Deepfakes, i.e. realistic yet manipulated videos using AI techniques, are increasingly used to spread false narratives, influence public opinion, and undermine trust in authentic media. The term deepfake itself serves as an umbrella concept encompassing a wide range of techniques, focusing primarily on facial manipulations such as face-swapping, lip-syncing, and facial reenactment, each presenting unique technical challenges. This diversity complicates the development of robust detection models, especially those capable of generalising effectively across datasets and methodologies. The variations in deepfake techniques introduce inconsistencies in artifacts and motion dynamics, making it challenging for detection models to identify universal features that reliably distinguish real content from synthetic. When deployed “in the wild”, challenges are even greater, as videos often include compression artifacts, diverse lighting conditions, and adversarial manipulations, all of which make detection significantly harder. This chapter discusses these challenges, presents state of the art approaches and explores how various factors impact the generalisation of deepfake detection models. Finally, we provide a detailed description of available datasets and discuss their limitations in assessing model performance and training robust systems capable of handling content from arbitrary sources.