Neural Network:
Signal Bias Analysis
Are machine learning models in Radio Frequency Fingerprinting truly learning hardware characteristics, or simply the wireless channel?
The Promise & Challenge of RFF
The increase of Internet of Things (IoT) devices has created a high demand for new security methods. Radio Frequency Fingerprinting (RFF) identifies devices by unique, subtle imperfections in their radio signals, much like a human fingerprint. These fingerprints are often found in the signal's turn-on transient—the initial burst of energy when a device begins to transmit.
While promising, a major challenge is that the wireless channel, or the environment the signal travels through, can distort the signal and obscure the device's true fingerprint.
Experimental Setup
Algorithmic learning models may primarily learn to differentiate between channel conditions rather than transmitter hardware. The subtle variations in the hardware fingerprint may be overshadowed by the effects of the channel.
To isolate the effects of the wireless channel, a controlled experiment was conducted using a single transmitter. A HackRF One SDR transmitted a simple on-off keying (OOK) signal, which was captured by a BladeRF 2.0 micro xA4 sampling at 60 MS/s.
Data was collected at five distinct physical locations. The receiver was initially placed at a starting position and then moved 3 inches farther away from the transmitter for each subsequent collection. Crucially, the data was labeled based on the receiver's physical position (0, 1, 2, 3, 4), forcing the model to classify the channel and not the device.
A set of 11 statistical features (including energy, standard deviation, skewness, and kurtosis) was engineered from the raw IQ data to capture the shape and spectral properties of the transient's energy envelope. This was fed into a Random Forest classifier configured with 100 estimators.
Summary of Findings
- >>
Our model achieved 91.68% accuracy in classifying the signal's origin based solely on the receiver's physical location, using a single transmitter.
- >>
The model learned to identify channel-sensitive features like signal energy and amplitude variation, not a hardware fingerprint.
- !!
Conclusion: This result strongly supports our hypothesis that machine learning models can easily mistake channel effects for device signatures.
The primary implication is that high accuracy alone is not a reliable benchmark for RFF systems. This is a call to action to develop more robust, channel-invariant methods.