A Class Incremental Learning Method With Forward-Compatible and Covariance-Aware for Specific Emitter Identification

Specific Emitter Identification (SEI) is essential for IoT security. Due to the continuous emergence of new communication devices in the real world, SEI needs to cope with an increasing number of transmitter categories. A trained recognition model needs to possess the capability to continuously learn new devices. This paper proposes a novel class incremental learning method based on forward compatibility and covariance awareness, named FCCA. Specifically, this paper devises a virtual signal class generation approach and an integrated loss function to expand the feature space for incremental categories while preserving valid feature representations. During the incremental phase, FCCA uses a frozen feature extractor to obtain category feature embeddings and models feature covariance relationships, helping the classifier better differentiate between categories. Experimental results on benchmark datasets demonstrate that FCCA outperforms other methods. It also demonstrates excellent performance on few-shot class incremental problems.