Radio Frequency Classification and Anomaly Detection using Convolutional Neural Networks

Radio frequency (RF) spectrum is a limited and critical resource for radar systems. As RF devices become smaller, cheaper and increasingly deployed, they present significant challenges with spectrum sharing and radar interference mitigation. This demands the need for reliable RF sensing and classification techniques for cognitive radars that can dynamically adapt to avoid interference. Our research investigates convolutional neural networks (CNN) trained on waveform images to classify RF spectrum modulations, and two techniques that uses the activations of the last hidden layer of the CNNs to detect anomalies. The CNNs used were AlexNet, GoogleNet, Inception V3, and ResNet50. Our first approach shows that CNNs and Principle Component Analysis (PCA) are highly accurate for classification and anomaly detection. Our second approach shows that appending N-nodes with randomly assigned weights and biases to the last hidden layer may give reasonable performance provided that the weights and biases are carefully selected.