A Big Data-Driven DRL Technique Against Simultaneous Eavesdropping and Jamming Attacks in mmWave HetNets

Eavesdropping and jamming attacks pose significant security threats to the integrity and availability of 5G mm Wave wireless networks. Although reinforcement learning (RL)-based schemes are considered viable solutions, the most challenging aspect is obtaining accurate training data. This paper proposes a big data-driven, secrecy-aware achievable data rate maximization scheme for a two-tier fifth-generation heterogeneous network (5G-HetNet) to defend against simultaneous eavesdropping and jamming attacks. By leveraging massive data generated in real-time from 5G networks, the proposed approach efficiently handles large-scale optimization challenges. We present a joint optimization problem on power allocation and beamforming for SUs placement, given minimum requirements for secrecy and data rates. We, then transfrom the resulting non-convex problem as a multi-agent reinforcement learning (MARL) problem by utilizing the Markov decision process (MDP). In order to tackle the MDP's large state and action spaces that are inherent in two discrete event-dynamic 5G HetNets, we propose the multiagent deep reinforcement learning (MADRL) scheme in order to jointly optimize data rate and secrecy rates. The proposed approach utilizes a double Q-architecture based double Deep Q Network (DDQN) to jointly solve the beamforming and power allocation vectors for the SUs. We compare our proposed DDQN approach to both Q-learning and DQN in terms of data and secrecy rate performance. The results of our simulations show that our proposed approach significantly improves the attainable secrecy rate by 22 % and 33% compared to DQN and Q-Learning, respectively.