Robust signal classification using Siamese networks

Abstract

We propose a noise-robust signal classification approach using siamese convolutional neural networks (CNNs), which employ a linked parallel structure to rank similarity between inputs. Siamese networks have powerful capabilities that include effective learning with few samples and noisy inputs. This paper focuses on the advantages that siamese CNNs exhibit for classification of quite similar wireless signal emitters across signal-to-noise ratio (SNR) and dataset size. Without any a priori information, candidate siamese and baseline CNNs were trained on compressed spectrogram images to distinguish modulated signal pulses with randomized symbols and identical signal parameters, save for slight frequency offsets commonly exhibited in commercial RF emitter reference oscillator uncertainty distributions. Compared with baseline CNN approaches the proposed methods demonstrate improved classification performance under poor SNR. Moreover, this advantage holds the potential for superior, low-SNR, semi-supervised classification using embeddings from within the networks.