A High-Fidelity Partial Face Manipulation Dataset for Enhanced Deepfake Detection

Deepfake technology has already impacted the integrity of news and may grow to hugely destructive political and social force. The realistic and convincing nature of deepfakes poses a threat to the authenticity of information, alarming individuals and organizations. While many studies have explored the issue of deepfakes, the majority of them have focused on swapping entire faces rather than partially manipulating them, which can be more difficult to detect. In this paper, we introduce a high-fidelity partially manipulated face dataset, aiming to fill the gap in the existing deepfake research by providing a comprehensive benchmark for partially manipulated face detection. Our dataset includes a diverse set of partially manipulated faces which is generated from high-quality facial images. Our proposed alignment pipeline ensures that the partially manipulated faces may be realistically integrated into the original images, providing a more challenging evaluation environment for deepfake detection models. Both objective and subjective evaluations of our proposed dataset have shown promising results, indicating its potential to become a significant benchmark for partially manipulated face detection.