WIR-Transformer: Using Transformers for Wireless Interference Recognition

Abstract

Wireless interference recognition (WIR) is one of the most indispensable technologies in non-cooperative communication systems. Recently, compared with convolutional networks, Transformers enjoy the ability of extracting global features and have achieved striking performance for WIR. However, the self-attention module in Transformers brings huge computational overhead that hinders deployment on resource-constrained devices. In this letter, we aim to alleviate the problem and propose WIR-Transformer. Specifically, WIR-Transformer introduces the division of regions and independently calculates self-attention in each region, which effectively reduces the complexity. To overcome the bottleneck of information blocking between regions, we propose an information exchange module (IEM). To further reduce the complexity, we introduce a novel patch aggregation module (PAM) to reduce the number of patches and fuse the local information. Experiments demonstrate that the proposed WIR-Transformer achieves higher accuracy as compared to conventional methods for WIR.