Improved Deepfake Classification Based on Feature Space Selection for Ensemble Models

Abstract

This paper proposes an improved approach to deepfake classification by integrating a multi-level feature representation with a systematic, classifier-specific feature space optimization. The paper represents audio signals through three-level model: tabular spectral features, spectrogram-based images, and temporal sequences. To maximize the effectiveness of base models, a Backward Feature Elimination procedure is implemented to adapt and optimize feature subsets for each base classifier, that works with tabular data. These optimized models are then integrated into a stacking ensemble architecture utilizing various aggregation strategies, such as hard voting, soft voting, and soft voting with Gompertz fuzzy ranking. Experimental results using the Fake-or-Real dataset demonstrate that the BFE procedure improves base classifier performance by up to 5 percent. The best-performing ensemble, consisting of K-nearest neighbors, Convolutional Neural Network and XGBoost with soft voting, achieved an Accuracy and F1-score of 98 percent, significantly outperforming the base classifiers (3.6-3.8 percent). The study confirms that combining heterogeneous classifiers with optimized feature spaces enhances the classification quality of audio deepfake detection.