Applying Distributed Learning of Deep Neural Networks to Improve Their Classification Accuracy on Radio-Frequency Datasets
This thesis aims to improve on the current classification capabilities of deep neural networks on two types of radio-frequency data: radar and OFDM packets. In radar, applying neu- ral networks to Automatic Target Recognition problems is a well-developed field, especially using the MSTAR database. However, existing state-of-the-art methods require precise pre- conditioning of radar data and are unsuited to applications with a large number of radar target classes. Therefore, we asked whether distributed learning can increase the generaliz- ability and scalability of neural networks in these tasks. To test this, we applied distributed learning via Multi-Stage Training and a new network architecture, the Convolutional Multi- Stage Network, to provide a scalable, generalized treatment of radar data for more practical applications. This method was shown to outperform traditional neural network architectures on a new radar dataset. A similar approach was applied to the OFDM data with the goal of identifying specific radio-frequency transmitters for network security purposes. The task of identifying OFDM packet transmitters has previously been performed successfully, though with precise data collection methods. Data collection methods on a live network will likely include imperfect recording times, so we sought to improve network robustness to time- shifted OFDM packets. It was shown that the Convolutional Multi-Stage Network improved robustness to time-shifting of the radio-frequency data over the Multi-Stage Network, which was the previous-best method. Simple preconditioning of the data using variations of the dis- crete wavelet transform further improved robustness to time-shifting of the radio-frequency data using both network architectures. These results are significant, as they provide a new avenue for applying neural networks to radio-frequency in difficult, real-world applications.