Learning Secured Modulation With Deep Adversarial Neural Networks (2005.13694v1)

Abstract: Growing interest in utilizing the wireless spectrum by heterogeneous devices compels us to rethink the physical layer security to protect the transmitted waveform from an eavesdropper. We propose an end-to-end symmetric key neural encryption and decryption algorithm with a modulation technique, which remains undeciphered by an eavesdropper, equipped with the same neural network and trained on the same dataset as the intended users. We solve encryption and modulation as a joint problem for which we map the bits to complex analog signals, without adhering to any particular encryption algorithm or modulation technique. We train to cooperatively learn encryption and decryption algorithms between our trusted pair of neural networks, while eavesdropper's model is trained adversarially on the same data to minimize the error. We introduce a discrete activation layer with a defined gradient to combat noise in a lossy channel. Our results show that a trusted pair of users can exchange data bits in both clean and noisy channels, where a trained adversary cannot decipher the data.