[Paper Review] Deep Learning for Interference Identification: Band, Training SNR, and Sample Selection

Xiwen Zhang, Tolunay Seyfi+4 more|arXiv (Cornell University)|May 16, 2019

Wireless Signal Modulation Classification8 references4 citations

TL;DR

This paper proposes a fast, accurate deep learning approach for wireless interference identification using 10 MHz of the 2.4 GHz ISM band, achieving ~89.5% classification accuracy across four deep neural networks (CNN, ResNet, CLDNN, LSTM). By applying band selection (4 MHz), SNR selection (-2 dB), and PCA/sub-Nyquist sampling, training time is reduced by up to 30x with minimal accuracy loss, enabling practical deployment in real-time cognitive radio systems.

ABSTRACT

We study the problem of interference source identification, through the lens of recognizing one of 15 different channels that belong to 3 different wireless technologies: Bluetooth, Zigbee, and WiFi. We employ deep learning algorithms trained on received samples taken from a 10 MHz band in the 2.4 GHz ISM Band. We obtain a classification accuracy of around 89.5% using any of four different deep neural network architectures: CNN, ResNet, CLDNN, and LSTM, which demonstrate the generality of the effectiveness of deep learning at the considered task. Interestingly, our proposed CNN architecture requires approximately 60% of the training time required by the state of the art while achieving slightly larger classification accuracy. We then focus on the CNN architecture and further optimize its training time while incurring minimal loss in classification accuracy using three different approaches: 1- Band Selection, where we only use samples belonging to the lower and uppermost 2 MHz bands, 2- SNR Selection, where we only use training samples belonging to a single SNR value, and 3- Sample Selection, where we try various sub-Nyquist sampling methods to select the subset of samples most relevant to the classification task. Our results confirm the feasibility of fast deep learning for wireless interference identification, by showing that the training time can be reduced by as much as 30x with minimal loss in accuracy.

Motivation & Objective

• To reduce training time for deep learning-based wireless interference identification while maintaining high classification accuracy.
• To evaluate the impact of band selection, SNR selection, and sample selection on model training efficiency and performance.
• To optimize a CNN architecture for faster training compared to state-of-the-art methods on the same dataset.
• To assess the robustness of classification accuracy under dimensionality reduction and sub-Nyquist sampling techniques.
• To enable practical deployment of deep learning in real-time cognitive radio systems through training time acceleration.

Proposed method

• Trained four deep neural network architectures—CNN, ResNet, CLDNN, and LSTM—on 225,225 labeled samples from 15 wireless channels in the 2.4 GHz ISM band.
• Used a 10 MHz bandwidth subset centered on 2422–2432 MHz, with signals from Bluetooth, Zigbee, and WiFi at SNR levels from -20 dB to 20 dB in 2 dB steps.
• Applied band selection by restricting training to the lower and upper 2 MHz bands, reducing the effective bandwidth to 4 MHz.
• Implemented SNR selection by training exclusively on data from a single SNR value (e.g., -2 dB), minimizing data diversity and training time.
• Employed PCA and sub-Nyquist sampling (random, uniform, high-magnitude rank) to reduce input dimensionality and sample count while preserving classification-relevant features.
• Compared training time per epoch and total training time across configurations, with accuracy evaluated on a held-out test set across SNR values.

Experimental results

Research questions

• RQ1 Can deep learning models achieve high interference classification accuracy with significantly reduced training time using selective data sampling?
• RQ2 How does restricting the frequency band to 4 MHz affect classification accuracy and training efficiency compared to full 10 MHz bandwidth?
• RQ3 What is the optimal SNR value for training that balances accuracy and training speed across diverse test SNR conditions?
• RQ4 To what extent can PCA and sub-Nyquist sampling reduce model training time with minimal accuracy degradation?
• RQ5 Can a combination of band selection, SNR selection, and dimensionality reduction yield a 30x reduction in training time while maintaining high accuracy for high-SNR signals?

Key findings

• The proposed CNN architecture achieved ~89.5% average test accuracy across 15 channels, using only 40% of the training time required by the state-of-the-art CNN in [8].
• Training with only a single SNR value (-2 dB) reduced total training time by 92.3% compared to using all SNR values, while maintaining >90% accuracy for high-SNR test data.
• Using only the 4 MHz outer bands (lower and upper 2 MHz) reduced training time by 74.8% and maintained 86.14% accuracy, with a 90.9% reduction in total training time.
• Applying PCA with a 16x compression rate on the 10 MHz dataset resulted in only 5.3% accuracy drop (from 89.62% to 84.11%) at high SNR, demonstrating robustness to dimensionality reduction.
• Random subsampling at 1/4 rate preserved ~83% accuracy at high SNR, while PCA at 1/8 rate reduced training time per epoch to 1.33s (from 16.37s) with only 2.5% accuracy loss.
• Combining 4 MHz band selection with 8x PCA compression reduced training time by up to 30x compared to the baseline, with negligible accuracy loss for SNR > -4 dB.