Wireless Interference Identification: A Bayesian-Multimodal Learning Approach

Abstract

Accurate interference identification is crucial for ensuring the security of wireless communications. Current deep learning-based methods achieve good performance but suffer from accuracy degradation due to the difference between the training and inference stages. Meanwhile, point-wise estimation lacks uncertainty quantification in the predictions. To tackle these issues, this letter proposes multimodal models for interference classification. First, both statistical and time–frequency (T-F) modalities are used to better represent the interference signals. Then, a multimodal model based on an attention mechanism is proposed to suppress irrelevant features at a low interference-to-noise ratio (INR), utilizing a transformer-based fusion scheme. Second, the Bayesian multimodal model is proposed for uncertainty measurement, therefore enhancing the robustness of the model. We finally evaluate the performance of the proposed model on the two constructed datasets. Experimental results show that the proposed Bayesian multimodal model reaches 99.83% accuracy when INR is greater than 0 dB. Moreover, it achieves a high classification confidence compared to existing models, demonstrating the effectiveness of the proposed Bayesian multimodal interference classification method.