effective strategy for deepfake detection across various data qualities.