Low complexity wireless interference identification in antagonistic environments based on multi-task learning

As the fundamental premise of anti-interference communication, wireless interference identification (WII) has garnered extensive research and yielded substantial results, especially for deep learning (DL)-enabled WII. However, existing studies are conducted typically under the closed-set assumption, whereby the type of interference is known. This aspect poses a challenge in identifying the unknown interference signals under the d-set assumption. To tackle this issue, this paper proposes a multi-task learning-enabled WII (MTL-WII) algorithm in antagonistic environments. Firstly, we generate the semantic feature space of known classes, forming the mapping relationship from the original signal data to the semantic feature space while calculating the semantic center vectors of each known class through multi-task learning. Secondly, we obtain the semantic feature space of the test set incorporating interference signals of known and unknown classes through the established mapping relationships. Subsequently, the similarity between the semantic feature vectors of the interference signals under test and the semantic center vectors of the interference signals within each class gets calculated using the clustering method. Finally, the attributed class is determined by comparing the similarity against a threshold. Nevertheless, the computational complexity of the proposed method limits its applicability across various scenarios. To this end, this paper further proposes a binarized MTL-WII (BMTL-WII) algorithm. By binarizing the semantic spatial generative network in MTL-WII, both the weights and activation of the semantic spatial generative network replace the 32-bit floating-point numbers with 1-bit fixed-point numbers, which can save memory and improve the speed of the operations. Experimental results show that the MTL-WII model achieves an average identification accuracy of 95.4%, with an 89.3% accuracy for unknown interference signals. Compared to the MTL-WII algorithm, the BMTL-WII algorithm reduces the number of floating-point operations by 64.2%, and the amount of memory access is reduced by 88%, at the cost of identification accuracy decrease by only 1.8% of the binarized part of the network structure.