<![CDATA[Deepfake technology has emerged as a significant challenge to digital security because it can generate highly realistic manipulated facial images and videos. The rapid spread of deepfake content has increased the risks of identity fraud, misinformation, privacy violations, and various forms of cybercrime. This study proposes an Android-based human face deepfake detection system using the YOLOv8 algorithm. The dataset consisted of authentic facial images collected from Universitas Malikussaleh students and deepfake facial images generated using artificial intelligence techniques. The research methodology included data collection, image preprocessing, annotation using Roboflow, YOLOv8 model training, model evaluation, TensorFlow Lite (TFLite) conversion, and Android application development. Experimental results demonstrated that the proposed model achieved 95% precision and 95% recall in detecting real and manipulated facial images from both images and videos, indicating reliable detection performance. Nevertheless, several limitations remain. The dataset does not fully represent real-world facial variations, including differences in ethnicity, illumination, facial expressions, head poses, and occlusion caused by masks, glasses, or other objects covering facial regions. These limitations may reduce the model's generalization capability when deployed in real-world environments outside the testing dataset. Furthermore, the deepfake dataset only includes several manipulation techniques and has not been evaluated using more recent deepfake generation methods, such as diffusion model-based face swapping or other advanced generative approaches. Consequently, the model's performance may decrease when encountering manipulation techniques that were not included during training. In addition, the evaluation has not comprehensively considered challenging imaging conditions, such as motion blur, image noise, low-bitrate video compression, and quality variations introduced by different mobile device cameras, which may affect the robustness of the proposed deepfake detection system in practical applications.]]>
Copyrights © 2026
