Deepfake Detection Performance Evaluation and Enhancement Through Parameter Optimization

Deepfake technology has become a subject of concern due to its potential for spreading misinformation and facilitating deceptive activities. To address these issues, various deepfake detection approaches have been developed with similar training paradigms. Then a natural question is which parameters are critical to achieving better detection performance. This study aims to evaluate and optimize the performance of existing deepfake detection systems by analyzing key parameters in the training paradigm. Specifically, we systematically analyze four crucial factors: image cropping, sampling rate, data augmentation, and transfer learning. The impact of different image scopes, such as utilizing the entire image or only the cropped face region, is investigated. We also explore how varying the sampling rate and employing data augmentation techniques can enhance the diversity of the training dataset. Additionally, transfer learning with pre-trained models is leveraged to improve detection accuracy. Through comprehensive experiments and evaluations of several popular and state-of-the-art detection methods, optimal configurations within each factor are identified, providing valuable insights to enhance the efficiency and effectiveness of deepfake detection systems. Given the widespread use and potential negative consequences of deepfake technology, reliable detection systems are crucial in combatting the harmful effects of manipulated media.