Feature Learning for Enhanced Security in the Internet of Things

Identifying Internet of Things (IoT) devices by their Radio Frequency (RF) fingerprint has important security implications. As the number of connected devices grows, current authentication mechanisms are becoming more susceptible to device spoofing attacks. To combat this, we exploit hardware imperfections in the RF transmit chain to extract device-specific features that uniquely identify an emitter, providing an additional layer of security. This is accomplished with a complex-valued Variational Autoencoder that has a Gaussian Mixture (GMVAE) prior on the latent variables' marginal distribution. By exploiting sequential information in the RF time-series data, we achieve processing gain by integrating multiple latent-space representations from a single device. We test and analyze the proposed approach on real WiFi data and obtain excellent classification results. We also test the proposed model on an Out-of-Distribution (OOD) detection task.