DeepSense: Fast Wideband Spectrum Sensing Through Real-Time In-the-Loop Deep Learning

Spectrum sharing will be a key technology to tackle spectrum scarcity in the sub-6 GHz bands. To fairly access the shared bandwidth, wireless users will necessarily need to quickly sense large portions of spectrum and opportunistically access unutilized bands. The key unaddressed challenges of spectrum sensing are that (i) it has to be performed with extremely low latency over large bandwidths to detect tiny spectrum holes and to guarantee strict real-time digital signal processing (DSP) constraints; (ii) its underlying algorithms need to be extremely accurate, and flexible enough to work with different wireless bands and protocols to find application in real-world settings. To the best of our knowledge, the literature lacks spectrum sensing techniques able to accomplish both requirements. In this paper, we propose DeepSense, a software/hardware framework for real-time wideband spectrum sensing that relies on real-time deep learning tightly integrated into the transceiver's baseband processing logic to detect and exploit unutilized spectrum bands. DeepSense uses a convolutional neural network (CNN) implemented in the wireless platform's hardware fabric to analyze a small portion of the unprocessed baseband waveform to automatically extract the maximum amount of information with the least amount of I/Q samples. We extensively validate the accuracy, latency and generality performance of DeepSense with (i) a 400 GB dataset containing hundreds of thousands of WiFi transmissions collected "in the wild" with different Signal-to-Noise-Ratio (SNR) conditions and over different days; (ii) a dataset of transmissions collected using our own software-defined radio testbed; and (iii) a synthetic dataset of LTE transmissions under controlled SNR conditions. We also measure the real-time latency of the CNNs trained on the three datasets with an FPGA implementation, and compare our approach with a fixed energy threshold mechanism. Results show that our learning-based approach can deliver a precision and recall of 98% and 97% respectively and a latency as low as 0.61ms. For reproducibility and benchmarking purposes, we pledge to share the code and the datasets used in this paper to the community.