Unsupervised Wireless Spectrum Anomaly Detection With Interpretable Features

Sreeraj Rajendran, Wannes Meert, Vincent Lenders, Sofie Pollin

TL;DR

This paper introduces SAIFE, a novel unsupervised deep learning model for detecting and localizing anomalies in wireless spectrum data. SAIFE leverages an Adversarial Autoencoder (AAE) to achieve high accuracy in anomaly detection and signal classification, even with limited labeled data, while also offering valuable interpretability and compression capabilities.

What Problem Does This Paper Address?

The increasing complexity and heterogeneity of wireless communication technologies, from 5G to IoT devices, make effective wireless spectrum management a significant challenge. Anomalies in the spectrum, such as unexpected signals or the absence of expected ones, are difficult to detect and label manually due to the sheer volume and complexity of the data. Existing methods often require band-specific models or expert feature engineering, limiting their generality and scalability. This paper aims to address these limitations by developing a unified, unsupervised approach for robust anomaly detection and localization in wireless spectrum data.

Methods

The core of SAIFE is an Adversarial Autoencoder (AAE). An AAE is a type of generative model that learns to represent data in a compressed latent space and then reconstruct it. The "adversarial" aspect comes from a discriminator network that tries to distinguish between real data distributions and those generated by the autoencoder, forcing the autoencoder to learn more robust and meaningful representations.

Key methodological contributions include:

  • Unsupervised Anomaly Detection: The AAE is trained to reconstruct normal spectrum data. Anomalies are detected by a high reconstruction error, indicating the model struggles to reconstruct data that deviates from its learned normal patterns.
  • Interpretable Feature Extraction: SAIFE is trained in a semi-supervised manner to extract interpretable features like signal bandwidth, class, and center frequency. This allows for a deeper understanding of why an anomaly is occurring.
  • Anomaly Localization: By analyzing the reconstruction error and the learned interpretable features, SAIFE can pinpoint the specific regions in the spectrum where anomalies are occurring.
  • Unified Model: Unlike previous approaches that required separate models for different frequency bands, SAIFE is designed to be trained on multiple bands simultaneously, making it more generalizable.

Datasets

The paper uses a combination of synthetic and real-world datasets for training and evaluation:

  • Synthetic Spectrum Dataset: Consists of four signal types (single-cont, single-rshort, mult-cont, dethop) with varying parameters like bandwidth, SNR, and frequency.
  • Synthetic Anomaly Dataset: Includes signals like scont, randpulses, wpulse, and oclass, used to simulate various anomaly scenarios.
  • Real Wireless Datasets:
    • SDR Dataset: Collected using a HackRF SDR in Belgium, covering frequencies from 10 MHz to 3 GHz across 12 selected frequency bands.
  • Electrosense Dataset: PSD sensor data from Electrosense sensors across Europe, using 7 selected frequency bands.

Key Insights and Ideas

  • AAE for Spectrum Anomalies: Adversarial Autoencoders are highly effective for unsupervised anomaly detection in complex, high-dimensional data like wireless spectrum, outperforming simpler autoencoders by learning more robust latent representations.
  • Interpretability is Key: Extracting interpretable features (bandwidth, class, frequency) alongside anomaly detection provides crucial context for understanding the nature of anomalies, moving beyond simple detection to diagnosis.
  • Unsupervised Learning for Generalization: Training a single model across multiple frequency bands without explicit labeling significantly enhances generalization and reduces the burden of creating band-specific models.
  • Reconstruction Error as a Powerful Anomaly Indicator: The model's inability to accurately reconstruct anomalous signals serves as a reliable indicator, even when anomalies are subtle or overlap with normal signals.
  • Compression as a Byproduct: The learned compressed features not only aid in anomaly detection but also enable significant data compression, which is vital for resource-constrained wireless sensing platforms.

Key Results

  • Anomaly Detection: SAIFE achieves good anomaly detection and localization capabilities across synthetic and real-world datasets. ROC curves demonstrate strong performance, particularly at higher SNRs.
  • Signal Classification: The model achieves near-perfect classification accuracy (close to 100%) on real SDR and Electrosense datasets, and 92.86% on the synthetic dataset, even when trained with only 20% labeled samples.
  • Data Compression: SAIFE offers significant lossy compression ratios, up to 120X on the Electrosense dataset, by using only 20 compressed features to represent the input PSD frame.
  • Anomaly Localization: The model successfully localizes anomalies, providing visual examples of detected anomalies in the spectrum.
  • Real-World Performance: Tested on a 500-hour Electrosense dataset, SAIFE detected various types of anomalies, including missing transmissions and out-of-band signals, demonstrating its practical utility.

Limitations and Future Work

  • Limitations:
    • The paper notes that for certain anomalies (e.g., scont in srshort band), the model might misattribute the anomaly to the wrong class label, although it still correctly identifies it as anomalous.
  • The current model is trained for three semi-supervised features (class label, position, bandwidth).
  • Future Work:
    • Detailed comparisons with other prediction-based anomaly detection models.
  • Evaluation of performance gains using raw IQ samples instead of PSD.
  • Development of similarity scores to select closely located or similar spectrum scanning sensors for unified model deployment.
  • Incorporating user feedback to enable automated retraining and incremental learning for anomaly detection.
  • Addressing the automated spectrum monitoring dream by making the training process fully automated.

Summary updated

Abstract

Detecting anomalous behavior in wireless spectrum is a demanding task due to the sheer complexity of the electromagnetic spectrum use. Wireless spectrum anomalies can take a wide range of forms from the presence of an unwanted signal in a licensed band to the absence of an expected signal, which makes manual labeling of anomalies difficult and suboptimal. We present, spectrum anomaly detector with interpretable features (SAIFE), an adversarial autoencoder (AAE)-based anomaly detector for wireless spectrum anomaly detection using power spectral density (PSD) data. This model achieves an average anomaly detection accuracy above 80% at a constant false alarm rate of 1% along with anomaly localization in an unsupervised setting. In addition, we investigate the model's capabilities to learn interpretable features, such as signal bandwidth, class, and center frequency in a semi-supervised fashion. Along with anomaly detection the model exhibits promising results for lossy PSD data compression up to 120× and semi-supervised signal classification accuracy close to 100% on three datasets just using 20% labeled samples. Finally, the model is tested on data from one of the distributed electrosense sensors over a long term of 500 h showing its anomaly detection capabilities.
BibTeX
@article{rajendran2019unsupervised,
  author = {Sreeraj Rajendran and Wannes Meert and Vincent Lenders and Sofie Pollin},
  title = {{Unsupervised Wireless Spectrum Anomaly Detection With Interpretable Features}},
  journal = {IEEE Trans. Cogn. Commun. Netw.},
  year = {2019},
  doi = {10.1109/tccn.2019.2911524},
}
Figure