Device authentication of wireless devices at the physical layer could augment security enforcement before fully decoding packets. At the upper layers of the stack, this is conventionally handled by cryptographic schemes. However, the associated computing overhead may make such regular approaches unsuitable for the emerging class of Internet of Things devices, which are typically resource-constrained and embedded in areas that make them difficult to retrieve and re-program. In contrast, radio frequency fingerprint identification (RFFI) exploits the unique hardware features as device identifiers at the physical layer. This article reviews both the state-of-the-art in engineered feature-based RFFI protocol design and advances in recent deep learning-based protocols, as well as a hybrid protocol that combines their advantages. Specifically, the hybrid approach leverages two methods: a more versatile distance-based classifier and an automatic feature extractor. This article also summarizes the goals of identification, verification and classification as applicable to RFFI, and how they can be achieved by the above protocols.