Meta-Learning for Wireless Interference Identification
Ali Owfi, Fatemeh Afghah, Jonathan Ashdown · 2023
Deep learning-based (DL-based) models have shown to be powerful tools for wireless interference identification (WII). However, one of the key concerns toward using these models in practical systems is that they perform poorly when they are encountered with signals coming from new sources not previously observed during the training phase. In a real-world communication system, the interference identifier will frequently face new unknown signals due to the existence of many wireless transmitters. This renders the conventional DL-based models impractical as a WII tool unless they go through a new training phase. Retraining the model is not only inefficient, but it can also be not feasible in some cases (e.g., at end-user devices) as the training phase consumes time and resources and requires large amounts of data. We present a new approach for data-driven WII systems using meta- learning to address the lack of adaptability in conventional DL-based models to new (not previously seen) signals. We show that by using meta-learning, we are able to identify signals coming from not previously observed technologies and frequencies using just a handful of new samples, a task that is not generally possible with conventional DL models. Finally, we analyze and compare the performance of the presented meta-learning model in multiple different settings using raw I/Q samples and Fast Fourier Transform of I/Q samples. Based on our experiments, we show that the proposed meta-learning scheme outperforms the conventional deep learning models for WII when there are just a few samples available for training1.