Artificial noise-based generative adversarial network for secure and robust wireless device-to-device signal transmissions

This paper presents a secure and noise-resilient communication scheme which has been designed for proximity-based device-to-device networks. A set of 50 original signals having the power range between 23–25 dB were generated and used as a set of real signals. A generative adversarial network (GAN) was then used to synthesize 500 deep-fake signals that were used to emulate the statistical structure of the original 50 signals. The signals were passed through a noisy channel which had the noise of additive white Gaussian noise (AWGN). At the receiving end, a supervised denoising neural network was trained using the clean and noisy signals to recover the original signal. The receiver was able to regenerate the original signal without any error and the correlation of the original signal and the signals regenerated by the receiver was high. In contrast, an eavesdropper (C) was trained only on one intercepted signal without access to clean references and the GAN network coefficients. The eavesdropper failed to reconstruct the meaningful signals and had very low correlation with the original signal. The strong denoising ability of the receiver and the communication security is validated by the evaluation metrics, signal-to-noise ratio, correlation coefficient, and Shannon capacity.