Secure Initial Access and Beam Alignment Using Deep Learning in 5G and Beyond Systems

5G and beyond networks will require fast, energy efficient, and secure initial access. In this study, a deep learning-based secure initial beam selection method is proposed that ranks the beam pairs between a transmitter and a legitimate user aiming to maximize the signal strength the user receives, while keeping the signal strength that the eavesdropper sees below a threshold. Instead of an exhaustive search, the initial beam selection is performed over a limited number of the top beam pairs, leading to reduced communication overhead and energy consumption. The proposed scheme is evaluated using data obtained from a real-life mobile network topology as well as a synthetic data set based on the same geographical site but with statistical system-level environment variables. Utilizing a multi-layer perceptron model, the neural network takes receiver locations as input and produces a ranked list of beam pairs between transmitter and receiver based on the specified coverage criteria. Numerical results show that the signalling overhead can be reduced by 75% with 99.66% accuracy in terms of the best beam pair, and 99.89% of the achievable signal strength. In terms of security, the proposed method has been shown to improve secure coverage probability by 68.12% compared to the best-coverage beam selection scenario.