Abstract: Aiming at the problem of self-interference cancellation when the full-duplex mode is adopted by the legitimate receiver, a scheme is presented for signal combining and self-interference cancellation based on neural networks at the full-duplex legitimate receiver in a single-input multi-output system. The legitimate receiver, while receiving signals, sends artificial noise to interfere with the eavesdropper. Two neural networks are designed, one combining the signals received by multiple antennas, and the other reconstructing self-interference for the self-interference cancellation of the received signals. The bit error rates of the legitimate receiver and the eavesdropper and the achievable secrecy rate are simulated. The simulation results show that the signal combination and self-interference cancellation scheme is feasible and effective. They also show that a considerable secrecy rate can be achieved when the transmitting antennas and receiving antennas at the legitimate receiver are properly allocated.

Key words: full duplex, physical layer security, neural network, signal combining, self-interference cancellation