HiNoVa: A Novel Open-Set Detection Method for Automating RF Device Authentication

Luke Puppo, Weng-Keen Wong, Bechir Hamdaoui, Abdurrahman Elmaghbub

Abstract

Deep learning has revolutionized wireless network security by enabling the identification of devices through unique radio frequency (RF) patterns. However, most existing models operate under a "closed-set" assumption, failing to identify unauthorized devices not present during training. We introduce HiNoVa (Hidden Node Value fingerprinting), a novel open-set detection method that leverages the internal hidden states of a Convolutional Neural Network (CNN) combined with a Long Short-Term Memory (LSTM) network. By analyzing the temporal dependencies and hardware-specific signatures within these hidden states, HiNoVa significantly improves the Area Under the Precision-Recall Curve (AUPRC) across LoRa and WiFi datasets, providing a robust solution for monitoring unauthorized network access.

Introduction

As the Internet of Things (IoT) expands into critical infrastructure, military installations, and government facilities, the need for hardware-level authentication has never been greater. Traditional security protocols, such as MAC address filtering or cryptographic keys, are vulnerable to cloning and man-in-the-middle attacks. RF fingerprinting offers a physical-layer alternative by identifying the unique, unintentional hardware imperfections—"fingerprints"—created during the manufacturing of radio transmitters.

The primary challenge in deploying RF fingerprinting is the open-set problem. In real-world scenarios, a security system will inevitably encounter devices it has never seen before. While standard machine learning models excel at classifying known devices (closed-set), they often struggle to distinguish between a known device and an unauthorized intruder. Furthermore, RF signals are complex time-series data with non-linear dependencies, making traditional open-set methods designed for image processing (like MaxLogit or OpenMax) less effective.

The HiNoVa Methodology

HiNoVa addresses these challenges by monitoring the internal "memory" of a recurrent neural network rather than relying solely on the final classification layer. The methodology consists of four primary stages:

1. Signal Preprocessing

Raw In-phase and Quadrature (IQ) signals are first converted into auto-correlation matrices. This step emphasizes cyclostationary features—patterns that repeat over time due to the specific internal circuitry and clock timings of the transmitter—providing a more stable foundation for fingerprinting than raw waveforms.

2. Hybrid CNN-LSTM Architecture

The system utilizes a hybrid deep learning model. A CNN layer first extracts local spatial features from the signal representations. These features are then passed to an LSTM network, which is specifically designed to capture the long-term temporal dependencies inherent in radio signals. This architecture ensures the model understands both the instantaneous "shape" of the signal and how it evolves over time.

3. Hidden State Fingerprinting

Unlike traditional methods that look at the final output probabilities (logits), HiNoVa looks inside the LSTM. As the model processes a signal, the values of its hidden nodes fluctuate. HiNoVa aggregates these values into a histogram for each known device. This histogram serves as a high-dimensional, unique hardware signature that is much harder to spoof than a simple classification label.

4. Open-Set Detection via Rank Correlation

To determine if a device is authorized, the system generates a test histogram from the incoming signal and compares it against the stored fingerprints of known devices. This comparison is performed using Kendall’s $\tau$, a non-parametric correlation measure. Kendall’s $\tau$ is particularly effective here because it measures the rank-order association between distributions rather than linear relationships, making it robust against noise and signal scaling.

Experimental Results

The researchers evaluated HiNoVa using three distinct RF datasets collected from a testbed of 15 PyCom IoT devices: LoRa, Wireless-WiFi, and Wired-WiFi (the latter using direct SMA connections to isolate hardware signatures from environmental interference).

Performance Metrics

The primary metric used was the Area Under the Precision-Recall Curve (AUPRC), which is the gold standard for "needle-in-a-haystack" detection scenarios where unauthorized devices are rare compared to authorized ones.

  • LoRa Performance: HiNoVa achieved an AUPRC of 0.95, a massive improvement over traditional baselines like MaxLogit, which struggled at approximately 0.42.
  • WiFi Performance: On the Wireless-WiFi dataset, HiNoVa reached a near-perfect 1.00 AUPRC in several experimental configurations.
  • Computational Efficiency: HiNoVa proved significantly more practical than existing time-series methods. For instance, the "Akar" method (based on Dynamic Time Warping) failed to complete the evaluation within 24 hours, whereas HiNoVa processed the data in real-time on standard hardware.

Discussion and Future Work

HiNoVa demonstrates that the internal hidden states of an LSTM contain significantly richer information for hardware identification than the final output layers. By treating RF fingerprinting as a distribution-matching problem rather than a simple classification task, HiNoVa provides a scalable and accurate way to detect unauthorized devices.

Future research will focus on the impact of dynamic environmental factors, such as multi-path fading and high interference, on these hidden-state signatures over long periods. Additionally, the authors suggest exploring attention-based architectures, such as Transformers, to determine if they can capture even more nuanced temporal patterns for enhanced security in increasingly crowded RF environments.

Summary updated

Abstract

New capabilities in wireless network security have been enabled by deep learning, which leverages patterns in radio frequency (RF) data to identify and authenticate devices. Open-set detection is an area of deep learning that identifies samples captured from new devices during deployment that were not part of the training set. Past work in open-set detection has mostly been applied to independent and identically distributed data such as images. In contrast, RF signal data present a unique set of challenges as the data forms a time series with non-linear time dependencies among the samples. We introduce a novel open-set detection approach based on the patterns of the hidden state values within a Convolutional Neural Network (CNN) Long Short-Term Memory (LSTM) model. Our approach greatly improves the Area Under the Precision-Recall Curve on LoRa, Wireless-WiFi, and Wired-WiFi datasets, and hence, can be used successfully to monitor and control unauthorized network access of wireless devices.
BibTeX
@inproceedings{puppo2023hinova,
  author = {Luke Puppo and Weng-Keen Wong and Bechir Hamdaoui and Abdurrahman Elmaghbub},
  title = {{HiNoVa: A Novel Open-Set Detection Method for Automating RF Device Authentication}},
  booktitle = {ISCC},
  year = {2023},
  url = {https://openreview.net/forum?id=ojUbhHAz5C},
}
Figure