Open-Set Interference Signal Recognition Using Boundary Samples: A Hybrid Approach

The applications of neural network based classifier to the real world tasks have to deal with the problem of open-set recognition (OSR), where during the testing phase there may appear unknown classes not seen in the training phase. This paper presents a novel method for OSR problem and its potential application on wireless interference signal recognition. The proposed method uses more precise and stable boundary samples to imitate unknown classes. To be sepecific, the proposed method modifies and combines the methods of intra-class splitting (ICS) and adversarial samples generation to construct precise boundary samples. Numerical experiments are conducted with three commonly used image datasets and one wireless interference signal dataset collected at 2.4GHz ISM band by LimeSDR, and the analytical results demonstrate the effectiveness of the proposed OSR method.