Spoof Testing of Synthetic Radio Frequency Waveforms via Generative Machine Learning Methods
Open Access DepositedRadio Frequency (RF) research and development is an expensive endeavor due tohigh costs in educated labor and expensive hardware. This research poses to provide some relief to that expense via proposal of generative machine learning to generate synthetic RF waveforms. RF waveforms are generated through training and use of a generative adversarial network (GAN) while assessing 4 different loss functions, 3 different data transformations, and using 2 different datasets in training. The loss functions comprise of using short term Fourier transforms in combination with a Frobenius norm, a Euclidean norm, a phase correlation, and a waveform ambiguity function. The data transformations were used to reduce noise in the waveforms, and the two different transformations used are the log function and the Wiener function. The third transformation is the composition of the two functions. Finally, 2 different datasets were assessed, where each dataset contains waveforms of different modulations schemes. This presents a total of 24 training scenarios which were run 3 times over differing number of training and evaluation iterations, first reducing the training iterations from 400,000 iterations to 60,000 iterations, and second, keeping iterations at 60,000 and increasing the evaluation dataset from 16% to 26%. After training the model and recording the results 72 times, once for each of the permutation described, it was found that the best performing model, statistically, was trained on dataset 2, transformed by the Wiener function using the Frobenius loss function trained for 400,000 iterations. The worst performing result was found to be trained on dataset 1 having undergone the composition log and Wiener transformation, trained via the phase correlation loss function for all vi iteration and evaluation parameters. The results were then analyzed via RF spoofing detector to better understand the value and performance as RF waveforms. In spoofing it was seen that the model and training data producing the statistically best synthetic data performed worst at spoofing, and the model and training data producing the statistically worst data performed best at spoofing. These results align with the statistical analysis since greater spoofing performance implies greater errors.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.