
Research Article
Effective WiTech Identification Using Deep Transfer Learning with SNR as an Additional Feature
@INPROCEEDINGS{10.1007/978-3-030-97124-3_28, author={Sachin Nayak and Amitesh Singh Sisodia and Subrahamanya Swamy Peruru}, title={Effective WiTech Identification Using Deep Transfer Learning with SNR as an Additional Feature}, proceedings={Simulation Tools and Techniques. 13th EAI International Conference, SIMUtools 2021, Virtual Event, November 5-6, 2021, Proceedings}, proceedings_a={SIMUTOOLS}, year={2022}, month={3}, keywords={Deep learning Wireless technology identification Signal to noise ratio CNNs LSTMs FCNs}, doi={10.1007/978-3-030-97124-3_28} }
- Sachin Nayak
Amitesh Singh Sisodia
Subrahamanya Swamy Peruru
Year: 2022
Effective WiTech Identification Using Deep Transfer Learning with SNR as an Additional Feature
SIMUTOOLS
Springer
DOI: 10.1007/978-3-030-97124-3_28
Abstract
Efficient sensing of wireless technology is critical in today’s congested wireless ecosystem for effective utilization of limited resources. This paper presents a novel approach to wireless technology identification using deep transfer learning techniques, which are known to outperform conventional methods. A thorough study is required to understand the format of the data which is best suited for the task of technology identification, since wireless technology data is available in multiple formats like time-domain and frequency-domain representation. More importantly, which neural network architecture works best for each of these representations is to be studied. In this work, we show that fully connected neural networks and convolutional neural networks work best for classifying frequency-domain data and long short-term memory networks for time-domain data. Further, wireless signals with different signal-to-noise ratios (SNRs) may require a different strategy for efficient classification. In particular, a model that works well with signals having a high SNR may not perform well on signals that have low SNR. In this work, we study how the concept of transfer learning can be leveraged to design neural networks that work across different SNRs. In particular, we study questions like whether a neural network pre-trained on high-SNR data can improve the performance on low-SNR data? Our experimental results show that leveraging transfer learning can give additional gains of 4 to 20% in accuracy, depending on the SNR of the signal.