Abstract We present a convolutional neural network for identifying radio frequency devices from signal data, in order to detect possible interference sources for wireless local area networks. Collecting training data for this task is difficult because the interference is often intermittent, leading to noisy labels. We propose a semi-supervised learning approach that uses a small amount of labeled data and a large amount of unlabeled data. Our approach uses a structured pseudo-labeling technique that takes advantage of the temporal continuity of the interference. We evaluate our approach on a dataset of 13 different interference sources and show that it achieves an accuracy of 97%.
Introduction
The 2.4GHz and 5GHz unlicensed bands are the "Wild West" of wireless communication. While these bands are primarily known for Wi-Fi, they are also home to a vast array of other devices, including Bluetooth headsets, Zigbee sensors, microwave ovens, and wireless cameras. When a wireless network experiences performance degradation, identifying the specific source of interference is a critical first step toward mitigation.
Traditional interference identification systems, such as AirShark, rely on hand-crafted features like pulse signatures or duty cycles. These features are difficult to design and require constant updates as new devices enter the market. Furthermore, collecting high-quality training data is challenging; for example, a microwave oven may only emit interference when its magnetron cycles, leading to "noisy" labels where the recorded data does not always match the "on" timestamp.
Methodology
The proposed approach treats radio frequency (RF) sweeps as 2D images, allowing the system to leverage the power of Convolutional Neural Networks (CNNs) for automatic feature discovery.
Data Representation and Labeling
The researchers transformed 1D RF sweeps into 2D spectrograms. Each image represents a 2-second window, with frequency on one axis (3000 bins) and time on the other (50 slices). To address the labeling challenge, the study employs a semi-supervised framework.
As shown in Figure 1, data is collected in "off-on-off" cycles. Only a small portion of the "Device on" period is used as a labeled training set, while the rest remains unlabeled or is reserved for testing.
CNN Architecture
The model uses a relatively shallow CNN architecture to learn spectral signatures automatically. This design avoids the need for human-engineered features and allows the model to become "frequency invariant"—learning the visual signature of a device regardless of the specific channel it occupies.
The architecture (Figure 2) consists of two convolutional layers (with 15 and 30 filters respectively), each followed by max-pooling, and two fully connected layers. The final layer outputs a probability distribution across the device classes.
Structured Pseudo-Labels
To improve performance with limited labeled data, the authors introduced Structured Pseudo-Labels. Standard pseudo-labeling involves a model predicting labels for unlabeled data and using high-confidence predictions as ground truth for further training. The authors enhanced this by adding a Markovian constraint: since radio interference is temporally continuous, the sequence of labels should make physical sense. By applying the Viterbi algorithm, the model penalizes erratic "flickering" between device types, ensuring temporal consistency.
Experimental Setup
The dataset was collected using an Ekahau spectrum analyzer and included 13 different interference sources, such as wireless video cameras, baby monitors, microwave ovens, and jammers. Each device was recorded for approximately three minutes (one minute background, one minute active, one minute background), resulting in roughly 10,000 two-second instances.
Key Results and Insights
- High Accuracy: The CNN achieved 97% accuracy across the 13 devices, significantly outperforming the AirShark baseline (79%).
- Data Efficiency: The model reached near-peak performance with only tens of seconds of labeled data per device, demonstrating the effectiveness of the semi-supervised approach.
- Robustness to Noisy Labels: For devices like microwave ovens that cycle power, the CNN correctly identified "off" periods during the "on" cycle, effectively correcting the initial human-provided ground truth.
- Transfer Learning: The network learns a "universal" representation of RF signals. The authors found that adding a new device type only required re-training the final layer, making the system highly scalable.
Conclusion and Future Work
This research demonstrates that treating RF interference as a computer vision problem significantly improves identification accuracy while reducing the burden of manual feature engineering and data labeling. However, the current study focuses on single-source interference. Future work will explore multi-source scenarios where multiple devices overlap, as well as localization techniques to estimate the signal strength and distance of the interferer.
Summary updated