Recent advances in Generative AI (GenAI) make them able to generate high-quality realistic content (i.e., images and videos). It is difficult for humans to discriminate between real and fake content. As a result, GenAI models could be used for malicious purposes, such as impersonating or even misleading political opinions. These activities are referred to as “Deepfake” (DF).  In the case that humans cannot reliably recognize DF content, it is imperative to develop a DF detection method. However, current works in DF detection can not generalize to samples generated by unknown generators or unknown post-processing techniques. In this project, I aim to develop a generalized DF detection system that can generalize to out-of-distribution (OOD) DFs without depending on only supervision signals.