With recent advancements in artificial intelligence, generative AI technologies such as Text-to-Speech and Voice Conversion have become highly sophisticated, making it increasingly difficult to distinguish between deepfake audio and genuine human voices. As this technology becomes more accessible, its malicious use in crimes like social engineering is increasing, highlighting the urgent need for robust detection technologies. This study presents a comprehensive performance analysis of various deepfake audio detection models. We utilize large-scale Self-Supervised Learning models–specifically XLSR, HuBERT, and WavLM as feature extractors. These are paired with various classifier architectures, including AASIST, Conformer, and Transformer, which have demonstrated strong performance in audio-related tasks. Through experiments on the ASVspoof 2019 LA and ASVspoof 5 datasets, we demonstrate that SSL-based models significantly outperform traditional detection systems. On the ASVspoof 2019 dataset, the WavLM-Transformer combination achieved the best performance with an Equal Error Rate of 2.22%. For the more challenging ASVspoof 5 dataset, the XLSR-AASIST combination proved most robust, achieving an EER of 8.41%. Our experiments confirm that overall performance is not determined by the standalone superiority of a single component, but by the optimal pairing of an extractor and classifier, which varies depending on dataset characteristics. These findings provide a clear direction for developing more advanced and robust deepfake audio detection models.