Novel Devices Identification with Deep Clustering

The rapid growth in the number of wireless devices has made robust emitter identification approaches essential for several applications, including network security. Recent studies have explored methods for identifying unknown devices, but many of these rely on labeled data for training, which require significant labeling effort, or assume prior knowledge, such as the exact number of devices that need to be identified. In this work, we propose a new zero-shot self-supervised learning (ZS-SSL) approach for recognizing unknown transmitters that eliminates the need for prior knowledge about the number of devices. The proposed method achieves high classification performance, with an F1 score of 81%, and demonstrates remarkable robustness. It handles a varying number of clusters, up to 20 times more than the actual number of transmitters, and maintains good ZSL performance even when the training set is reduced to just two transmitters, achieving an F1 score of 80%.