CONIX Publication

Deep learning approaches have proven to be a viable method for identification of RF devices based on their RF fingerprints, which can be used for low layer security protocols and enable new applications at the edge. Deep-learning-based classification of transmitters based on their RF fingerprints is not resilient to the impairments introduced to the radio signals at the receiver. For this reason, a deep-learning transmitter classifier trained on some receiver A, will not perform well when operating on a new receiver B. To solve this problem, we developed a novel approach based on a feature extractor (FE) that extracts receiver-invariant features from IQ samples, i.e. features that are not susceptible to receiver impairments. The receiver-invariant FE is a deep neural network trained using novel approaches that use a calibration dataset of signals transmitted by N transmitters and received at M receivers. The FE is only trained once and can be deployed on new receivers (other than M receivers that were used to train the FE) to assist in classification of new transmitters (other than N transmitters used to train the FE). We evaluate several different approaches for training of receiver-invariant FEs inspired by the state-of-the-art algorithms in domain-invariant deep learning. The developed approaches show improvement in transferability of deep-learning transmitter classifiers from one receiver to another, compared to the baseline approach which does transmitter classification by directly using the IQ samples. Furthermore, we evaluate other signal-processing methods that can improve the transferability of transmitter classifiers, such as applying channel-equalization on the IQ samples or using data collected over multiple days to train the transmitter classifier. These methods help negate the receiver-specific channel and background interference effects.